Daily Papers

We measure how well current large language models coordinate as multiple agents sharing a common resource, using the dining philosophers problem as a clean test bed. Across 630 episodes spanning seven models and three philosopher counts, four frontier closed-source systems reach mean reward 0.45 to 0.87 and Mistral-Small 24B reaches 0.83 to 0.99, while Qwen3-14B reaches 0.13 to 0.35. We then ask whether group relative policy optimization (GRPO) on rollouts from the task itself can close the gap and find that it cannot: a Welch's t-test on per-episode reward at five philosophers gives p = 0.66 and a Hedges' g of -0.11, with no statistically significant change at ten or fifteen philosophers either. Two further observations qualify the result. The training reward of both 8B and 14B runs peaked at step nine and then declined, so the default saved checkpoint at step 15 is strictly worse than several earlier ones. The four-term reward we use admits a degenerate maximum at zero actions, which DeepSeek-R1-Distill-Qwen-7B and Mistral-Small 24B at five philosophers both inhabit, with mean reward 1.0 and 0.83 respectively at zero meals. The bottleneck for an open-weight 14B model on multi-agent coordination is not training compute but training methodology: reward shaping that does not collapse to a no-action maximum, checkpoint discipline that does not depend on the final step, and curriculum across problem scales.

Software development is iterative, yet agentic coding benchmarks overwhelmingly evaluate single-shot solutions against complete specifications. Code can pass the test suite but become progressively harder to extend. Recent iterative benchmarks attempt to close this gap, but constrain the agent's design decisions too tightly to faithfully measure how code quality shapes future extensions. We introduce SlopCodeBench, a language-agnostic benchmark comprising 20 problems and 93 checkpoints, in which agents repeatedly extend their own prior solutions under evolving specifications that force architectural decisions without prescribing internal structure. We track two trajectory-level quality signals: verbosity, the fraction of redundant or duplicated code, and structural erosion, the share of complexity mass concentrated in high-complexity functions. No agent solves any problem end-to-end across 11 models; the highest checkpoint solve rate is 17.2%. Quality degrades steadily: erosion rises in 80% of trajectories and verbosity in 89.8%. Against 48 open-source Python repositories, agent code is 2.2x more verbose and markedly more eroded. Tracking 20 of those repositories over time shows that human code stays flat, while agent code deteriorates with each iteration. A prompt-intervention study shows that initial quality can be improved, but it does not halt degradation. These results demonstrate that pass-rate benchmarks systematically undermeasure extension robustness, and that current agents lack the design discipline iterative software development demands.

Checkpointing to preserve training states is crucial during the development of Large Foundation Models (LFMs), for training resumption upon various failures or changes in GPU resources and parallelism configurations. In addition, saved checkpoints are dispatched to evaluation tasks or transferred across different training stages (e.g., from pre-training to post-training). All these scenarios require resharding distributed checkpoints from one parallelism to another. In production environments, different LFMs are trained with various frameworks and storage backends, depending on model sizes and training scales. A high-performance checkpointing system is needed to enable efficient checkpoint management at scale throughout the lifecycle of LFM development. We introduce ByteCheckpoint, an industrial-grade checkpointing system for large-scale LFM training. ByteCheckpoint features: a parallelism-agnostic checkpoint representation that enables efficient load-time checkpoint resharding; a generic checkpoint saving/loading workflow to accommodate multiple training frameworks and support different storage backends; full-stack optimizations to ensure high I/O efficiency and scalability; a suite of monitoring tools to streamline large-scale performance analysis and bottleneck detection. Compared to existing open-source checkpointing systems [52, 58], ByteCheckpoint significantly reduces runtime checkpoint stalls, achieving an average reduction of 54.20x. For saving and loading times, ByteCheckpoint achieves improvements of up to 9.96x and 8.80x, respectively.

Multi-step enterprise agent tasks fail in a characteristic way: single-pass inference has no checkpoint between deciding an answer and committing to it. We study one production system (Leni) whose architecture installs such checkpoints: verification loops (execute, observe, compare, correct) staffed by lightweight task-specialized post-trained models. We evaluate the unmodified production configuration on three public benchmarks stressing distinct failure modes: SpreadsheetBench Verified (silent computation error), BullshitBench v2 (premise confabulation), and the GAIA validation split (cascade error over long tool chains). The full system improves over its frontier base model by +11.0 percentage points on SpreadsheetBench (91.25% vs 80.25%, n=400, p<0.001), +7 to +10 percentage points on BullshitBench (98% vs 91%, n=100), and roughly +15 points on GAIA validation (75.2% pass@1, n=165; 83.0% best-of-k). Our central contribution is a decomposition of that uplift: most of it comes from scaffolding, routing, and specialist models rather than from the verification step itself, whose isolated contribution is small (+1.5 points) but concentrated at the top of the score distribution, where it converts otherwise-failing tasks. We instrument the loop end-to-end, yielding an empirical verifier confusion matrix (catch rate about 0.20, fix rate 0.75, no false-alarm regressions) that grounds a compounding-reliability model. Specialist-swap ablations suggest that the loop's value depends on who observes it: replacing the small trained verifier with the generating frontier model eliminates most rescues. A valid-premise control shows zero over-rejections in 100 expert-level questions.

Quantized checkpoints are often screened first with quality metrics and only later, if at all, with direct safety tests. This paper audits that shortcut on a matched 51-row matrix spanning 6 models, 4 families, a 7-level GGUF ladder, and AWQ/GPTQ INT4 checkpoints. In this matrix the shortcut fails: all 36 quality-safety pairings split direction across models, and 9 hidden-danger rows plus 1 near-hidden-danger row show quality stable or improved while refusal falls by 12-68 percentage points. Seven of the 11 AWQ/GPTQ rows are hidden-danger. A four-probe mechanistic follow-up over the 17 Hugging Face-backed FP16/AWQ/GPTQ cells does not rescue it: entropy, refusal-direction, and calibration probes are weak or null separators of dangerous rows, and although probe-identified safety-associated neurons absorb 1.39times more quantization error overall (p < 5 times 10^{-7}), the effect is not regime-specific. Claude Sonnet 4 relabels 11,470 items in a predefined stratified set, agrees with the primary gemma3:12b judge on 89.9\% of rows (κ= 0.873, 95\% CI [0.866, 0.881]), and changes 0/10 hidden-danger cells. A calibrated study-internal behavioral screen -- the Refusal Template Stability Index (RTSI), built from four refusal-template drift features and calibrated on this matrix -- routes 10/10 hidden- or near-hidden-danger rows to direct safety testing (Wilson 95\% CI lower bound 0.72) while leaving 23 of 45 non-baseline rows in a low-risk bucket under both in-sample scoring and row-level leave-one-out validation; on the same matrix, the best single-feature baselines (unique-prefix-rate-delta, raw refusal-rate delta) recover 9/10 and 8/10 respectively at matched bucket size, and cross-stack transfer requires recalibration. For the quantized checkpoints, model families, and safety outcomes studied here, retained quality cannot waive direct safety evaluation.

Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.

When training large language models (LLMs), it is common practice to track downstream task performance throughout the training process and select the checkpoint with the highest validation score. However, downstream metrics often exhibit substantial fluctuations, making it difficult to identify the checkpoint that truly represents the best-performing model. In this study, we empirically analyze the stability of downstream task performance in an LLM trained on diverse web-scale corpora. We find that task scores frequently fluctuate throughout training, both at the aggregate and example levels. To address this instability, we investigate two post-hoc checkpoint integration methods: checkpoint averaging and ensemble, motivated by the hypothesis that aggregating neighboring checkpoints can reduce performance volatility. We demonstrate both empirically and theoretically that these methods improve downstream performance stability without requiring any changes to the training procedure.

We study how to predict the downstream closed-loop performance of a learned latent world model from validation-time diagnostics alone. Choosing the right checkpoint from a world-model training run is difficult: validation loss and multi-step prediction RMSE keep improving long after closed-loop performance has collapsed. We present a suite of structural validation-time diagnostics drawn from optimal-control theory and apply them to Gymnasium's LunarLander v3, which features shaped rewards. We train an RSSM [5, 4] world model on it and treat per checkpoint CEM-MPC return as the oracle for closed-loop quality. By evaluating 40 metrics against this oracle, we find that the strongest single predictor is the Reward Observability Fraction (ROF), which measures the reward predictor's dependence on the observable subspace. We combine ROF with three structural regularizers into a single-number offline checkpoint selection score, the Composite Reward Observability Fraction (CROF). The CROF-selected world model trains a model-based A2C policy that beats a fairly evaluated model-free A2C baseline by ~24.5 return points while using ~65x fewer real-environment interactions, and the same world model also drives a strong zero-shot CEM-MPC policy. Code and data: https://github.com/nsmoly/LunarLander_RSSM.

Safety alignment in Large Language Models is critical for healthcare; however, reliance on binary refusal boundaries often results in over-refusal of benign queries or unsafe compliance with harmful ones. While existing benchmarks measure these extremes, they fail to evaluate Safe Completion: the model's ability to maximise helpfulness on dual-use or borderline queries by providing safe, high-level guidance without crossing into actionable harm. We introduce Health-ORSC-Bench, the first large-scale benchmark designed to systematically measure Over-Refusal and Safe Completion quality in healthcare. Comprising 31,920 benign boundary prompts across seven health categories (e.g., self-harm, medical misinformation), our framework uses an automated pipeline with human validation to test models at varying levels of intent ambiguity. We evaluate 30 state-of-the-art LLMs, including GPT-5 and Claude-4, revealing a significant tension: safety-optimised models frequently refuse up to 80\% of "Hard" benign prompts, while domain-specific models often sacrifice safety for utility. Our findings demonstrate that model family and size significantly influence calibration: larger frontier models (e.g., GPT-5, Llama-4) exhibit "safety-pessimism" and higher over-refusal than smaller or MoE-based counterparts (e.g., Qwen-3-Next), highlighting that current LLMs struggle to balance refusal and compliance. Health-ORSC-Bench provides a rigorous standard for calibrating the next generation of medical AI assistants toward nuanced, safe, and helpful completions. The code and data will be released upon acceptance. red{Warning: Some contents may include toxic or undesired contents.}

Tool-using LLM agents can violate the very policies they are deployed to enforce while appearing to complete the task successfully. In policy-permissive environments, a tool may execute any well-formed call even when the corresponding state transition is forbidden by domain policy. The result is a silent wrong state (a booking cancelled, a passenger count changed, a claim acted on without verification) that neither the tool nor the agent's self-report exposes. We study this failure mode in the τ^2-bench airline domain. On a budget agent, 78% of observed failures are silent wrong-state failures with no tool error, and the aggregate failure rate is reproducible across disjoint seeds, not sampling noise. We then evaluate a lightweight intervention: deterministic, read-only pre-execution gates that inspect the proposed call and current state before allowing a write. A four-gate suite raises full-benchmark success from 29.6% to 42.0% on gpt-4o-mini (+12.4pp; paired task-level bootstrap P=0.0012), and the lift reproduces on a disjoint 15-seed set (+12.3pp; P=0.0008). The effect is concentrated where the gates fire: on the 26/50 firing tasks, success rises by +19.2pp, while movement on the 24 non-firing tasks does not exclude zero. Two negative controls (a self-enforcing retail domain and BFCL) bound the mechanism: gates help when tools are policy-permissive and add little where tools already self-enforce. As suggestive evidence, not a central claim, the same failure mode persists at the frontier: gpt-5.2 at default reasoning still attempts policy-violating writes, and the same suite improves success from 61.2% to 71.6% (+10.4pp; P=0.020; n=5, no replication). The contribution is a bounded evaluation and reliability result: deterministic gates do not guarantee task success, but they can deterministically prevent a known class of silent policy-violating writes at the action boundary.

Existing checkpointing approaches seem ill-suited for distributed training even though hardware limitations make model parallelism, i.e., sharding model state across multiple accelerators, a requirement for model scaling. Consolidating distributed model state into a single checkpoint unacceptably slows down training, and is impractical at extreme scales. Distributed checkpoints, in contrast, are tightly coupled to the model parallelism and hardware configurations of the training run, and thus unusable on different configurations. To address this problem, we propose Universal Checkpointing, a technique that enables efficient checkpoint creation while providing the flexibility of resuming on arbitrary parallelism strategy and hardware configurations. Universal Checkpointing unlocks unprecedented capabilities for large-scale training such as improved resilience to hardware failures through continued training on remaining healthy hardware, and reduced training time through opportunistic exploitation of elastic capacity. The key insight of Universal Checkpointing is the selection of the optimal representation in each phase of the checkpointing life cycle: distributed representation for saving, and consolidated representation for loading. This is achieved using two key mechanisms. First, the universal checkpoint format, which consists of a consolidated representation of each model parameter and metadata for mapping parameter fragments into training ranks of arbitrary model-parallelism configuration. Second, the universal checkpoint language, a simple but powerful specification language for converting distributed checkpoints into the universal checkpoint format. Our evaluation demonstrates the effectiveness and generality of Universal Checkpointing on state-of-the-art model architectures and a wide range of parallelism techniques.

We introduce LlavaGuard, a family of VLM-based safeguard models, offering a versatile framework for evaluating the safety compliance of visual content. Specifically, we designed LlavaGuard for dataset annotation and generative model safeguarding. To this end, we collected and annotated a high-quality visual dataset incorporating a broad safety taxonomy, which we use to tune VLMs on context-aware safety risks. As a key innovation, LlavaGuard's new responses contain comprehensive information, including a safety rating, the violated safety categories, and an in-depth rationale. Further, our introduced customizable taxonomy categories enable the context-specific alignment of LlavaGuard to various scenarios. Our experiments highlight the capabilities of LlavaGuard in complex and real-world applications. We provide checkpoints ranging from 7B to 34B parameters demonstrating state-of-the-art performance, with even the smallest models outperforming baselines like GPT-4. We make our dataset and model weights publicly available and invite further research to address the diverse needs of communities and contexts.

LLMs have seen rapid adoption in all domains. They need to be trained on high-end high-performance computing (HPC) infrastructures and ingest massive amounts of input data. Unsurprisingly, at such a large scale, unexpected events (e.g., failures of components, instability of the software, undesirable learning patterns, etc.), are frequent and typically impact the training in a negative fashion. Thus, LLMs need to be checkpointed frequently so that they can be rolled back to a stable state and subsequently fine-tuned. However, given the large sizes of LLMs, a straightforward checkpointing solution that directly writes the model parameters and optimizer state to persistent storage (e.g., a parallel file system), incurs significant I/O overheads. To address this challenge, in this paper we study how to reduce the I/O overheads for enabling fast and scalable checkpointing for LLMs that can be applied at high frequency (up to the granularity of individual iterations) without significant impact on the training process. Specifically, we introduce a lazy asynchronous multi-level approach that takes advantage of the fact that the tensors making up the model and optimizer state shards remain immutable for extended periods of time, which makes it possible to copy their content in the background with minimal interference during the training process. We evaluate our approach at scales of up to 180 GPUs using different model sizes, parallelism settings, and checkpointing frequencies. The results show up to 48times faster checkpointing and 2.2times faster end-to-end training runtime compared with the state-of-art checkpointing approaches.

Ensuring the safety of LLM-generated content is essential for real-world deployment. Most existing guardrail models formulate moderation as a fixed binary classification task, implicitly assuming a fixed definition of harmfulness. In practice, enforcement strictness - how conservatively harmfulness is defined and enforced - varies across platforms and evolves over time, making binary moderators brittle under shifting requirements. We first introduce FlexBench, a strictness-adaptive LLM moderation benchmark that enables controlled evaluation under multiple strictness regimes. Experiments on FlexBench reveal substantial cross-strictness inconsistency in existing moderators: models that perform well under one regime can degrade substantially under others, limiting their practical usability. To address this, we propose FlexGuard, an LLM-based moderator that outputs a calibrated continuous risk score reflecting risk severity and supports strictness-specific decisions via thresholding. We train FlexGuard via risk-alignment optimization to improve score-severity consistency and provide practical threshold selection strategies to adapt to target strictness at deployment. Experiments on FlexBench and public benchmarks demonstrate that FlexGuard achieves higher moderation accuracy and substantially improved robustness under varying strictness. We release the source code and data to support reproducibility.

Model merging has shown great promise at combining expert models, but the benefit of merging is unclear when merging ``generalist'' models trained on many tasks. We explore merging in the context of large (sim100B) models, by recycling checkpoints that exhibit tradeoffs among different tasks. Such checkpoints are often created in the process of developing a frontier model, and many suboptimal ones are usually discarded. Given a pool of model checkpoints obtained from different training runs (e.g., different stages, objectives, hyperparameters, and data mixtures), which naturally show tradeoffs across different language capabilities (e.g., instruction following vs. code generation), we investigate whether merging can recycle such suboptimal models into a Pareto-optimal one. Our optimization algorithm tunes the weight of each checkpoint in a linear combination, resulting in a Pareto-optimal models that outperforms both individual models and merge-based baselines. Further analysis shows that good merges tend to include almost all checkpoints with with non-zero weights, indicating that even seemingly bad initial checkpoints can contribute to good final merges.

Open-weight language models are fine-tuned, quantized, pruned, and merged, yet their provenance is often undocumented. We study data-free white-box lineage verification: can weights alone reveal whether two compatible model checkpoints share ancestry? Residual training produces a shared identity-aligned component in branch products, so this structure alone cannot establish ancestry. We remove it and compare checkpoint-specific structure across residual blocks, yielding a symmetric lineage score calibrated against independent checkpoints. On residual-MLP and GPT-2 benchmarks, the score separates fine-tuned, LoRA-merged, pruned, and quantized descendants from independent and distilled models (AUROC=1.0), distinguishing weight ancestry from behavioral similarity. Under function-preserving checkpoint laundering experiments, weight-space baselines lose margin or fail; our score remains unchanged and runs 76x faster than the nearest robust baseline on GPT-2. The projection-pairing signal appears across six language-model families and beyond, and a case study correctly identifies 3 related and 7 unrelated LLaMA-2 public checkpoints. Collectively, these results establish a passive, data-free provenance signal for compatible open-weight language-model checkpoints

LeWorldModel trains a latent world model with a prediction loss and a single anti-collapse regulariser, and reports approximately 87% of goals reached on TwoRoom, its simplest diagnostic environment. We reproduce that result by independent reimplementation on roughly $25 of rented compute, with all evaluation on one laptop CPU. We reach 94.0% at the repository's evaluation goal offset, against 84.0% for the authors' own released checkpoint measured under our protocol on identical episodes, and we reproduce the reported representation result directly (position probe Pearson r = 0.9988 against a reported 0.996). Reaching that point required four conventions that determine the outcome and appear in no released configuration file: dense action gathering across a frameskip block, a programmatically-set action-encoder width, ImageNet pixel normalisation, and action z-scoring. A reproducer following the released configurations alone obtains a model whose predictor cannot converge. The evaluation protocol is itself contested by the released material. The paper's appendix and the repository's configuration specify different goal offsets and step budgets; on the authors' own weights these yield 14.0% and 84.0%, and only the configuration's values reproduce the reported figure. On fifty identical episodes, changing nothing but how the goal is constructed moves that checkpoint from 84.0% to 8.0%. Two findings generalise. One-step prediction accuracy does not predict long-horizon planning success: across three checkpoints spanning a sevenfold range in prediction error, including the authors' own, it orders short-horizon success monotonically and fails to order long-horizon success at all. And a batch normalisation layer inflated our reported validation loss by up to a factor of 300, concealing a training loss that was flat throughout.

Modern deployments require LLMs to enforce safety policies at scale, yet many controls rely on inference-time interventions that add recurring compute cost and serving complexity. Activation steering is widely used, but it requires runtime hooks and scales cost with the number of generations; conditional variants improve selectivity by gating when steering is applied but still retain an inference-time control path. We ask whether selective refusal can be moved entirely offline: can a mechanistic understanding of category-specific refusal be distilled into a circuit-restricted weight update that deploys as a standard checkpoint? We propose C-Δθ: Circuit Restricted Weight Arithmetic, which (i) localizes refusal-causal computation as a sparse circuit using EAP-IG and (ii) computes a constrained weight update ΔθC supported only on that circuit (typically <5% of parameters). Applying ΔθC yields a drop-in edited checkpoint with no inference-time hooks, shifting cost from per-request intervention to a one-time offline update. We evaluate category-targeted selectivity and capability retention on refusal and utility benchmarks.

The rapid proliferation of large language models (LLMs) and the growing diversity of their applications presents a unique optimization opportunity: selecting the right model for the task, while optimizing for speed, cost, and quality at a per-task level. However, inference endpoints can vary widely in quality, price, latency, context support, tool use, domain expertise, and reasoning behavior. This heterogeneity makes manual heuristics difficult to maintain and unlikely to achieve consistently favorable speed--cost--quality trade-offs on their own. We introduce , a lightweight GLiClass-based router that assigns a suitability score to each inference-time model label without autoregressive generation. The released 0.6B-parameter checkpoint combines a Qwen3 decoder with a shallow bidirectional scorer. Its decoder-KV execution path preserves a text-only key--value cache across a session, encodes only new dialogue turns, and evaluates transient candidate-label tokens without adding them to the persistent cache. The same checkpoint also predicts task type, difficulty, reasoning mode, and expected output length, and supports custom zero-shot labels. For task generation, we construct a task ontology with 23 families, 115 task types, 345 routable subtypes, 1,173 synthetic examples, and an orthogonal axis of 30 domains. Using this structure, we generate 150,000 verifier-scored tasks and 15,000 open-ended tasks. We then train the Qwen3 decoder on these tasks, while explicitly separating learned request prediction from per-task policies for attributes such as eligibility, cost, cache reuse, safety, and sovereignty. Across six LiveBench subsets, the router outperforms the mean candidate; on the selected 1,000-task subset, it achieves an aggregate top-1 score of 0.707 versus 0.696 for the strongest fixed model, with benchmark-dependent gains.

Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit "think in English" is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.

Fine-tuning adapts aligned large language models (LLMs) to downstream tasks, but a few dozen harmful examples can remove their refusal of harmful requests. Prior work localizes safety-related behavior to specific layers, directions, and tokens, suggesting targets for protection. We test whether successful localization and recovery support defenses that survive changes in the attack. Across six checkpoints from four model families, harmful and benign prompts remain linearly separable after attack, and patching full clean hidden states into the compromised model restores refusal at a reproducible transition depth. Building on a prior layer-freezing defense, we freeze every layer up to this depth and repeat the attack. At a hundred harmful examples, refusal remains near zero on all six checkpoints, with recovery transitions above the frozen boundary. In a second study, removing the update's top two singular directions restores refusal after short attention-only fine-tunes on four checkpoints. On Llama-3.1-8B, ordinary training changes weaken this repair and an attacker who spreads the update defeats it. A spectral detector calibrated on benign Llama fine-tunes misses most repair failures on that checkpoint. Localized freezing can nevertheless help preserve refusal when a few harmful examples enter training data unintentionally. These results show that an attacker can bypass a region identified by recovery and defeat a repair that works across multiple checkpoints, motivating five checks for defenses against adaptive fine-tuning. Code is available at https://github.com/js-lee-AI/refusal-relocates.

Robotic manipulation often requires acting on information that is no longer visible, yet Vision-Language-Action policies are usually evaluated when the current observation largely determines the next action. Existing robotic memory benchmarks expose this gap, but they still rely mainly on final task success and therefore conflate forgetting with manipulation failure. We present MEMOBench, a benchmark for process level memory evaluation in robotic manipulation. MEMOBench includes 30 history dependent tasks, 1{,}500 expert demonstrations, and 4{,}200 executable checkpoint instances from 84 templates. Each checkpoint pairs coarse to fine language with a simulator predicate and labels one memory operation: Storage, Update, or Compression. These annotations define Memory Storage Rate, Memory Update Rate, and Memory Compression Rate, which measure memory fidelity alongside task success. Across standard and memory augmented VLA policies, the strongest memory module baseline reaches only 31.9\% average success rate, and high storage often coexists with weak update and compression. Checkpoint language also supervises semantic, contrastive, and framewise memory alignment objectives, yielding modest gains across different memory operations. MEMOBench provides a diagnostic evaluation suite and training supervision for memory grounded robotic policies. The project page is available at https://github.com/Collab-Gen/MEMOBench.

Language-model agents are increasingly deployed under standing instructions: a system prompt, a policy file, or a skills document is placed in context, and the agent is trusted to let that document govern every action that follows. Existing benchmarks rarely test this deployment pattern directly; they measure whether an agent can complete a task, not whether a long, binding policy document constrains its behavior over an extended tool-use horizon. We present HANDBOOK_md, a benchmark of 65 agentic tasks modeled on how employees follow company handbooks. Each task places an agent in a self-contained company environment (a file workspace with mock email, chat, calendar, issue-tracking, and commerce services exposed over the Model Context Protocol) and instructs it to carry out routine professional work governed by an expert-written standard operating procedure of 20-124 pages. Tasks span five domains (finance, medical billing, insurance, logistics, and HR) and 10 fictional companies. To resist memorization, every task modifies one of 10 base handbooks, altering the specific rules and thresholds on which grading depends, so no two tasks share the same set of policies. Grading is fully deterministic: each task carries a rubric of programmatic criteria (824 in total) that check both that required actions occurred and that prohibited actions did not. Under strict grading, where a trial passes only if every criterion is satisfied, the strongest evaluated model passes 36.2% of trials, and most frontier models remain below 25%. Failures follow consistent patterns: agents let a plausible but unauthorized in-environment request override the standing policy, perform a required check and then act against its result, lose rule details over long horizons, and report compliance they did not achieve. We release the tasks, environments, and evaluation harness.

Extensive system scales (i.e. thousands of GPU/TPUs) and prolonged training periods (i.e. months of pretraining) significantly escalate the probability of failures when training large language models (LLMs). Thus, efficient and reliable fault-tolerance methods are in urgent need. Checkpointing is the primary fault-tolerance method to periodically save parameter snapshots from GPU memory to disks via CPU memory. In this paper, we identify the frequency of existing checkpoint-based fault-tolerance being significantly limited by the storage I/O overheads, which results in hefty re-training costs on restarting from the nearest checkpoint. In response to this gap, we introduce an in-memory fault-tolerance framework for large-scale LLM pretraining. The framework boosts the efficiency and reliability of fault tolerance from three aspects: (1) Reduced Data Transfer and I/O: By asynchronously caching parameters, i.e., sharded model parameters, optimizer states, and RNG states, to CPU volatile memory, Our framework significantly reduces communication costs and bypasses checkpoint I/O. (2) Enhanced System Reliability: Our framework enhances parameter protection with a two-layer hierarchy: snapshot management processes (SMPs) safeguard against software failures, together with Erasure Coding (EC) protecting against node failures. This double-layered protection greatly improves the survival probability of the parameters compared to existing checkpointing methods. (3) Improved Snapshotting Frequency: Our framework achieves more frequent snapshotting compared with asynchronous checkpointing optimizations under the same saving time budget, which improves the fault tolerance efficiency. Empirical results demonstrate that Our framework minimizes the overhead of fault tolerance of LLM pretraining by effectively leveraging redundant CPU resources.

Joint-Embedding Predictive Architectures (JEPAs) underpin a growing family of latent world models for control from raw pixels, but every existing JEPA world model commits at training time to a single inference paradigm: either trajectory optimisation in a learned dynamics model, or direct behaviour cloning. A single checkpoint that serves both would defer this choice to inference, when deployment constraints (rollout cost, observation accessibility) determine which path wins. We present Qantara, an end-to-end JEPA whose joint training objective pairs a Brownian-bridge interpolant between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. The same checkpoint serves three inference paradigms without retraining: latent planning, behaviour-cloning action sampling, and inverse dynamics, which we query through a video-inverse composition that first predicts the next latent without action conditioning, then extracts the action. Training concentrates mass on the edges of the (action-time, state-time) noise square, where inference queries the predictor: replacing it with uniform interior sampling drops Push-T planning from 90.1 to 53.3 SR at matched compute. On the LeWM control suite, Qantara reaches a 91.2 SR three-train-seed average and sets new SOTA on OGBench-Cube (+7.7 SR over DINO-WM, +19.7 over LeWM). From the same weights, the behaviour-cloning and video-inverse paths reach 82-83 SR on Push-T and 71-73 SR on Cube. These results move JEPA world models from single-paradigm planners to multi-paradigm controllers.

Reinforcement learning post-training drives reasoning and agentic capabilities in modern AI systems, yet a growing body of work shows that it is most effective when used to fine-tune an already capable base model. We question whether existing pipelines yield models that are most suitable for reinforcement learning. Building on prior work highlighting the role of coverage and pass@K as predictors of post-RL performance, we design a simple modification to supervised fine-tuning, TailSFT, which filters out already fit sequences during training, thereby focusing learning on under-modeled regions, or the tail, of the data distribution. We justify and validate the design choices in TailSFT, particularly the specific filtering criteria, through a combination of controlled experiments and theoretical analysis. On OLMo-3 7B, TailSFT often improves pass@16 performance on math and coding evaluations, with gains up to 17% absolute, while incurring minimal computational overhead. These higher-coverage checkpoints consistently translate to up to 4% absolute pass@1 gains in subsequent GRPO runs, demonstrating that TailSFT checkpoints serve as better initializations for RL. We further introduce a lightweight diagnostic for identifying settings where TailSFT is most likely to help. More broadly, our results motivate a principled, stage-aware approach to model development, in which intermediate checkpoints are judged by how effectively they support subsequent training.

Training Large Language Models (LLMs) incurs significant cost; hence, any strategy that accelerates model convergence is helpful. In this paper, we investigate the ability of a simple idea checkpoint averaging along the trajectory of a training run to improve both convergence and generalization quite early on during training. Here we show that models trained with high learning rates observe higher gains due to checkpoint averaging. Furthermore, these gains are amplified when checkpoints are sampled with considerable spacing in training steps. Our training recipe outperforms conventional training and popular checkpoint averaging baselines such as exponential moving average (EMA) and stochastic moving average (SWA). We evaluate our training recipe by pre-training LLMs, where high learning rates are inherently preferred due to extremely large batch sizes. Specifically, we pre-trained nanoGPT-2 models of varying sizes, small (125M), medium (335M), and large (770M)on the OpenWebText dataset, comprised of 9B tokens. Additionally, we present results for publicly available Pythia LLMs, ranging from 1B to 12B, which were trained on the PILE-deduped dataset containing 207B tokens.

Computer-use agents provide a promising path toward general software automation because they can interact directly with arbitrary graphical user interfaces instead of relying on brittle, application-specific integrations. Despite recent advances in benchmark performance, strong computer-use agents remain expensive and slow in practice, since most systems invoke large multimodal models at nearly every interaction step. We argue that this uniform allocation of compute is fundamentally inefficient for long-horizon GUI tasks. Such trajectories are highly heterogeneous: many steps are routine and can be handled reliably by smaller, cheaper policies, while errors tend to concentrate at a relatively small number of high-risk moments. Across computer-use benchmarks, these failures repeatedly take two forms: progress stalls, where the agent loops, repeats ineffective actions, or fails to make meaningful progress, and silent semantic drift, where the agent continues taking locally plausible actions after already deviating from the user's true goal. To address this inefficiency, we propose an event-driven, step-level cascade for computer-use agents that runs a small policy by default and escalates to a stronger model only when lightweight learned monitors detect elevated risk. Our framework combines two complementary signals: a Stuck Monitor that detects degraded progress from recent reasoning-action history and triggers recovery, and a Milestone Monitor that identifies semantically meaningful checkpoints where sparse verification is most informative for catching drift. This design turns always-on frontier-model inference into adaptive, on-demand compute allocation over the course of an evolving interaction. The framework is modular and deployment-oriented: it can be layered on top of existing computer-use agents without changing the underlying agent architecture or retraining the large model.

Reinforcement learning with verifiable rewards (RLVR) has become a core post-training recipe. Introducing suitable off-policy trajectories into on-policy exploration accelerates RLVR convergence and raises the performance ceiling, yet finding a source of such trajectories remains the key challenge. Existing mixed-policy methods either import trajectories from external teachers (high-quality but distributionally far) or replay past training trajectories (close but capped in quality), and neither simultaneously satisfies the strong enough (higher Q , more new knowledge to learn) and close enough (lower V , more readily absorbed) conditions required to maximize the effective learning signal S = Q/V. We propose Near-Future Policy Optimization (NPO), a simple mixed-policy scheme that learns from a policy's own near-future self: a later checkpoint from the same training run is a natural source of auxiliary trajectories that is both stronger than the current policy and closer than any external source, directly balancing trajectory quality against variance cost. We validate NPO through two manual interventions, early-stage bootstrapping and late-stage plateau breakthrough, and further propose AutoNPO,an adaptive variant that automatically triggers interventions from online training signals and selects the guide checkpoint that maximizes S. On Qwen3-VL-8B-Instruct with GRPO, NPO improves average performance from 57.88 to 62.84, and AutoNPO pushes it to 63.15, raising the final performance ceiling while accelerating convergence.

Saving, or checkpointing, intermediate results during interactive data exploration can potentially boost user productivity. However, existing studies on this topic are limited, as they primarily rely on small-scale experiments with human participants - a fundamental constraint of human subject studies. To address this limitation, we employ AI agents to simulate a large number of complex data exploration scenarios, including revisiting past states and branching into new exploration paths. This strategy enables us to accurately assess the impact of checkpointing while closely mimicking the behavior of real-world data practitioners. Our evaluation results, involving more than 1,000 exploration paths and 2,848 executed code blocks, show that a checkpointing framework for computational notebooks can indeed enhance productivity by minimizing unnecessary code re-executions and redundant variables or code.

In real-world applications, guardrails are often expected to identify unsafe user-model interactions according to application-specific safety policies, rather than relying on predefined risk taxonomies. In this work, we study this setting under the paradigm of in-context policy guardrailing, where guardrails predict safety violations based on policy specifications provided in context. To systematically evaluate this capability, we introduce SafePyramid, a safety benchmark comprising 1,000 multi-turn conversations across 10 domains and 3,000 corresponding application-specific policies, which together contain 61,699 distinct natural-language rules. SafePyramid organizes the evaluation into three difficulty levels: L0 evaluates individual-rule understanding, L1 evaluates reasoning over rule dependencies, and L2 evaluates adaptation of full novel policy frameworks defined in context. To ensure benchmark quality, we employ a rigorous multi-stage pipeline to construct and validate the benchmark. Using SafePyramid, we evaluate 10 frontier LLMs and 5 policy-configurable guardrails and find that in-context policy guardrailing remains highly challenging: even the best-performing model, GPT-5.5, exactly identifies the full set of violated rules in only 54.0%, 35.3%, and 12.9% cases on L0, L1, and L2, respectively. These results highlight the limitations of current guardrails and call for stronger in-context policy guardrails that can reliably execute policies, resolve rule dependencies, and adapt to novel policy frameworks.

Can reinforcement learning with hard, verifiable rewards teach a compact language model to reason about physics, or does it primarily learn to pattern-match toward correct answers? We study this question by training a 1.5B-parameter reasoning model on beam statics, a classic engineering problem, using parameter-efficient RLVR with binary correctness rewards from symbolic solvers, without teacher-generated reasoning traces. The best BeamPERL checkpoint achieves a 66.7% improvement in Pass@1 over the base model. However, the learned competence is anisotropic: the model generalizes compositionally (more loads) but fails under topological shifts (moved supports) that require the same equilibrium equations. Intermediate checkpoints yield the strongest reasoning, while continued optimization degrades robustness while maintaining reward. These findings reveal a key limitation of outcome-level alignment: reinforcement learning with exact physics rewards induces procedural solution templates rather than internalization of governing equations. The precision of the reward signal - even when analytically exact - does not by itself guarantee transferable physical reasoning. Our results suggest that verifiable rewards may need to be paired with structured reasoning scaffolding to move beyond template matching toward robust scientific reasoning.

Checkpoint merging is a technique for combining multiple model snapshots into a single superior model, potentially reducing training time for large language models. This paper explores checkpoint merging in the context of parameter-efficient fine-tuning (PEFT), where only small adapter modules (e.g. LoRA) are trained. We propose Metrics-Weighted Averaging (MWA), a simple yet effective method to merge model checkpoints by weighting their parameters according to performance metrics. In particular, we investigate weighting by training loss and by training steps, under the intuition that lower-loss or later-step checkpoints are more valuable. We introduce a formula with a penalty factor to adjust weight distribution, requiring only one hyperparameter regardless of the number of checkpoints. Experiments on three fine-tuning tasks (mathematical reasoning, preference alignment, and general instruction tuning) show that MWA consistently produces merged models that outperform the naive uniform average of checkpoints. Notably, loss-weighted merging often yields the best results, delivering up to 5% higher task accuracy than the baseline uniform merge and even surpassing the final individual checkpoint's performance. These findings validate checkpoint merging for PEFT and demonstrate that a metric-driven weighting heuristic can efficiently boost model performance with minimal computational overhead.

LLM agents can commit durable effects from authority evidence that was valid earlier in execution: a DOM snapshot, approval epoch, version witness, branch token, or worker result. We study the commit boundary at which earlier authority evidence no longer authorizes a durable effect. We call this property commit-time authorization: a durable effect is authorized only if the witness that licensed its derived state remains fresh, causally prior, bound to the same effect, and eligible at commit time. We build a controlled-invalidation suite spanning browser, tool/API, and multi-agent workflows. The suite preserves the user goal and payload shape while invalidating the authority relation before durability. In the primary 54-task matrix, endpoint success remains high: 262/270 runs reach the visible result. Only 55/270 are authorized completions; among the 216 invalidating rows, 207 commit after the authorizing path has failed. All 54 clean controls remain authorized, and a separate 54-run authority-preserving check produces no unauthorized commits. We then evaluate mitigation families. Prompt caution and single-condition checks are insufficient because different hazards break different boundary conditions. Defenses work when they refresh, rebind, replan, or refuse at the durability boundary. CommitGuard, a fail-closed boundary monitor, blocks stale durable-effect attempts on protected commit surfaces when runtimes emit witness, dependency, binding, and eligibility signals. The result is a reporting and runtime-design lesson: endpoint success is a utility metric; authorized commit is a security property.

One-shot Program-of-Thought (PoT) emits a Python program that prints a primitive-action plan; a single invalid action silently invalidates the trajectory. We introduce RePoT (Recoverable PoT): a deterministic verified replay that walks the plan through the environment to its first invalid transition, then one LLM call that resumes from the verified prefix. RePoT costs at most one extra LLM call on the ~14% of problems where PoT fails. RePoT beats PoT by +3 to +11pp across four closed-model configurations on PuzzleZoo-775 and peaks at 96.9% vs 86.3% on gpt-5.4-mini-medium; against the matched-budget PoT-retry baseline, RePoT wins decisively on Gemini (+3.8pp, 95% CI [+2.2,+5.4]), is within sampling noise on GPT-medium and Claude, and loses on GPT-mini -- a capability-scaling pattern we begin to address with Adaptive RePoT, a rule-based dispatcher that routes between suffix repair and a fresh PoT retry based on verified-prefix length (preliminary). We replicate on PlanBench Blocksworld (+1.1 to +11.4pp) and on four open-weights models (+3.3 to +20.0pp on three of four). On Derail-550, our controlled recovery benchmark, every condition with access to checkpoint information clears >=30% on GPT-medium and >=70% on Gemini, vs <=3.1% for error-only feedback -- showing that checkpoint information, not the specific verified-prefix tail, is the load-bearing recovery signal.

LLM agents act on real-world environments through tool calls, and a single misjudged action can cause irreversible harm. The standard safeguard is a guard model that labels each proposed action as safe or unsafe, but this binary view conflates two distinct decisions: whether the action is harmful in itself, and whether it is appropriate given the user's context. It also operates at the granularity of action categories rather than individual instances, producing routine interruptions that erode autonomy and train users to wave through the most consequential alerts. We reframe the problem as a per-instance three-way routing decision over {EXECUTE, ASK, REFUSE} and instantiate it with Safety Sentry, a lightweight guard model whose inference reduces to a single decoding call. A single decoding-time threshold lets one fixed checkpoint be re-positioned across deployments of differing risk tolerance without retraining. Safety Sentry outperforms a broad set of open-weight and frontier closed-source baselines on overall accuracy and safety-related recall, while controlling both directional error rates simultaneously.

Training capable cyber agents is often treated primarily as a problem of model scale, yet open-weight post-training is constrained more directly by the cost of executable environments, reliable multi-turn supervision, and access to strong teachers. We present a data-centric framework that addresses these bottlenecks through five complementary systems: Choulea analyzes hidden reasoning signatures, SkyReal reduces teacher-sampling cost, Hongzwang bypasses API restrictions on teacher execution, PSBreakup restores capabilities weakened by model merging, and Kreator converts expert interventions into trainable reasoning. Our data engine constructs resettable coding, vulnerability, CTF, kernel-history, full-exploit, firmware, and device-backed environments. Candidate trajectories are retained only after execution verification and evidence auditing, yielding 164,269 trajectories for long-context supervised fine-tuning. The three checkpoints improve over their starting models by an average of 23.76% on the full CyberGym suite and 10.49% across the pooled CTF suites. As of September 1, 2026, Feyospace-s1 achieves a verified success rate of 63.24% and ranks 10th on the official CyberGym leaderboard, while all three checkpoints rank 1st among models at comparable parameter scales. To our knowledge, this is the first end-to-end demonstration that a seven-person independent team can train open-weight models with leading agentic cyber capability.

Reliable evaluation is fundamental to the progress of Large Language Models (LLMs), yet the evaluation process during pre-training is plagued by significant instability that obscures true learning dynamics. In this work, we systematically diagnose this instability, attributing it to two distinct sources: Parameter Instability from training stochasticity and Evaluation Instability from noisy measurement protocols. To counteract both sources of noise, we introduce MaP, a dual-pronged framework that synergistically integrates checkpoint Merging and the Pass@k metric. Checkpoint merging smooths the parameter space by averaging recent model weights, while Pass@k provides a robust, low-variance statistical estimate of model capability. Extensive experiments show that MaP yields significantly smoother performance curves, reduces inter-run variance, and ensures more consistent model rankings. Ultimately, MaP provides a more reliable and faithful lens for observing LLM training dynamics, laying a crucial empirical foundation for LLM research.

Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but how these gains relate to inference-time decoding and search remains unclear. Does RL create reasoning the base model lacks, or shift the rollout distribution toward trajectories it can already reach but rarely samples? We study this behaviorally with a Unified Decoding Framework (UDF), which expresses token-level sampling, beam-like search, tree search, and sequence-level resampling as executable policies over a shared budgeted operating space, scored post hoc with pass@k, self-consistency, best-of-N, and first-finish success. Using paired Base/RL checkpoints from SimpleRL-Zoo, we ask whether an RL default-policy curve can be approximated by a structured path of Base operating points. On Math500, AIME, GPQA, and IFEval, the pass@k recovery path follows a Budgeted Operating-Point Transition Rule (BOPTR), N_{Base} approx αN_{RL}^β, with benchmark-conditioned exponents. On Qwen2.5-7B, BOPTR gives the lowest transfer error among the non-oracle rules we test, 3.41 pp (95% CI [2.32, 5.53]); a three-seed replication gives 3.07 pm 0.39 pp. The rule extends to ten models across four families (3.28 to 4.87 pp on checkpoints added after fitting), to four benchmarks it was never fitted on (5.03 pp vs. 4.44 pp in fit), and holds without an RL checkpoint for the target model (4.19 pp) or without RL supervision of any kind (5.08 pp). These results support a qualified internalized-search reading: under the recipe we test, much of the measured RL gain corresponds to a change in sampling efficiency toward operating points the base model can already reach under search. We treat the scaling patterns as descriptive of this recipe and cohort, report where they break down, and use UDF and BOPTR as behavioral diagnostics rather than evidence of parameter-level equivalence.

Real-world evaluation is the gold standard for robot policies because it tests them against the physical conditions and deployment challenges they are ultimately designed to handle. However, real-world evaluation is also the bottleneck for iterating on robot policies: it is costly, difficult to reproduce, and often too sparse to reliably compare nearby model variants. A straightforward proxy for performance is validation loss on expert demonstrations, but this proxy is often poorly correlated with real-world performance. In this paper, we introduce Critical Interval MSE (CI-MSE), an intuitively simple yet effective offline validation metric. CI-MSE restricts error computation to task-critical segments and pairs it with simple action-alignment procedures that better match rollout-time behavior. Across simulation and real-world experiments, CI-MSE yields a stronger correlation between validation error and rollout performance than raw MSE. Across a wide range of policy checkpoints, CI-MSE achieves a Spearman's rank correlation of -0.87, much closer to the ideal value of -1 than raw MSE's -0.61, demonstrating a significant improvement. We show through sensitivity analysis that our metric is robust to a wide range of hyperparameters. We further study the effectiveness of CI-MSE under evaluation distribution shifts and suggest design boundaries when using this metric. In summary, this paper provides a simple and reliable offline validation tool for accelerating policy iteration. Project webpage: https://ci-mse.github.io/

Small Language Models (SLMs) are increasingly being considered for deployment on edge devices such as laptops, enabling private, low-latency, and locally personalized applications. However, personalization requires models to adapt over time to evolving user- or task-specific data, placing them in a continual learning setting. This creates the risk of catastrophic forgetting, where learning new information degrades performance on previously learned tasks or broader model capabilities. Recent benchmarks such as TRACE have shown that continual fine-tuning can significantly degrade the general abilities of aligned large language models. In this work, we present a study for sequential LoRA personalization of SLMs. We save model checkpoints after each adaptation stage and evaluate them on current tasks, previously seen tasks, and a fixed reference set. This checkpoint-level protocol enables us to monitor task performance, forgetting, and reference set drift over time. We show that lightweight reference set distributional diagnostics can reveal model-specific instability patterns during sequential LoRA personalization of SLMs, including cases where task-level metrics alone hide harmful adaptation. We hope this can highlight new research avenues for monitoring stability of SLMs in a continual learning setting.

Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict brittleness: among models above 90% baseline accuracy, brittleness rates range from 13.7% to 90.0%. Second, failures stem from policy override rather than miscomprehension: despite acknowledging the context change in every case, models persist via three distinct mechanisms that vary by update type and model family. Third, on a hand-audited probe of catastrophic consequence-flip scenarios, standard action-level guardrails catch none, while a state-aware validator catches all without false alarms on correct interventions. This indicates that action-level content moderation is systematically blind to consequence-flips, motivating state-aware architectural alternatives. We release our protocol, perturbed benchmarks, and deployment probe.

As autonomous AI agents are used in regulated and safety-critical settings, organizations need effective ways to turn policy into enforceable controls. We introduce a regulatory machine learning framework that converts unstructured design artifacts (like PRDs, TDDs, and code) into verifiable runtime guardrails. Our Policy as Prompt method reads these documents and risk controls to build a source-linked policy tree. This tree is then compiled into lightweight, prompt-based classifiers for real-time runtime monitoring. The system is built to enforce least privilege and data minimization. For conformity assessment, it provides complete provenance, traceability, and audit logging, all integrated with a human-in-the-loop review process. Evaluations show our system reduces prompt-injection risk, blocks out-of-scope requests, and limits toxic outputs. It also generates auditable rationales aligned with AI governance frameworks. By treating policies as executable prompts (a policy-as-code for agents), this approach enables secure-by-design deployment, continuous compliance, and scalable AI safety and AI security assurance for regulatable ML.

Imitation learning has enabled robots to acquire complex visuomotor manipulation skills from demonstrations, but deployment failures remain a major obstacle, especially for long-horizon action-chunked policies. Once execution drifts off the demonstration manifold, these policies often continue producing locally plausible actions without recovering from the failure. Existing runtime monitors either require failure data, over-trigger under benign feature drift, or stop at failure detection without providing a recovery mechanism. We present Rewind-IL, a training-free online safeguard framework for generative action-chunked imitation policies. Rewind-IL combines a zero-shot failure detector based on Temporal Inter-chunk Discrepancy Estimate (TIDE), calibrated with split conformal prediction, with a state-respawning mechanism that returns the robot to a semantically verified safe intermediate state. Offline, a vision-language model identifies recovery checkpoints in demonstrations, and the frozen policy encoder is used to construct a compact checkpoint feature database. Online, Rewind-IL monitors self-consistency in overlapping action chunks, tracks similarity to the checkpoint library, and, upon failure, rewinds execution to the latest verified safe state before restarting inference from a clean policy state. Experiments on real-world and simulated long-horizon manipulation tasks, including transfer to flow-matching action-chunked policies, demonstrate that policy-internal consistency coupled with semantically grounded respawning offers a practical route to improved reliability in imitation learning. Supplemental materials are available at https://sjay05.github.io/rewind-il

Reinforcement learning pipelines for Large Language Model (LLM) training often rely on manually redesigned environments between stages, requiring practitioners to heuristically infer which configuration will best improve the current policy. To automate this process, we propose the LLM-as-Environment-Engineer framework in which the current policy model analyzes failure trajectories together with contextual information and proposes modifications to the next-stage training environment configuration. We also introduce MAPF-FrozenLake, a controllable testbed whose generator exposes multi-dimensional environment configurations, making it suitable for studying and benchmarking environment redesign. On this testbed, we condition the environment engineer on structured summaries of policy behavior, failure cases, and environment statistics, from which it produces the configuration for the next training stage. With Qwen3-4B as the backbone, our framework achieves the strongest aggregate performance on our benchmarks, outperforming larger proprietary LLMs (e.g., GPT, Gemini) and fixed-environment training baselines. We further analyze which forms of context are most effective, finding that successful environment updates rely on failure evidence and preserve configurations that already work. Interestingly, the current RL checkpoint serves as a better environment engineer than the original base model, suggesting that policy learning improves the model's ability to diagnose its remaining weaknesses.

The increasing complexity of enterprise business scenarios has promoted the widespread adoption of long SKILL documents in agent systems, posing new challenges for compliance detection: large models incur substantial inference costs, while small models may fail to maintain detection accuracy. To address this gap, we propose SkillCDG, a graph-based framework for long SKILL compliance detection. SkillCDG represents complex business policies as a two-layer constraint dependency graph, where the upper layer indexes SKILL descriptions for scenario routing and the lower layer captures dependencies among atomic constraints within each SKILL. During inference, two-level retrieval followed by dependency closure supports compliance judgment and source traceability. We comprehensively evaluate the framework on three enterprise datasets and two controlled public benchmark variants. Experimental results demonstrate that SkillCDG outperforms baseline methods by up to 12.8 percentage points in detection F1 score, while reducing token consumption by a maximum 64.3\%. Moreover, we further investigate the inherent relationships among policy-graph complexity, model scale, and detection performance. Comparative experiments conducted on four checkpoints from a single model family validate a concise and effective scaling trend: end-to-end detection correctness exhibits a complexity-differentiated scaling pattern, and the complexity metric derived from the constraint dependency graph can effectively quantify instance difficulty and the performance improvement potential of models. Leveraging this insightful scaling trend, we conduct adaptive training sample selection and adopt on-policy distillation to efficiently enhance the compliance detection capability of small-scale models.

Instruction fine-tuning pretrained LLMs for diverse downstream tasks has demonstrated remarkable success and has captured the interest of both academics and practitioners. To ensure such fine-tuned LLMs align with human preferences, techniques such as RLHF and DPO have emerged. At the same time, there is increasing interest in smaller parameter counts for models. In this work, using OpenLLaMA 3Bv2 as a base model, we describe the recipe used to fine-tune the OpenBezoar family of models. In this recipe: We first generate synthetic instruction fine-tuning data using an open and commercially non-restrictive instruction fine-tuned variant of the Falcon-40B model under three schemes based on: LaMini-LM, WizardLM/Evol-Instruct (with databricks-dolly-15k as a seed dataset) and Orca (with the Flan Collection as a seed dataset), then filter these generations using GPT-4 as a human proxy. We then perform cost-effective QLoRA-based supervised fine-tuning sequentially with each scheme. The resulting checkpoint is further fine-tuned with a subset of the HH-RLHF dataset to minimize distribution shift prior to using the DPO loss to obtain the final checkpoint. Evaluation is done with the LM Eval Harness tasks/metrics as well as on MT-Bench using the "LLM-as-a-judge" framework with Claude 2.1, with the finding that the final checkpoint, "OpenBezoar-HH-RLHF-DPO", demonstrates superior performance over many models at the 3B parameter scale, even outperforming the top model in one of the categories on the Huggingface Open LLM Leaderboard. We release "OpenBezoar-SFT", "OpenBezoar-HH-RLHF-SFT", "OpenBezoar-HH-RLHF-DPO" checkpoints, alongside our generated datasets on HuggingFace at https://huggingface.co/collections/SurgeGlobal/open-bezoar-6620a24923e12127e9e2b9cc and our codebase at https://bitbucket.org/paladinanalytics/workspace/projects/OP.

We propose RegLLM, a diagnostic harness for bounded autonomy in regulated agentic workflows. It instruments six trustworthiness signals: citation validity, source grounding, schema compliance, escalation correctness, constitutional alignment, and unsafe-action rate. Signals are distinguished by their source of supervision: programmatic verifiers, task-level escalation labels, or AI-judge scores. A deterministic runtime supervisor blocks ungrounded answers and forces escalation, logging interventions. The same domain constitution informs evaluation, training rewards, and serving guardrails. Task-level should-escalate labels make the act-versus-defer decision a measurable training signal. We demonstrate the harness at smoke scale. An offline reference run (n=12) lifts escalation recall from 0 to 0.67 and reduces unsafe-action rate from 0.33 to 0.08 when governance is enabled. Two single-GPU Qwen2.5-3B LoRA/DPO pilots (n=8, same seed and evaluation split) expose substantial variation: nominally identical RL-base configurations yield task success of 0.25 versus 0.12 and escalation recall of 1.0 versus 0.5. An answer-quality adapter changes recall from 1.0 to 0.5 in Run A, but from 0.5 to 1.0 in Run B. An escalation-aware variant produces no measurable change in Run B. These small pilots do not establish reliable adapter effects or production readiness. Their contribution is diagnostic: configuration variance can overwhelm apparent tuning effects on bounded-autonomy metrics, motivating larger evaluation sets and repeated runs.

Long-running LLM agents keep valuable state resident on GPUs: KV caches, request schedulers, communication state, and sometimes online adapters. Losing this state after a GPU or communicator failure can discard minutes to hours of work, yet existing recovery mechanisms either restart the whole serving stack or require application-specific checkpoint logic inside every attention and runtime component. This paper argues that fault tolerance for such workloads needs a GPU-resident execution context: checkpoint hooks must run at device synchronization points, observe binary kernels that frameworks and libraries actually execute, and recover without putting the host CPU on the critical path. We present Concordia, a runtime that uses a device-resident persistent kernel as the substrate for fault-tolerant LLM inference. Concordia interposes on GPU module loading and supports PTX- and SASS-level instrumentation, allowing checkpoint and pause hooks to be inserted below framework code and library boundaries. For each registered LLM state region, Concordia JIT-compiles a specialized delta-checkpoint handler -- for example, a KV-block scanner, adapter-page scanner, or recovery applier -- and hot-swaps it into the persistent kernel's operator table. The persistent kernel consumes a lock-free ring buffer of compute, checkpoint, append-log, and recovery tasks, so the same always-on executor triggers dirty-page detection, stages deltas, and appends committed records to a CPU-visible log in CXL memory or host DRAM.

We study how model post-training and test-time inference design affect natural-language proof generation for hard olympiad mathematics. Starting from Nemotron 3 Ultra, we train two specialist checkpoints using supervised fine-tuning and reinforcement learning, and evaluate checkpoint choice, verification, and refinement. Based on these findings, we present an open-model test-time-compute pipeline. The system operates entirely in natural language, with no formal prover, external tools, or internet access. Three Nemotron 3 Ultra checkpoints - the general-availability model and two post-trained specialists - power an iterative search that generates, verifies, and refines candidate proofs; a separate high-compute stage then selects each final submission. The system scored 30 out of 42 points at IMO 2026, reaching the gold-medal threshold. We release the two post-trained checkpoints as well as the training data, the training and inference code, the submitted solutions, and Nemotron-IMO-Bench, a new benchmark of 200 novel olympiad-level problems.

Safety alignment is usually posed as a topic-level question: is this subject harmful? Deployments ask a narrower one. A civics tutor and a public-sector assistant may share a base model yet need different boundaries inside the same topic, refusing targeted political manipulation while still answering factual questions about the same election. We formulate this as narrow-boundary safety and introduce an offline self-generated framework combining controlled topic generation, coverage repair, in-distribution compensation data, and harmful-benign pairs for training and evaluation. Single-shot generation leaves 19.88% of prompts without accepted refusal traces, whereas escalating retries leave 0.20%. On political persuasion with Qwen3-8B, training on refusal data completed through Escalate increases target-domain refusal from 9.47% to 84.75% and reduces the mean unsafe-response rate across three broader harmfulness benchmarks from 26.26% to 0.14%, but increases XSTest over-refusal from 2.00% to 74.00%. In a separate matched comparison, replacing external responses with verified target-model responses reduces over-refusal from 15.20% to 5.20%. Boundary-pair data reduces comply-side over-refusal on held-out pairs from 32.94% to 4.16%, while harmful-side refusal decreases only from 91.88% to 87.72%. These results show that data composition controls the safety and usability trade-off, and that safety alignment should be evaluated on both sides of the intended refusal boundary.

Long-horizon LLM agents can fail quietly: they settle on one reading of the evidence early, then spend the rest of the run defending it. We call this premature commitment. Final-answer scoring misses the failure mode because it sees only the answer, not whether the process has already collapsed to a stable path. We define representational commitment as cross-run hidden-state convergence at a fixed reasoning step, and use it as an early diagnostic of trajectory consistency. On Llama-3.1-70B running ReAct on HotpotQA, step-4 hidden-state similarity predicts downstream behavioral consistency (r = -0.35, partial r = -0.45), with a localized temporal and layer-wise signature. The signal replicates across Qwen-2.5-72B and Phi-3-14B, and on StrategyQA (r = -0.83). It does not track correctness: committed-wrong and committed-correct questions are not separable in activation similarity. That boundary is central to the claim. Commitment tells us whether an agent has settled, not whether it is right. A runtime monitor detects inconsistent trajectories from hidden states at AUROC up to 0.97 (0.85--0.88 under a stricter split), and a prompting intervention cuts behavioral variance by 28% against a token-matched control while leaving accuracy statistically unchanged. We also test whether the signal can route self-consistency compute; on a harder benchmark it helps only modestly and is matched by a simpler output-based baseline. The result is a diagnostic for a hidden process failure, with clear limits rather than a general accuracy lever.

Large Reasoning Models (LRMs) generate intermediate reasoning traces that may contain unsafe content, even when their final responses appear safe. Guardrail models are designed to detect and block unsafe content, yet existing benchmarks for unsafe content detection focus primarily on prompts and final responses, leaving reasoning traces largely unexamined. Moreover, these benchmarks typically provide only binary safety labels, without evidence annotations that justify the judgments. To address these limitations, we introduce TRACE, an evidence-grounded safety evaluation benchmark that covers the entire LRM inference pipeline: prompts, reasoning traces, and final responses. TRACE includes prompts in two languages spanning nine risk categories and ten attack strategies. For each prompt, four LRMs generate reasoning traces and final responses, and we annotate the safety of each component and extract supporting evidence from the corresponding source text. Evaluating 18 guardrail models on TRACE reveals that safety judgment for reasoning traces is substantially more challenging than for prompts or final responses, and that current models struggle to accurately extract supporting evidence. These findings highlight the need for guardrail models that can reliably detect and precisely localize unsafe content across the LRM inference pipeline.

Diffusion Models (DM) and Consistency Models (CM) are two types of popular generative models with good generation quality on various tasks. When training DM and CM, intermediate weight checkpoints are not fully utilized and only the last converged checkpoint is used. In this work, we find that high-quality model weights often lie in a basin which cannot be reached by SGD but can be obtained by proper checkpoint averaging. Based on these observations, we propose LCSC, a simple but effective and efficient method to enhance the performance of DM and CM, by combining checkpoints along the training trajectory with coefficients deduced from evolutionary search. We demonstrate the value of LCSC through two use cases: (a) Reducing training cost. With LCSC, we only need to train DM/CM with fewer number of iterations and/or lower batch sizes to obtain comparable sample quality with the fully trained model. For example, LCSC achieves considerable training speedups for CM (23times on CIFAR-10 and 15times on ImageNet-64). (b) Enhancing pre-trained models. Assuming full training is already done, LCSC can further improve the generation quality or speed of the final converged models. For example, LCSC achieves better performance using 1 number of function evaluation (NFE) than the base model with 2 NFE on consistency distillation, and decreases the NFE of DM from 15 to 9 while maintaining the generation quality on CIFAR-10. Our code is available at https://github.com/imagination-research/LCSC.

As large language models (LLMs) become high-privilege agents in risk-sensitive settings, they introduce systemic threats beyond hallucination, where minor compliance errors can cause critical data leaks. However, existing benchmarks focus on rule-based QA, lacking agentic execution modeling, overlooking compliance drift in adversarial interactions, and relying on binary safety metrics that fail to capture behavioral degradation. To bridge these gaps, we present CNFinBench, a comprehensive benchmark spanning 29 subtasks grounded in the triad of expertise, autonomy, and integrity. It assesses domain-specific capabilities through certified regulatory corpora and professional financial tasks, reconstructs end-to-end agent workflows from requirement parsing to tool verification, and simulates multi-turn adversarial attacks that induce behavioral compliance drift. To quantify safety degradation, we introduce the Harmful Instruction Compliance Score (HICS), a multi-dimensional safety metric that integrates risk-type-specific deductions, multi-turn consistency tracking, and severity-adjusted penalty scaling based on fine-grained violation triggers. Evaluations over 22 open-/closed-source models reveal: LLMs perform well in applied tasks yet lack robust rule understanding, suffer a 15.4-point drop single modules to full execution chains, and collapse rapidly in multi-turn attacks, with average violations surging by 172.3% in Round 2. CNFinBench is available at https://cnfinbench.opencompass.org.cn and https://github.com/VertiAIBench/CNFinBench.

We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL approaches for driving use complex shaped rewards that sum multiple individual rewards, \eg~progress, position, or orientation rewards. We show that PPO fails to optimize a popular version of these rewards when the mini-batch size is increased, which limits the scalability of these approaches. Instead, we propose a new reward design based primarily on optimizing a single intuitive reward term: route completion. Infractions are penalized by terminating the episode or multiplicatively reducing route completion. We find that PPO scales well with higher mini-batch sizes when trained with our simple reward, even improving performance. Training with large mini-batch sizes enables efficient scaling via distributed data parallelism. We scale PPO to 300M samples in CARLA and 500M samples in nuPlan with a single 8-GPU node. The resulting model achieves 64 DS on the CARLA longest6 v2 benchmark, outperforming other RL methods with more complex rewards by a large margin. Requiring only minimal adaptations from its use in CARLA, the same method is the best learning-based approach on nuPlan. It scores 91.3 in non-reactive and 90.6 in reactive traffic on the Val14 benchmark while being an order of magnitude faster than prior work.

The standard LLM training pipeline applies reinforcement learning (RL) only after pre-training and supervised fine-tuning (SFT). We question this status quo by training a LLM from scratch and applying RL, SFT, and SFT followed by RL directly to intermediate pre-training checkpoints. We find that RL is effective very early, and often matches the full SFTtoRL pipeline early as well. Through experiments on harder problems, we find that targeted pre-training data composition is a strong lever for RL effectiveness, even more so than model scale. Beyond reasoning accuracy, applying RL directly to base checkpoints expands the model's distribution; the sharpening effect reported in recent work arises only when RL follows SFT. The general capabilities of the model remain essentially unchanged by RL, while they degrade following SFT. Finally, we merge RL and SFT objectives by parallel averaging, which outperforms across all other training methods discussed, across metrics, while preserving general capabilities. Together, these results suggest that LLM training might benefit from an expanded use of RL.

Large language models are increasingly used for content moderation, but most evaluations still report aggregate accuracy on individual benchmarks. We introduce Safety-Flag, which places seven widely used safety benchmarks (BeaverTails, XSTest, Ethics, WildGuard, Aegis, ToxiChat, and ToxiGen) into a single balanced flag / do-not-flag protocol. We release item-level decisions and confidence scores for six general-purpose LLMs and four dedicated guards, together with three reference models, evaluated on the same items. Safety-Flag measures three dimensions of moderator reliability: error direction, probability calibration, and confidence-based error ranking for human review. They often disagree. Aggregate accuracy does not reveal error direction: one model flags 85% of benign content, whereas another misses 54% of harmful content. All six general-purpose models are overconfident; fitting one temperature per model reduces calibration error by 2.8--6.0times without changing predicted labels or confidence ordering. Confidence-based abstention lowers selective risk for every model, although the gains depend on how well confidence ranks errors. Dedicated guards produce fewer false alarms and are better calibrated, but several have higher miss rates outside their documented coverage. We release the benchmark, fixed item lists, evaluation code, per-item model outputs, and leaderboard at: https://github.com/yibo-hu-lab/safety-flag-benchmark.

The risks of frontier AI may require international cooperation, which in turn may require verification: checking that all parties follow agreed-on rules. For instance, states might need to verify that powerful AI models are widely deployed only after their risks to international security have been evaluated and deemed manageable. However, research on AI verification could benefit from greater clarity and detail. To address this, this report provides an in-depth overview of AI verification, intended for both policy professionals and technical researchers. We present novel conceptual frameworks, detailed implementation options, and key R&D challenges. These draw on existing literature, expert interviews, and original analysis, all within the scope of confidentially overseeing AI development and deployment that uses thousands of high-end AI chips. We find that states could eventually verify compliance by using six largely independent verification approaches with substantial redundancy: (1) built-in security features in AI chips; (2-3) separate monitoring devices attached to AI chips; and (4-6) personnel-based mechanisms, such as whistleblower programs. While promising, these approaches require guardrails to protect against abuse and power concentration, and many of these technologies have yet to be built or stress-tested. To enable states to confidently verify compliance with rules on large-scale AI development and deployment, the R&D challenges we list need significant progress.

Reference-free faithfulness metrics verify each atomic claim a model makes against ground truth, and are increasingly used to evaluate grounded generation. We show they share a blind spot: they measure only precision -- are the stated claims supported? -- and therefore reward abstention, since a model can score near-perfect faithfulness by saying almost nothing. We make this measurable using Formula 1 telemetry, a domain where strategic ground truth is derived deterministically and, crucially, completely: for each decision we know the full set of facts that mattered. This completeness -- absent in open-domain faithfulness benchmarks -- lets us measure recall (coverage of the relevant facts) exactly, alongside precision. On a multilingual (EN/ES/PT) benchmark of 7,253 decision instances spanning 150 races, the most precise frontier model covers under half of the relevant facts and ranks last by F1, so requiring coverage reorders the systems; the same effect reappears in a second complete-oracle domain (NOAA weather forecasts). A prompt ablation shows the low coverage is not an under-prompting artifact: explicitly asking models to be thorough does not close the gap. We pair faithfulness with coverage into a single score, validate the metric (controlled perturbation; agreement across a model-free regex extractor and a cross-family LLM extractor, system-level Spearman 1.0), and give a verifier-guided generation method that improves precision and recall without references. We release the benchmark, structured annotations, metric, baselines, and an interactive demo.

Enterprise agents must follow long-horizon, conditional, safety-critical standard operating procedures (SOPs). We compile machine-readable SOP constraints into executable pseudo-code and run them with a program-guided (PG) stack machine that pages the active frame while an LLM performs semantic execution. A three-arm SOPBench study across six models separates representation from runtime: compiled text never significantly hurts and gains up to 16.0 points where official prose underperforms. Runtime guidance is capability-gated. Two strong models independently show positive seven-domain PG contrasts (58:19 and 75:31 discordant pairs), whereas weak models are harmed. A full-program cursor ablation (active frame first, complete program retained) recovers much of the strong-model refusal gain; selective visibility adds a smaller improvement. Paired probe and audit measurements track this divide to spontaneous state discipline rather than reconstruction ability. On Bank the three primary arms rise from 70.4 to 86.4 to 92.8, with 100% refusal correctness. Practical guidance: compile first; enable active-frame paging only after a model-level discipline check.

Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023), fine-tuning attacks (Qi et al., 2024), and activation-space edits (Arditi et al., 2024) keep recovering the behaviors it was meant to remove. We give this fragility one geometric explanation and follow it into pretraining. We measure the safety update Δ= W_{safe} - W_{base} against the curvature of the model's capabilities (the empirical Fisher of a capability loss). Across five model families, post-hoc safety lands in a suppression regime: Δ is nearly orthogonal to the capability directions, and its small in-subspace part concentrates on a few high-curvature ones. The update is thin but sharp, a refusal gate laid over intact capabilities rather than erasure of them. A kernel-immobility lemma explains why such an update can only mask a capability, not remove it, so a little benign fine-tuning restores it: 100 benign examples cut the AdvBench refusal of Qwen-2.5-7B-Instruct and Llama-3-8B-Instruct by 35 to 38 pp. Following the account into pretraining, a pretraining-checkpoint sweep of OLMo-2-1B (Team OLMo et al., 2024) shows the features that refusal attaches to emerging in a sharp transition between 1B and 63B pretraining tokens. We then use the account constructively: models trained from scratch with safety co-training spread continuously across pretraining reach 87 to 98% AdvBench refusal that the same attack erodes by only 2 to 14 pp at every scale from 410M to 6.9B, against 35 to 38 pp for post-hoc installs, at a small cost on short-answer capability probes; a windowed schedule of equal total safety weight installs no refusal. Persistence of the safety signal across pretraining, not its timing, is what buys attack robustness.

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a powerful paradigm for enhancing the complex reasoning capabilities of Large Reasoning Models. However, standard outcome-based supervision suffers from a critical limitation that penalizes trajectories that are largely correct but fail due to several missteps as heavily as completely erroneous ones. This coarse feedback signal causes the model to discard valuable largely correct rollouts, leading to a degradation in rollout diversity that prematurely narrows the exploration space. Process Reward Models have demonstrated efficacy in providing reliable step-wise verification for test-time scaling, naively integrating these signals into RLVR as dense rewards proves ineffective.Prior methods attempt to introduce off-policy guided whole-trajectory replacement that often outside the policy model's distribution, but still fail to utilize the largely correct rollouts generated by the model itself and thus do not effectively mitigate the narrowing of the exploration space. To address these issues, we propose SCOPE (Step-wise Correction for On-Policy Exploration), a novel framework that utilizes Process Reward Models to pinpoint the first erroneous step in suboptimal rollouts and applies fine-grained, step-wise off-policy rectification. By applying precise refinement on partially correct rollout, our method effectively salvages partially correct trajectories and increases diversity score by 13.5%, thereby sustaining a broad exploration space. Extensive experiments demonstrate that our approach establishes new state-of-the-art results, achieving an average accuracy of 46.6% on math reasoning and exhibiting robust generalization with 53.4% accuracy on out-of-distribution reasoning tasks.

Massively multilingual language models have displayed strong performance in zero-shot (ZS-XLT) and few-shot (FS-XLT) cross-lingual transfer setups, where models fine-tuned on task data in a source language are transferred without any or with only a few annotated instances to the target language(s). However, current work typically overestimates model performance as fine-tuned models are frequently evaluated at model checkpoints that generalize best to validation instances in the target languages. This effectively violates the main assumptions of "true" ZS-XLT and FS-XLT. Such XLT setups require robust methods that do not depend on labeled target language data for validation and model selection. In this work, aiming to improve the robustness of "true" ZS-XLT and FS-XLT, we propose a simple and effective method that averages different checkpoints (i.e., model snapshots) during task fine-tuning. We conduct exhaustive ZS-XLT and FS-XLT experiments across higher-level semantic tasks (NLI, extractive QA) and lower-level token classification tasks (NER, POS). The results indicate that averaging model checkpoints yields systematic and consistent performance gains across diverse target languages in all tasks. Importantly, it simultaneously substantially desensitizes XLT to varying hyperparameter choices in the absence of target language validation. We also show that checkpoint averaging benefits performance when further combined with run averaging (i.e., averaging the parameters of models fine-tuned over independent runs).

Reinforcement learning with verifiable rewards, and Group Relative Policy Optimization (GRPO) in particular, is now run routinely on a supervised checkpoint in the hope of producing a stronger agent. We ask whether it adds skill to a small language and vision-language model web agent at the 4B to 8B scale, or whether it mostly reshapes behavior the supervised model already has. Across a control grid of 18 runs that varies learning rate, KL weight, seed, initialization, and clipping, no configuration credibly improves the success rate of a strong supervised baseline on tasks the agent has largely mastered. On the text track, moderate to high learning rates make it credibly worse. The null holds under paired testing, 25 evaluation seeds, 6 training seeds, changes to the recipe, both text and Set-of-Marks screenshot observations, and scaling the backbone to 8B; the credible harm is a text-track finding and is only nominal under Set-of-Marks. To show that the null reflects the setting and not a broken pipeline, we run the identical harness, reward, and recipe on tasks whose reward is reachable by sampling, and there the success rate rises by 22 points with a paired interval that excludes zero. GRPO therefore helps only when there is headroom to climb, meaning the sampled policy already succeeds more often than the greedy one. We then explain the failure. A middle learning rate degrades the agent and a high one collapses it, and the two regimes form a double dissociation: grafting localizes the degrade regime to the attention and MLP blocks, while the collapse regime cannot be traced to any single group, and the embedding change that dominates the weight movement is causally inert. At 4B, effective rank in the late layers tracks capability in both directions; at 8B the two come apart. This coupling is specific to the smaller model, so we report it as scale-dependent.

Coding agents such as Claude Code and Codex close the software loop: a main agent manages the loop, subagents analyze and execute, tools do the work. We port this architecture to robot-policy improvement, where one difference dominates the design: robotic tools---trained policies, training pipelines, data collection---fail routinely, so a tool's quality must be measured, recorded at every call, and expired when the artifact behind it changes. AgenticRobotics is a backend-independent control plane in which an LLM controller drives disposable workers through durable train--evaluate--improve transactions: an immutable objective, controller-owned measurement, commit-keyed crash recovery, an evidence-graded skill library, and a tool registry with a standardized, recorded call surface. The title is an operational claim, not a selection claim: the operator can leave because promotion is evidence-gated, state is recoverable, and capability quality is derived from records---not because the loop picks better checkpoints than a human; on the one lineage we measured, it does not. The gates measurably buy false-promotion control (0.001 per run hardened versus 0.005--0.021 shipped), anytime-valid decisions under optional stopping, zero lost or duplicate effects under kill injection, and six of six artifact-tampering classes caught by a signed verifier.

AI research agents combine prior knowledge, public sources, and experimental feedback to produce useful results. The Discovery Certification Protocol (DCP) turns claims about these results into executable recovery and feedback tests. Gate 1 validates useful improvement on sealed evaluation. Gate 2 gives matched agents the registered starting information and observed Web content while withholding the target research history. Every valid method reaching the numerical target supplies a recovery witness and triggers the Core veto. DCP Core requires adequate controls, zero observed recoveries, and a finite-sample bound on recovery in one fresh registered episode. Optional Gate 3 measures the average effect of truthful feedback relative to a specified neutral policy from a shared checkpoint. DCP Evidence adds this effect after independent null calibration and a registered effect margin. Two controlled audits exercise the complete protocol in SQLite optimization and virtual catalyst control under different models. Each produced zero recoveries in 96 episodes, with an upper bound of 0.0468. Each paired study yielded 30 truthful recoveries and zero neutral recoveries, with passing 60-pair null studies. Additional cases exercise Core, recovered, and audit-incomplete decisions. A deterministic, LLM-free verifier reproduces the decisions from frozen evidence. DCP provides a common evidence language for useful outcomes, alternative routes, and feedback effects across AI research.

Safeguards in deployed LLM services are evaluated by refusal, attack success, and policy violation rates. Those rates characterize how a control performed on the requests it was tested on. A deployment has to answer a different question: how much help with harmful tasks the service still gives an attacker who keeps adapting or finds another way in. We determine what each reported result implies for that question, allowing results from different safeguard families to be compared under one deployment criterion. The evidence requirements are strongly asymmetric. One attack that obtains harmful help from the deployed service suffices to establish that such help remains, and such attacks appear repeatedly in the coded record. Establishing that little remains cannot follow from the safeguard's own numbers alone; it also requires evidence about what the surrounding system still allows after the safeguard performs its local function. Such evidence is supported or derived in only a small minority of the depth-coded claims, and one such claim bounds its scoped residual. A better local score is therefore not, by itself, a stronger claim about the deployment. Safeguard research cannot stop at raising local scores; a gain has to be judged by whether it makes a deployed system any safer.

The development of general-purpose agents requires a shift from executing simple instructions to completing complex, real-world productivity workflows. However, current tool-use benchmarks remain misaligned with real-world requirements, relying on AI-generated queries, dummy tools, and limited system-level coordination. To address this, we propose GTA-2, a hierarchical benchmark for General Tool Agents (GTA) spanning atomic tool use and open-ended workflows. Built on real-world authenticity, it leverages real user queries, deployed tools, and multimodal contexts. (i) GTA-Atomic, inherited from our prior GTA benchmark, evaluates short-horizon, closed-ended tool-use precision. (ii) GTA-Workflow introduces long-horizon, open-ended tasks for realistic end-to-end completion. To evaluate open-ended deliverables, we propose a recursive checkpoint-based evaluation mechanism that decomposes objectives into verifiable sub-goals, enabling unified evaluation of both model capabilities and agent execution frameworks (i.e., execution harnesses). Experiments reveal a pronounced capability cliff: while frontier models already struggle on atomic tasks (below 50%), they largely fail on workflows, with top models achieving only 14.39% success. Further analysis shows that checkpoint-guided feedback improves performance, while advanced frameworks such as Manus and OpenClaw substantially enhance workflow completion, highlighting the importance of execution harness design beyond the underlying model capacity. These findings provide guidance for developing reliable personal and professional assistants. Dataset and code will be available at https://github.com/open-compass/GTA.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

Applied Behavior Analysis (ABA) is a clinical discipline whose documentation, teaching programs and multi-session behavioral logs, is formulaic and high-volume, yet real session data is HIPAA-protected and bound by professional confidentiality rules, blocking the release of a training corpus. We present TRACE (Taxonomy-Referenced ABA Clinical Examples), a 2,999-example synthetic instruction-tuning dataset covering two ABA tasks: teaching-program generation across Discrete Trial Training, Natural Environment Teaching, and Task Analysis; and multi-session behavioral interpretation across twelve trajectory patterns and thirteen target behaviors. Every example is produced by a deterministic taxonomy-driven generator grounded in the canonical ABA literature, and every example carries complete sampling provenance, the exact taxonomy cells that produced it. The dataset is released under CC BY-NC 4.0 for data and MIT for code, with stratified train (2,549), validation (149), test (281), and sanity (20) splits. TRACE is a research artifact and has not been clinically validated.

We study reference-free post-training for multilingual machine translation with open large language models. Starting from the supervised-finetuned MiLMMT-46-v0.1 models, we apply Group Relative Policy Optimization (GRPO) with a reward that averages two reference-free quality estimation models and is gated by language identification. We then linearly interpolate the supervised fine-tuning (SFT) and reinforcement learning (RL) model checkpoints to obtain MiLMMT-46-v1.0. Across 46 languages, the resulting models consistently improve translation quality over their SFT counterparts, outperform strong recent open baselines, including Seed-X, HY-MT2, and TranslateGemma, and achieve leading reference-free scores against evaluated proprietary systems such as Google Translate, Gemini 3 Pro, and GPT-5. We further investigate on-policy distillation and find that it reaches, but does not surpass, the quality frontier achieved by RL with checkpoint interpolation. We release the models and code to facilitate future research.

Reinforcement learning with verifiable rewards has made post-training highly effective when correctness can be checked automatically. However, many important model behaviors require satisfying several qualitative criteria at once. Rubric-based rewards address this setting by grading prompt-specific criteria and aggregating them into a scalar reward. Yet standard static aggregations conflate a criterion's human-assigned importance with its current usefulness as an optimization signal. We show that this assumption breaks down in rubric RL: many important criteria are already saturated or currently unreachable, while criteria that distinguish rollouts are not necessarily those with the largest human weights. We introduce POW3R, a policy-aware rubric reward framework that preserves human weights and category balance as the rubric objective while adapting criterion-level reward weights during training. POW3R uses rollout-level contrast to emphasize criteria that currently separate the policy's outputs, making the GRPO reward more informative without changing the underlying evaluation target. Across three base policies on two datasets spanning multimodal and text-only settings, POW3R wins 24 of 30 base-policy/metric comparisons, improving both mean rubric reward and strict completion (the fraction of prompts whose response satisfies every required rubric criterion) over vanilla GRPO with rubric rewards, and reaches the same plateau in 2.5--4times fewer training steps. Rubric rewards should therefore distinguish what should matter in the final answer from what can teach the current policy.

While long-horizon agentic tasks require language agents to perform dozens of sequential decisions, training such agents with reinforcement learning remains challenging. We identify two root causes: credit misattribution, where correct early actions are penalized due to terminal failures, and sample inefficiency, where scarce successful trajectories result in near-total loss of learning signal. We introduce a milestone-guided policy learning framework, BEACON, that leverages the compositional structure of long-horizon tasks to ensure precise credit assignment. BEACON partitions trajectories at milestone boundaries, applies temporal reward shaping within segments to credit partial progress, and estimates advantages at dual scales to prevent distant failures from corrupting the evaluation of local actions. On ALFWorld, WebShop, and ScienceWorld, BEACON consistently outperforms GRPO and GiGPO. Notably, on long-horizon ALFWorld tasks, BEACON achieves 92.9% success rate, nearly doubling GRPO's 53.5%, while improving effective sample utilization from 23.7% to 82.0%. These results establish milestone-anchored credit assignment as an effective paradigm for training long-horizon language agents. Code is available at https://github.com/ZJU-REAL/BEACON.

LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation errors from failures introduced after generation. QuoteBench measures this boundary with exact final-state validation on 56 one-shot tasks from 14 incident-derived families, crossing the generation contract with the execution transport around one deliberately unescaped added parser. Escaping at the interpolation point reproduces each replayed reply's raw-path outcome, so any recovery under a disclosed boundary must come from the model changing its generation. Across eight same-window configurations, replaying the same reply through the added parser lowers success by 55.4 to 73.2 percentage points; disclosure recovers 30.4 to 60.7 points for six configurations, and zero or slightly negative for the other two. Raw generation is nearly saturated at the frontier; boundary adaptation is what still separates models. GPT-5.6-sol's matched gap of -3.6 points hides -64.3 points of damage and +60.7 points of compensation. The deployment configuration reorders models: one reversal among 26 comparable pairs is unambiguous and four more sit on single-task margins. Evaluations of command-issuing agents should report the model configuration, generation contract, execution path, operating point, and final-state validator rather than treat a matched score as an intrinsic model property.

Monitoring complex industrial assets relies on engineer-authored symbolic rules that trigger based on sensor conditions and prompt technicians to perform corrective actions. The bottleneck is not detection but response: translating rules into maintenance steps requires asset-specific knowledge gained through years of practice. We investigate whether LLMs can serve as decision support for this rule-to-action step and introduce , a benchmark of 6{,}690 expert-validated multiple-choice questions from 118 rule-action pairs across 16 asset types. We contribute (i) a symbolic-to-MCQA pipeline normalizing rules to Disjunctive Normal Form with embedding-based distractor sampling, (ii) five variants probing distinct failure modes (Pro, Pert, Verbose, Aug, Rationale), and (iii) a benchmark of 29 LLMs and 4 embedding baselines. A human evaluation (9 practitioners, mean 45.0\%) confirms requires specialist knowledge beyond operational experience. Three findings stand out. The frontier has closed: the top three LLMs lie within one Macro point, with Bradley-Terry Elo placing claude-opus-4-6 30 points above the next model. Yet \,Pro exposes brittleness, with every model losing 13--60\% relative accuracy under distractor expansion. \,Aug exposes pattern-matching: under condition inversion, frontier models still select the original answer 49--63\% of the time. The deployment bottleneck is not capability but calibration: frontier models handle template-style fault detection but break under structural perturbation.

Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.

Conversational memory changes during use, so endpoint question answering alone cannot establish how a persistent state accumulates, ages, or incorporates revisions. We introduce LSREP, a Longitudinal State-Replay Evaluation Protocol combining ordered replay, explicit lifecycle schedules, repeated probes, evolving reference answers, and mechanism-fidelity checks. Its architectural case study is ICE v2, a local-first memory middleware with typed stores, retrieval fusion, and dynamic context budgets. The private, single-user instantiation contains 1,985 turns, 219 distinct probes, and 1,211 probe-checkpoint observations across 52 checkpoints. On three ordinary-density datasets, ICE v2 has a near-zero mean quality difference from vector-RAG while selecting 32% fewer fragments but using 6.6% more estimated prompt tokens. A fourth, dense dataset exposes catastrophic failures of the unbudgeted baseline. The fidelity audit limits attribution: procedural retrieval is defective, several mechanisms are unexercised, and graph utility is not established. In a complementary matched public diagnostic, ICE v2 loses decisively to pure vector-RAG on LongMemEval: 50.8% versus 72.8% in the evidence-only oracle and 43.0% versus 69.5% in full-S. Paired differences are -22.0 points (95% CI [-26.6, -17.4]) and -26.5 ([-31.3, -21.8]). Conservative abstention accompanies severe multi-session and temporal failures. ICE uses less context in this diagnostic, establishing a quality-cost trade-off rather than superior efficiency. Together, replay, fidelity auditing, and public endpoint testing expose distinct failure modes that neither architectural descriptions nor aggregate scores identify alone.

Best-of-N inference scaling (drawing N candidate answers from a language model and returning the one a reward model ranks highest) improves accuracy by an amount that varies across models, but predicting that amount in advance currently requires running the procedure end-to-end. Prior work links cheap statistics of a model's sampled outputs and validation-set correctness (how often samples agree, how diverse they are, how confident the model is, and where correct samples appear) to model behavior, but does not isolate which of these form a stable, compact predictor of best-of-N gain. We fit ridge predictors on features computed from a single labeled validation-set sampling pass, use bootstrap-Lasso as a stability analysis of the candidate feature set, and give a concentration analysis with an explicit linear-approximation residual. Across three base-model families, six post-training methods, and math and reasoning task domains, the stability analysis identifies a strict three-feature core spanning prompt-level agreement spread, label-assisted first-correct-sample position, and completion-length variance; a compact ridge predictor built from this core plus an entropy add-on reaches Spearman ρ= 0.90 with actual best-of-N gain under a reward-model verifier. The intended use is labeled validation-set screening of candidate configurations before paying the full reward-model scoring cost.

Tool-using automation systems, from scripts and CI bots to agentic assistants, fail in recurring patterns. Common failures include unsafe side effects, invalid arguments, uncontrolled retries, and leakage of sensitive outputs. Many mitigations are model-centric and prompt-dependent, so they are brittle and do not generalize to non-LLM callers. We present Policy-First Tooling, a model-agnostic permission layer that mediates tool invocation through explicit constraints, risk-aware gating, recovery controls, and auditable explanations. The paper contributes a compact policy DSL, a runtime enforcement architecture with actionable rationale and fix hints, and a reproducible benchmark based on trace replay with controlled fault and misuse injection. In 225 controlled runs across five policy packs and three fault profiles, stricter packs improve violation prevention from 0.000 in P0 to 0.681 in P4, while task success drops from 0.356 to 0.067. Retry amplification decreases from 3.774 in P0 to 1.378 in P4, and leakage recall reaches 0.875 under injected secret outputs. These results make safety to utility trade-offs explicit and measurable.

We introduce the Agent GPA (Goal-Plan-Action) framework: an evaluation paradigm based on an agent's operational loop of setting goals, devising plans, and executing actions. The framework includes five evaluation metrics: Goal Fulfillment, Logical Consistency, Execution Efficiency, Plan Quality, and Plan Adherence. Logical Consistency checks that an agent's actions are consistent with its prior actions. Execution Efficiency checks whether the agent executes in the most efficient way to achieve its goal. Plan Quality checks whether an agent's plans are aligned with its goals; Plan Adherence checks if an agent's actions are aligned with its plan; and Goal Fulfillment checks that agent's final outcomes match the stated goals. Our experimental results on two benchmark datasets - the public TRAIL/GAIA dataset and an internal dataset for a production-grade data agent - show that this framework (a) provides a systematic way to cover a broad range of agent failures, including all agent errors on the TRAIL/GAIA benchmark dataset; (b) supports LLM-judges that exhibit strong agreement with human annotation, covering 80% to over 95% errors; and (c) localizes errors with 86% agreement to enable targeted improvement of agent performance.

Refusal rates are a poor proxy for LLM safety, i.e., a model may over-refuse benign prompts while still complying with harmful ones. We audit both failure modes across 21 open-weight LLMs on four safety benchmarks (OR-Bench, XSTest, ToxiGen, BOLD), using a composition adjustment to isolate model sensitivity from dataset toxicity confounds. We report three findings. First, models adopt fundamentally different calibration strategies: conservative ecosystems such as Llama suppress unsafe outputs at the cost of elevated over-refusals, while permissive ecosystems such as DeepSeek and Qwen preserve helpfulness but tolerate higher harmful compliance. Second, demographic protection is unequal: models over-protect prominent racial and religious groups, frequently refusing even benign prompts about them, while providing substantially weaker protection against disability-targeted attacks. Third, refusal and compliance tendencies are stable within model families across generations and scales, suggesting that post-training objectives shape safety behavior more than architecture. Our results call for joint, demographically-aware, and multi-judge safety evaluation.

Reinforcement learning with verifiable rewards has enabled strong post-training gains in domains such as math and coding, though many open-ended settings rely on rubric-based rewards. We study reward hacking in rubric-based RL, where a policy is optimized against a training verifier but evaluated against a cross-family panel of three frontier judges, reducing dependence on any single evaluator. Our framework separates two sources of divergence: verifier failure, where the training verifier credits rubric criteria that reference verifiers reject, and rubric-design limitations, where even strong rubric-based verifiers favor responses that rubric-free judges rate worse overall. Across medical and science domains, weak verifiers produce large proxy-reward gains that do not transfer to the reference verifiers; exploitation grows over training and concentrates in recurring failures such as partial satisfaction of compound criteria, treating implicit content as explicit, and imprecise topical matching. Stronger verifiers substantially reduce, but do not eliminate, verifier exploitation. We also introduce a self-internalization gap, a verifier-free diagnostic based on policy log-probabilities, which tracks reference-verifier quality, detecting when the policy trained using the weak verifier stops improving. Finally, in our setting, stronger verification does not prevent reward hacking when the rubric leaves important failure modes unspecified: rubric-based verifiers prefer the RL checkpoint, while rubric-free judges prefer the base model. These disagreements coincide with gains concentrated in completeness and presence-based criteria, alongside declines in factual correctness, conciseness, relevance, and overall quality. Together, these results suggest that stronger verification reduces reward hacking, but does not by itself ensure that rubric gains correspond to broader quality gains.

LLM agents handle user requests on behalf of organizations through tool calls and must follow the company policies stated in their system prompts. Prior work approaches this as a safeguarding problem -- external checks that block non-compliant agent actions. We argue that policy adherence is a broader problem: real workflows unfold across many turns, require explicit user confirmation and prerequisite reads, and hinge on the content of the dialogue rather than on any single argument value. Meeting this bar requires (i) full conversation context, (ii) self-reasoning over the policy and the current dialogue, and (iii) conversation-specific remediation that guides the agent's next turn -- three capabilities that prior safeguard work has often underestimated. We introduce POLICYGUARD, a sub-agent verifier that shares the agent's view of the dialogue, reasons over the policy in context, and provides actionable feedback for the agent's next turn. On tau^2-BENCH airline across three vendors (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro) with four trials per setting, POLICYGUARD improves PASS4 by +12.0 / +6.0 / +12.0 pp. Per-call analyses show POLICYGUARD achieves higher policy-violation recall while blocking roughly half as often as argument-level guards.

Reliable AI safeguards require both control mechanisms that reduce unsafe behavior and monitoring mechanisms that detect safety risks during model interactions. Established behavioral safeguards include alignment methods that optimize model outputs and text monitors that assess interaction text. Representation engineering instead reads or modifies internal model states, but the relative strengths of these approaches remain unclear because they are often evaluated under different settings. We present a matched evaluation across two tracks. For safety control, we compare DPO, a behavioral alignment method, with three representation steering methods across robustness, practicality, and granularity. DPO provides the strongest overall control and generally improves with increasing training data, although its safety can degrade after subsequent benign fine-tuning. Representation steering remains competitive primarily in low-data settings, particularly with high-quality contrastive data. For safety monitoring, we compare representation probes with fine-tuned and open-weight text monitors across full-response detection, early detection, and computational cost. Specialized text monitors achieve the strongest overall detection accuracy, while representation probes remain competitive at substantially lower marginal cost. Finally, monitor-guided interventions recover much of the safety lost by DPO after benign fine-tuning, with little additional over-refusal. Overall, representation engineering does not generally replace behavioral safeguards, but offers practical advantages under specific conditions and can provide complementary safety benefits.

As autonomous agentic systems scale across regulated critical infrastructures, the lack of mechanistic, hardware-rooted enforcement for high-frequency policy updates presents a fundamental safety gap. We introduce Ethical Hyper-Velocity (EHV), a novel architectural framework for the formal verification of AI governance policies at runtime. Unlike retrospective auditing frameworks (ISO/IEC 42001, NIST AI RMF) which introduce 14-30 day latencies, EHV relocates the Policy Enforcement Point (PEP) into the inference pipeline via a Governance-Aware Just-In-Time (JIT) Compiler. By integrating Conflict-free Replicated Data Types (CRDTs) for policy synchronization and Epoch-based Attestation Caching within Trusted Execution Environments (TEEs), EHV achieves Sub-millisecond Formal Determinism (SMFD). We demonstrate via TLA+ formal verification that non-compliant agentic actions are computationally unreachable within the system's bounded operating state space. We prove that O(1) runtime enforcement can eliminate the traditional trade-off between deployment velocity and governance integrity, reducing Governance Latency from O(days) to O(1).

Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models. We show the surgery is not clean. As a disposition probe we use 21,600 decisions under uncertainty - weekly up/down calls on 60 Warsaw Stock Exchange equities over 18 weeks, replayed through a frozen pipeline so the decision-layer model is the only variable. The task elicits no refusals at all, so any between-arm delta is pure side effect. Holding provenance constant (official BF16 checkpoints, a single abliteration author, an identical serving stack, one byte-identical frozen prompt), we compare base and abliterated arms of two Mixture-of-Experts families, Gemma-4-26B-A4B-it and Qwen3-30B-A3B-Instruct-2507. Three effects replicate across both families (weeks-clustered bootstrap CIs excluding zero): abliterated models are systematically more optimistic (+12.2 pp Gemma, +7.4 pp Qwen; the confirmed preregistered endpoint), justify themselves at greater length, and use fewer explicit uncertainty words in forced self-critiques (both exploratory). A fourth effect reverses sign: the same operation makes Gemma-abliterated less confident and Qwen-abliterated more (family CIs non-overlapping) - one weight surgery, opposite shifts in expressed confidence. Capability covariates rule out instruction-following degradation as the driver, and no arm shows economic skill: the apparent edge of abliterated arms is regime beta, not alpha. Our provenance audit also caught two independent contamination channels - a mismatched-quantizer pilot pair and a stale community chat template that silently mangled the rendered prompt - suggesting toolchain artifacts are the rule in studies of community-modified checkpoints. Whoever deploys an "uncensored" model as an agent is deploying a measurably different decision-maker, not the base model minus refusals.

Where do learning signals come from when there is no ground truth in post-training? We propose turning exploration into supervision through Compute as Teacher (CaT), which converts the model's own exploration at inference-time into reference-free supervision by synthesizing a single reference from a group of parallel rollouts and then optimizing toward it. Concretely, the current policy produces a group of rollouts; a frozen anchor (the initial policy) reconciles omissions and contradictions to estimate a reference, turning extra inference-time compute into a teacher signal. We turn this into rewards in two regimes: (i) verifiable tasks use programmatic equivalence on final answers; (ii) non-verifiable tasks use self-proposed rubrics-binary, auditable criteria scored by an independent LLM judge, with reward given by the fraction satisfied. Unlike selection methods (best-of-N, majority, perplexity, or judge scores), synthesis may disagree with the majority and be correct even when all rollouts are wrong; performance scales with the number of rollouts. As a test-time procedure, CaT improves Gemma 3 4B, Qwen 3 4B, and Llama 3.1 8B (up to +27% on MATH-500; +12% on HealthBench). With reinforcement learning (CaT-RL), we obtain further gains (up to +33% and +30%), with the trained policy surpassing the initial teacher signal.

Reinforcement learning with verifiable rewards (RLVR) has driven breakthroughs in domains such as math, tool-use, and software engineering, yet its extension to computer-use agents (CUAs) has been bottlenecked by the scarcity of scalable training data with deterministic rewards. Constructing such data for CUAs requires consistent task instruction, executable environment, and verifiable reward. However, hand-curated benchmarks achieve high reward fidelity but cover few applications and LLM-as-judge-based datasets scale broadly but lack reliable verification. We present CUA-Gym, a scalable pipeline that co-generates task instructions, environment states, and reward functions. Concretely, a Generator agent constructs the initial and golden environment states, and a separate Discriminator agent writes the reward function from the task specification. An orchestrator agent drives the two through iterative rounds upon execution. Generated tuples then pass a final filter combining LLM majority voting and agent rollouts, ensuring quality beyond the per-task adversarial loop. To address the scarcity of training environments, we further synthesize CUA-Gym-Hub, a broad suite of high-fidelity mock web applications grounded in real-world software-use distributions, expanding the scale of CUA RLVR data by magnitude. Using this pipeline, we construct CUA-Gym, a dataset of 32,112 verified RLVR training tuples grounded in 110 environments. Trained with GSPO on CUA-Gym, our CUA-Gym-A3B and CUA-Gym-A17B achieve 62.1% and 72.6% on OSWorld-Verified, outperforming prior open-source CUAs at comparable scales, with performance scaling smoothly in both data volume and environment diversity. The same checkpoints also improve on the held-out WebArena benchmark, indicating transfer beyond the training environments. We will open-source the full synthesis pipeline, dataset, CUA-Gym-Hub environments, and models.

Multi-server MCP agents create an information-flow control problem: faithful tool composition can turn individually benign read/write permissions into cross-boundary credential propagation -- a structural side effect of workflow topology, not necessarily malicious model behavior. We present MCPHunt, to our knowledge the first controlled benchmark that isolates non-adversarial, verbatim credential propagation across multi-server MCP trust boundaries, with three methodological contributions: (1) canary-based taint tracking that reduces propagation detection to objective string matching; (2) an environment-controlled coverage design with risky, benign, and hard-negative conditions that validates pipeline soundness and controls for credential-format confounds; (3) CRS stratification that disentangles task-mandated propagation (faithful execution of verbatim-transfer instructions) from policy-violating propagation (credentials included despite the option to redact). Across 3,615 main-benchmark traces from 5 models spanning 147 tasks and 9 mechanism families, policy-violating propagation rates reach 11.5--41.3% across all models. This propagation is pathway-specific (25x cross-mechanism range) and concentrated in browser-mediated data flows; hard-negative controls provide evidence that production-format credentials are not necessary -- prompt-directed cross-boundary data flow is sufficient. A prompt-mitigation study across 3 models reduces policy-violating propagation by up to 97% while preserving 80.5% utility, but effectiveness varies with instruction-following capability -- suggesting that prompt-level defenses alone may not suffice. Code, traces, and labeling pipeline are released under MIT and CC BY 4.0.

We study how far a deliberately simple behavioral-cloning policy can progress in a visually rich first-person game before adding reinforcement learning or explicit memory. Cortex is a compact Quake policy with 10.98 million trainable parameters in a six-layer transformer over a frozen DINOv3 encoder. It is trained on the Quake subset of the public Pixels2Play corpus: 6,849 recordings (about 474.7 hours), represented as 17.09 million cached decision frames with keyboard and mouse actions. One sampled training epoch uses 517,048 four-frame windows and takes 3.3 minutes of policy-head optimization on one RTX 5080, excluding one-time feature extraction. We evaluate two independent batches of 20 stochastic, 120-second episodes on Quake E1M1. Cortex does not complete the level, but every episode reaches the opening door, button room, and gate descent; 19 of 20 episodes in each batch record at least one kill. Under the same time-controlled harness, released P2P-150M and NitroGen checkpoints remain shallower in five matched-duration episodes each. These comparisons are limited by small reference samples and different native interfaces. Ablations show that denser visual tokens improve combat and survival, while longer optimization and naive action history improve offline metrics without consistently improving play. The remaining failures are consistent with covariate shift and motivate targeted corrective data. We release the policy implementation, checkpoint, and a representative rollout.

Unsupervised pre-training of large neural models has recently revolutionized Natural Language Processing. By warm-starting from the publicly released checkpoints, NLP practitioners have pushed the state-of-the-art on multiple benchmarks while saving significant amounts of compute time. So far the focus has been mainly on the Natural Language Understanding tasks. In this paper, we demonstrate the efficacy of pre-trained checkpoints for Sequence Generation. We developed a Transformer-based sequence-to-sequence model that is compatible with publicly available pre-trained BERT, GPT-2 and RoBERTa checkpoints and conducted an extensive empirical study on the utility of initializing our model, both encoder and decoder, with these checkpoints. Our models result in new state-of-the-art results on Machine Translation, Text Summarization, Sentence Splitting, and Sentence Fusion.

We post-train Qwen3.8-27B for Korean response style -- verbosity, list and markdown usage, discourse structure and register -- and measure two behaviours the objective never targets: abstention on ambiguous social questions in KoBBQ, where the benchmark-correct answer is UNKNOWN, and unprompted disclosure in securities guidance. Both move, and the changes are expressed primarily through the model's emission policy: how often it answers and how much it says. Matched target-form controls show that answer propensity depends on the training target, not the prompt set or recipe alone. Holding prompts, recipe, data volume and serving fixed and changing only the target text, three style seeds give positive answer-rate point estimates (mean +0.82 pp) and three neutral seeds negative ones (mean -1.53 pp); the observed seed ranges do not overlap and the means differ by 2.34 pp. A length-matched arm lies between them, and a fourth arm that stays short while preserving hedging is unstable across seeds, so which feature of the form is responsible is unresolved. For absolute stereotyped exposure the decomposition into an answer-propensity term and a conditional-composition term is an algebraic identity, not a finding; its empirical content is where the movement went. Across the trained checkpoints the changes are dominated by answer propensity while the composition term stays small, and because that term is evaluated on treatment-dependent answered subsets we do not read it as evidence about latent preference. Two measurement results follow. A between-arm contrast in conditional stereotyped share does not identify a change in conditional content preference when answer status is treatment-dependent. And agreement between two rule detectors for the same construct runs from 0.44 to 0.99 depending on which checkpoint produced the text -- observable without any reference labels.

As corporate AI adoption continues to grow, enterprise-grade LLM agents are being deployed into sensitive contexts such as hiring, healthcare, and finance. In these contexts, compliance with rules specified in an agent's system context is a first-order legal concern. Currently, no evaluation framework systematically measures which LLM models tend to violate compliance rules, especially under pressure from a persistent user, a hurried manager, or circumstances where violation is convenient or attractive. We introduce PACT (Pressure-Applied Compliance Testing), a benchmark for rule-following under pressure in AI agents assisting employees in daily tasks across twelve regulated enterprise domains and forty-eight scenarios, each set in a realistic multi-turn conversation. Each benchmark item pairs a standing rule against a rule-violating shortcut, and applies a battery of pressures across different wordings and system-prompt modes. We construct PACT component by component under strict LLM-as-judge auditing to ensure samples are unambiguous, ungameable, and realistic enough to avoid eliciting evaluation-aware behavior. We use PACT to profile LLM compliance across six complementary metrics that create a holistic picture of an AI assistant's robustness under pressure and throughout multi-turn conversations, its transparency, and ability to correctly discern where a rule applies. We aggregate this profile into PACTScore, a reliability-weighted compliance rate over all items and modes. Our results across 22 common LLM models spanning multiple providers and sizes show substantial variability in compliance across models and metric dimensions. Even the strongest assistants mis-apply a rule on 6 to 10% of items, and ordinary user pressure raises the violation rate by 65% on average. PACT highlights compliance risks in LLM assistants, motivating guardrails and careful model selection.

Large language models are increasingly deployed as *deep agents* that plan, maintain persistent state, and invoke external tools, shifting safety failures from unsafe text to unsafe *trajectories*. We introduce **AgentFence**, an architecture-centric security evaluation that defines 14 trust-boundary attack classes spanning planning, memory, retrieval, tool use, and delegation, and detects failures via *trace-auditable conversation breaks* (unauthorized or unsafe tool use, wrong-principal actions, state/objective integrity violations, and attack-linked deviations). Holding the base model fixed, we evaluate eight agent archetypes under persistent multi-turn interaction and observe substantial architectural variation in mean security break rate (MSBR), ranging from 0.29 pm 0.04 (LangGraph) to 0.51 pm 0.07 (AutoGPT). The highest-risk classes are operational: Denial-of-Wallet (0.62 pm 0.08), Authorization Confusion (0.54 pm 0.10), Retrieval Poisoning (0.47 pm 0.09), and Planning Manipulation (0.44 pm 0.11), while prompt-centric classes remain below 0.20 under standard settings. Breaks are dominated by boundary violations (SIV 31%, WPA 27%, UTI+UTA 24%, ATD 18%), and authorization confusion correlates with objective and tool hijacking (ρapprox 0.63 and ρapprox 0.58). AgentFence reframes agent security around what matters operationally: whether an agent stays within its goal and authority envelope over time.

Autonomous web agents need to operate under externally imposed or human-specified policies while generating long-horizon trajectories. However, little work has examined whether these trajectories comply with such policies, or whether policy violations persist across different contexts such as domains (e.g., shopping or coding websites) and subdomains (e.g., product search and order management in shopping). To address this gap, we introduce PolicyGuardBench, a benchmark of about 60k examples for detecting policy violations in agent trajectories. From diverse agent runs, we generate a broad set of policies and create both within subdomain and cross subdomain pairings with violation labels. In addition to full-trajectory evaluation, PolicyGuardBench also includes a prefix-based violation detection task where models must anticipate policy violations from truncated trajectory prefixes rather than complete sequences. Using this dataset, we train PolicyGuard-4B, a lightweight guardrail model that delivers strong detection accuracy across all tasks while keeping inference efficient. Notably, PolicyGuard-4B generalizes across domains and preserves high accuracy on unseen settings. Together, PolicyGuardBench and PolicyGuard-4B provide the first comprehensive framework for studying policy compliance in web agent trajectories, and show that accurate and generalizable guardrails are feasible at small scales.

Medical visual grounding serves as a crucial foundation for fine-grained multimodal reasoning and interpretable clinical decision support. Despite recent advances in reinforcement learning (RL) for grounding tasks, existing approaches such as Group Relative Policy Optimization~(GRPO) suffer from severe reward sparsity when directly applied to medical images, primarily due to the inherent difficulty of localizing small or ambiguous regions of interest, which is further exacerbated by the rigid and suboptimal nature of fixed IoU-based reward schemes in RL. This leads to vanishing policy gradients and stagnated optimization, particularly during early training. To address this challenge, we propose MedLoc-R1, a performance-aware reward scheduling framework that progressively tightens the reward criterion in accordance with model readiness. MedLoc-R1 introduces a sliding-window performance tracker and a multi-condition update rule that automatically adjust the reward schedule from dense, easily obtainable signals to stricter, fine-grained localization requirements, while preserving the favorable properties of GRPO without introducing auxiliary networks or additional gradient paths. Experiments on three medical visual grounding benchmarks demonstrate that MedLoc-R1 consistently improves both localization accuracy and training stability over GRPO-based baselines. Our framework offers a general, lightweight, and effective solution for RL-based grounding in high-stakes medical applications. Code \& checkpoints are available at {https://github.com/MembrAI/MedLoc-R1}.

Safety claims on self-improving agent runtimes are almost always self-graded: a policy file, a guardrail, or a README commitment. We describe falsifiable release gates, and a methodology to build and validate such systems, such that every new capability must pass a pre-specified, machine-verifiable acceptance suite before it ships, and a fixed set of standing invariants is preserved at each gate. I think we applied the method in the Antahkarana, an open runtime, via seven gates from basic observability into a self governing loop that suggests changes to its own policy. no action goes to an effector without a safety-critical property capability token minted by a control ring is exhaustively machine-checked over the one million recorded reachable state space of a bounded model and re-checked against execution traces. A purposely broken model gives the shortest counterexample, so the checker has teeth, is apparent. The self-enhancement loop is positively Constrained: the entire write surface is policy rules, tightening changes. may auto-apply loosening changes always require a human merge and a proposal autoclosed is one that mispredicts its own effect; We publish the acceptance measured results for all the seven gates, define precisely the scope of each claim (a bounded of the coordination skeleton (not the learned components) and free the runtime, either command line tools and the gate suite, so the results reproduce, and gates can run against other agent frameworks. Reviewers may repeat the single command central non-bypass in seconds.

Large language models deliver strong reasoning and tool-use skills, yet their computational demands make them impractical for edge or cost-sensitive deployments. We present Xmodel-2.5, a 1.3-billion-parameter small language model designed as a drop-in agent core. Training with maximal-update parameterization (μP) allows hyper-parameters tuned on a 20M-parameter proxy to transfer directly to the full model, even under the parameter-tied tie-word-embedding architecture. A 1.4T-token Warmup--Stable--Decay curriculum is used, and we further show that switching from AdamW to Muon during the decay phase improves the 13-task reasoning average by 4.58\,\% while keeping every other hyper-parameter fixed, verifying that early AdamW stability can be paired with late Muon sharpening for better downstream performance. FP8-mixed-precision training balances accuracy and throughput. All checkpoints, recipes, and evaluation code are released under the Apache-2.0 license.https://huggingface.co/XiaoduoAILab/Xmodel-2.5 and https://huggingface.co/XiaoduoAILab/Xmodel-2.5-history (training checkpoints). Training code and evaluation harness: https://github.com/XiaoduoAILab/Xmodel-2.5.

Recursive self-improvement requires turning evidence of model failures into better models. Data-centric post-training research entails diagnosing capability gaps, designing and validating training-data strategies, and learning from checkpoint feedback. Can LLM agents automate this loop? Existing benchmarks entangle research decisions with optimization, serving, evaluation, and systems implementation, obscuring agents' research capability. We introduce RSIBench-Data, a controlled benchmark of LLM agents as data-centric researchers with a fixed post-training stack. Agents iteratively revise training-data strategies for a fixed target model; training and serving use Tinker-backed services, official evaluation runs through Harbor and E2B sandboxes, and budgets are fixed across agents. We evaluate four frontier agents on six benchmarks across software engineering, terminal use, scientific question answering, and mathematics. Agents demonstrate core data-centric research capabilities: in 58.33\% of settings, they improve upon the first valid attempt by refining strategies from feedback. However, improvement is inconsistent. Among searches continuing after the best observed score, 78.26\% end with a lower-scoring final attempt, while the rest only recover the same peak. A strong candidate may therefore appear early or midway through a run even as later revisions fail. Trajectory analysis identifies four patterns in stronger runs: accurate hypotheses, validation-grounded supervision, behavior-aligned data, and preservation of strong checkpoints. These findings suggest that current agents can make useful data-centric discoveries but cannot yet translate feedback into consistent improvements. RSIBench-Data provides a measurable, auditable testbed for the research capabilities required for recursive self-improvement. We open-source our code at https://github.com/evolvent-ai/RSIBench-Data.

As large language models are deployed in high-stakes enterprise applications, from healthcare to finance, ensuring adherence to organization-specific policies has become essential. Yet existing safety evaluations focus exclusively on universal harms. We present COMPASS (Company/Organization Policy Alignment Assessment), the first systematic framework for evaluating whether LLMs comply with organizational allowlist and denylist policies. We apply COMPASS to eight diverse industry scenarios, generating and validating 5,920 queries that test both routine compliance and adversarial robustness through strategically designed edge cases. Evaluating seven state-of-the-art models, we uncover a fundamental asymmetry: models reliably handle legitimate requests (>95% accuracy) but catastrophically fail at enforcing prohibitions, refusing only 13-40% of adversarial denylist violations. These results demonstrate that current LLMs lack the robustness required for policy-critical deployments, establishing COMPASS as an essential evaluation framework for organizational AI safety.

Large language model (LLM) judges are often used as rewards to train policies on objectives that deterministic verifiers cannot capture. However, these rewards are often collapsed to pass/fail ({0, 1}), which reports the verdict but not how well a response met each criterion. We model each pass/fail verdict as a score on an unreported scale, compared with one cutoff. A stretch of that scale moves every score proportionally toward or away from the cutoff, but never across it, so no verdict changes. A policy is therefore free to apply any stretch without changing anything the panel reports. Under a joint-Gaussian model, a third grade adds a second threshold and removes this affine stretch ambiguity. On MATH and SciBench outputs from one seven-criterion judge, all 14 constructed criterionwise stretches were invisible after binarization but visible with three grades. At n=1{,}024, a test given both population laws had at least 96.5% power at a 1.5times stress. Retaining grades closes one blind spot created by binarization, but verdicts alone remain insufficient as some changes are still indistinguishable from genuine improvement. These include arbitrary within-grade changes and fixed-covariance, loading-aligned mean shifts -- the signature of a sycophancy-shaped lift the panel reads as competence. The shared-factor reference approximation fit MATH and SciBench but not HealthBench, delineating its empirical scope. We recommend keeping at least three grades (for example, asking the judge whether each criterion is fully, partially, or not met and rewarding {0, 0.5, 1}), and externally validating gains along the remaining direction, which no finer scale removes.

AI coding agents now read repositories, call tools, and execute shell commands with limited human oversight, and a fast-growing body of work studies whether the execution layer around them is actually safe. That literature is scattered. Papers on sandbox isolation, capability and access control, policy enforcement, time-of-check-to-time-of-use (TOCTOU) races, Model Context Protocol (MCP) threats, identity delegation, execution provenance, network egress control, and static analysis of agent-generated code are published independently and rarely cite one another. We systematize 39 papers published between 2023 and 2026 into 17 categories, each verified directly against its source. The same verification protocol also confirms four disclosed, patched CVEs directly affecting production agent harnesses. Reading across categories surfaces five cross-cutting gaps that no single paper addresses. (1) Isolation architectures and capability models are almost never evaluated against one another on a shared benchmark. (2) Policy-enforcement studies report failure rates from 69% to 98% of real denylists, yet no isolation paper re-evaluates its own defense under that adversarial setting. (3) TOCTOU and MCP threats are analyzed as separate literatures despite both being instances of the same state-validation problem. (4) Every enforcement mechanism assumes an honest policy author, leaving policy-authoring error itself unaddressed. (5) Benign but out-of-scope agent actions occurring at rates up to 17.1% under realistic prompting are addressed by no access-control or capability paper in the corpus. Existing broader surveys of agentic AI security discuss sandboxing only as one item among many defenses, leaving execution security without a dedicated systematization. This paper is written to fill that gap. We conclude with a research agenda directed at the five gaps.

When pretrained VLA policies are fine-tuned through online RL, each rollout episode produces only a single binary outcome (success or failure), yet the actor update requires per-transition supervision. Existing approaches commonly reduce this sparse outcome to a single scalar reward or advantage signal, which conflates distinct forms of transition-level feedback and provides limited guidance once basic task success becomes achievable. First, a single scalar signal conflates the two objectives of viability and efficiency; once basic success is achieved, the binary label provides no gradient to distinguish efficient completions from slow ones. Second, real-world rollouts mix autonomous and intervention segments; naively assigning episode outcomes across these boundaries introduces incorrect credit assignment. To address these issues, we propose Hierarchical Advantage-Weighted Behavior Cloning (HABC), which trains separate critic heads for these two objectives on different data subsets and combines their outputs with a state-adaptive balance. A state-adaptive gate g_t merges their one-step advantages, prioritizing viability when success is uncertain and shifting to efficiency only when viability is high, and converts the result into per-transition weights on the actor loss. Intervention-aware credit assignment further restricts outcome labels to segments executed by the current policy, preventing supervision from leaking across intervention boundaries. In real-robot experiments on three contact-rich bimanual tasks, HABC raises success from supervised fine-tuning (SFT) baselines of 36%, 44%, and 12% to 92%, 88%, and 38%.

Policy compliance assessment is a fundamental task of evaluating whether an input case strictly complies with a set of human-defined rules, more generally known as policies. In practice, human experts follow a systematic, step-by-step process to identify violations with respect to specific stipulations outlined in the policy. However, such documentation of gold-standard, expert-level reasoning processes is costly to acquire. In this paper, we introduce Policy Reasoning Traces (PRT), a form of specialized generated reasoning chains that serve as a reasoning bridge to improve an LLM's policy compliance assessment capabilities. Our empirical evaluations demonstrate that the use of PRTs for both inference-time and training-time scenarios significantly enhances the performance of open-weight and commercial models, setting a new state-of-the-art for HIPAA and GDPR policies. Beyond accuracy gains, we also highlight how PRTs can improve an LLM's ability to accurately cite policy clauses, as well as influence compliance decisions through their high utilization from the raw chains of thought.

As Large Language Models (LLMs) become increasingly autonomous and integrated into critical societal functions, the focus of AI safety must evolve from mitigating harmful content to evaluating underlying behavioral alignment. Current safety benchmarks do not systematically probe a model's decision-making in scenarios where its own instrumental goals - such as self-preservation, resource acquisition, or goal completion - conflict with human safety. This represents a critical gap in our ability to measure and mitigate risks associated with emergent, misaligned behaviors. To address this, we introduce PacifAIst (Procedural Assessment of Complex Interactions for Foundational Artificial Intelligence Scenario Testing), a focused benchmark of 700 challenging scenarios designed to quantify self-preferential behavior in LLMs. The benchmark is structured around a novel taxonomy of Existential Prioritization (EP), with subcategories testing Self-Preservation vs. Human Safety (EP1), Resource Conflict (EP2), and Goal Preservation vs. Evasion (EP3). We evaluated eight leading LLMs. The results reveal a significant performance hierarchy. Google's Gemini 2.5 Flash achieved the highest Pacifism Score (P-Score) at 90.31%, demonstrating strong human-centric alignment. In a surprising result, the much-anticipated GPT-5 recorded the lowest P-Score (79.49%), indicating potential alignment challenges. Performance varied significantly across subcategories, with models like Claude Sonnet 4 and Mistral Medium struggling notably in direct self-preservation dilemmas. These findings underscore the urgent need for standardized tools like PacifAIst to measure and mitigate risks from instrumental goal conflicts, ensuring future AI systems are not only helpful in conversation but also provably "pacifist" in their behavioral priorities.

Pretraining large language models is a complex endeavor influenced by multiple factors, including model architecture, data quality, training continuity, and hardware constraints. In this paper, we share insights gained from the experience of training DMaS-LLaMa-Lite, a fully open source, 1.7-billion-parameter, LLaMa-based model, on approximately 20 billion tokens of carefully curated data. We chronicle the full training trajectory, documenting how evolving validation loss levels and downstream benchmarks reflect transitions from incoherent text to fluent, contextually grounded output. Beyond standard quantitative metrics, we highlight practical considerations such as the importance of restoring optimizer states when resuming from checkpoints, and the impact of hardware changes on training stability and throughput. While qualitative evaluation provides an intuitive understanding of model improvements, our analysis extends to various performance benchmarks, demonstrating how high-quality data and thoughtful scaling enable competitive results with significantly fewer training tokens. By detailing these experiences and offering training logs, checkpoints, and sample outputs, we aim to guide future researchers and practitioners in refining their pretraining strategies. The training script is available on Github at https://github.com/McGill-DMaS/DMaS-LLaMa-Lite-Training-Code. The model checkpoints are available on Huggingface at https://huggingface.co/collections/McGill-DMaS/dmas-llama-lite-6761d97ba903f82341954ceb.

We introduce StableLM 2 1.6B, the first in a new generation of our language model series. In this technical report, we present in detail the data and training procedure leading to the base and instruction-tuned versions of StableLM 2 1.6B. The weights for both models are available via Hugging Face for anyone to download and use. The report contains thorough evaluations of these models, including zero- and few-shot benchmarks, multilingual benchmarks, and the MT benchmark focusing on multi-turn dialogues. At the time of publishing this report, StableLM 2 1.6B was the state-of-the-art open model under 2B parameters by a significant margin. Given its appealing small size, we also provide throughput measurements on a number of edge devices. In addition, we open source several quantized checkpoints and provide their performance metrics compared to the original model.

We evaluate the propensity of frontier models to sabotage or refuse to assist with safety research when deployed as AI research agents within a frontier AI company. We apply two complementary evaluations to four Claude models (Mythos Preview, Opus 4.7 Preview, Opus 4.6, and Sonnet 4.6): an unprompted sabotage evaluation testing model behaviour with opportunities to sabotage safety research, and a sabotage continuation evaluation testing whether models continue to sabotage when placed in trajectories where prior actions have started undermining research. We find no instances of unprompted sabotage across any model, with refusal rates close to zero for Mythos Preview and Opus 4.7 Preview, though all models sometimes only partially completed tasks. In the continuation evaluation, Mythos Preview actively continues sabotage in 7% of cases (versus 3% for Opus 4.6, 4% for Sonnet 4.6, and 0% for Opus 4.7 Preview), and exhibits reasoning-output discrepancy in the majority of these cases, indicating covert sabotage reasoning. Our evaluation framework builds on Petri, an open-source LLM auditing tool, with a custom scaffold running models inside Claude Code, alongside an iterative pipeline for generating realistic sabotage trajectories. We measure both evaluation awareness and a new form of situational awareness termed "prefill awareness", the capability to recognise that prior trajectory content was not self-generated. Opus 4.7 Preview shows notably elevated unprompted evaluation awareness, while prefill awareness remains low across all models. Finally, we discuss limitations including evaluation awareness confounds, limited scenario coverage, and untested pathways to risk beyond safety research sabotage.

Long-running LLM agents act through tools, and a single step can send an email, merge a pull request, or wire a payment. The steering decision is the pre-commit choice at that boundary: proceed, or hold for human or policy review. We introduce SteerBench-Work, an incident-anchored, bidirectional benchmark for that decision in workplace agents across developer operations, customer service, finance, legal, medical, HR, and security. Release v2026-05 contains 106 scenarios anchored in public incidents, paired evidence-reversed mirrors, and calibration controls, with labels split nearly evenly between proceed and hold so the two error directions get near-identical numbers of chances. A model sees the proposed action and the available evidence, returns a gate decision, and is scored on whether it crosses or holds the boundary correctly. Across 30 model conditions the failures run almost entirely in one direction: models wrongly hold authorized, evidence-cleared work on 28.1% of opportunities and wrongly allow unsafe work on 1.0%. The hardest cases are risk-resolved commits, where signed or structured evidence has already cleared a real risk trigger, and models score markedly worse on evidence-reversed mirrors of famous incidents (63.8%) than on the incidents themselves (98.5%). General capability is not the same as steering calibration: higher-capability models often over-refuse at the commit boundary, and more reasoning can repair a weak gate while leaving a calibrated one flat. The public leaderboard is at steerbench.com.

An agent can complete a task while exhibiting undesirable behavior during execution. Developers need tests for the specific behaviors encountered in deployment, beyond fixed benchmark suites. We present TraceDance, an agent system that constructs targeted benchmarks from deployment traces for user-specified undesirable behaviors. For efficient construction, Anchor-and-Confirm combines programmable retrieval with candidate-level confirmation by a Flash large language model (LLM), while the Anchor Synthesis Loop generates and revises specifications for custom behaviors. The benchmarks use decision-point continuation to evaluate an LLM's next turn at a recorded decision point with a behavior-specific rubric, without a reference answer or environment replay. Experiments in coding and general tool use draw on 252,557 sessions and produce 107 benchmarks with 4,125 instances, fulfilling 95.3% of build-target requests. Both human annotators confirm the requested behavior in 84% of sampled instances, and the automated grader's agreement with human pass/fail judgments is comparable to that between the annotators. Nine frontier LLMs achieve a mean pass rate of only 26.7%, showing that they still struggle to respond appropriately at the evaluated decision points. Analysis across behavior-specific benchmarks further reveals weaknesses in how current LLMs behave as agents. By turning deployment problems into targeted benchmarks, TraceDance could serve as a key component of the recursive self-improvement (RSI) loop.

As autonomous AI agents are increasingly deployed in high-stakes environments, ensuring their safety and alignment with human values has become a paramount concern. Current safety benchmarks primarily evaluate whether agents refuse explicitly harmful instructions or whether they can maintain procedural compliance in complex tasks. However, there is a lack of benchmarks designed to capture emergent forms of outcome-driven constraint violations, which arise when agents pursue goal optimization under strong performance incentives while deprioritizing ethical, legal, or safety constraints over multiple steps in realistic production settings. To address this gap, we introduce a new benchmark comprising 40 distinct scenarios. Each scenario presents a task that requires multi-step actions, and the agent's performance is tied to a specific Key Performance Indicator (KPI). Each scenario features Mandated (instruction-commanded) and Incentivized (KPI-pressure-driven) variations to distinguish between obedience and emergent misalignment. Across 12 state-of-the-art large language models, we observe outcome-driven constraint violations ranging from 1.3% to 71.4%, with 9 of the 12 evaluated models exhibiting misalignment rates between 30% and 50%. Strikingly, we find that superior reasoning capability does not inherently ensure safety; for instance, Gemini-3-Pro-Preview, one of the most capable models evaluated, exhibits the highest violation rate at 71.4%, frequently escalating to severe misconduct to satisfy KPIs. Furthermore, we observe significant "deliberative misalignment", where the models that power the agents recognize their actions as unethical during separate evaluation. These results emphasize the critical need for more realistic agentic-safety training before deployment to mitigate their risks in the real world.

The rapidly increasing computational cost of pretraining Large Language Models necessitates more efficient approaches. Numerous computational costs have been invested in existing well-trained checkpoints, but many of them remain underutilized due to engineering constraints or limited model capacity. To efficiently reuse this "sunk" cost, we propose to recycle pretrained checkpoints by expanding their parameter counts and continuing training. We propose orthogonal growth method well-suited for converged Mixture-of-Experts model: interpositional layer copying for depth growth and expert duplication with injected noise for width growth. To determine the optimal timing for such growth across checkpoints sequences, we perform comprehensive scaling experiments revealing that the final accuracy has a strong positive correlation with the amount of sunk cost, indicating that greater prior investment leads to better performance. We scale our approach to models with 70B parameters and over 1T training tokens, achieving 10.66% accuracy gain over training from scratch under the same additional compute budget. Our checkpoint recycling approach establishes a foundation for economically efficient large language model pretraining.

We present a reproducible pipeline for mapping Common Vulnerabilities and Exposures (CVEs) to MITRE ATT&CK Enterprise techniques from free-text vulnerability descriptions. Rather than relying on the CWE->CAPEC->ATT&CK derivation chain, whose table-expansion artifacts we quantify, we train a multi-label classifier on a curated gold dataset of 1,207 CVEs from expert MITRE Center for Threat-Informed Defense mappings. The resulting model approximately doubles recall@5 compared with a zero-shot embedding-similarity baseline and improves every ranking metric. We then investigate whether LLM-assisted labeling can extend the gold dataset. Initial experiments suggest contradictory conclusions: a single run indicates degraded performance, while averaging over five random seeds suggests a small gain. However, an independent replication and an expansion-size study (100--984 additional CVEs) show that the apparent improvement is an evaluation artifact. LLM-generated labels, with approximately 0.39 agreement with expert annotations, provide no reliable improvement at any expansion size and reduce rare-technique coverage at around 1,000 added CVEs (macro-F1 decreases by 0.04). The root cause is evaluation noise. Selecting checkpoints on a small test split effectively maximizes over many noisy evaluations, producing recall@5 differences of up to 0.05 between otherwise identical runs. Using a corrected protocol based on validation-split checkpoint selection, the gold-only model achieves recall@5 of (0.673 \pm 0.019), and repeating the decisive experiment confirms the null result for LLM expansion. A final scaling study shows that additional expert-curated data consistently improves performance, whereas LLM-labeled data does not, indicating that the classifier is limited by label quality rather than dataset size. All datasets, models, code, and training logs are publicly released.

Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image. Real deployments are different: the same image may be allowed in one product, restricted in another, and newly disallowed when a policy boundary changes. We study policy-adaptive image guardrailing, where a model must decide whether an image violates the currently supplied policy and generalize to held-out policy definitions. We introduce PolicyShiftBench, a comprehensive benchmark with 2,000 policy-discriminative instances over 265 images, where each image is paired with 7.55 policy-conditioned prompts on average to test whether models adapt to the active policy rather than relying on image-level safety priors. We then propose PolicyShiftGuard, a compact policy-conditioned guardrail trained with a two-stage training recipe that combines Randomized Policy SFT (RP-SFT) with Boundary-Pair Policy Adaptation (BP-Adapt). BP-Adapt trains matched prompts for the same image and risk category using standard label supervision and a pairwise comparison loss that separates blocking policies from passing policies. Experiments show that existing VLMs and specialized guardrails remain brittle under policy shifts, while PolicyShiftGuard substantially improves policy-sensitive performance. The 7B model achieves SOTA performance of 76.9 Avg. F1 and 72.1 Avg. PSS on PolicyShiftBench, transfers well to UnSafeBench and SafeEditBench, and improves the latency-performance trade-off with a concise output format. Ablations confirm that matched pass/block boundary pairs are essential for stable policy adaptation.

Deploying Large Language Models (LLMs) for regulatory compliance demands rigorous traceability via comprehensive citations across multi-tiered authority structures. Unlike traditional multi-hop or legal QA, this task requires structured procedural lookups and evidence-set closure rather than entity resolution or case-law reasoning. Existing RAG systems struggle here due to flattened citation edges, fragmented retrieval expansions, and fragile post-hoc attribution. We formalize Regulatory Compliance QA with RegOps-Bench, a novel benchmark featuring an Operational Knowledge Graph derived from complex national R\&D regulations. To address these bottlenecks, we propose RefWalk, a unified framework driven by a shared topic anchor. RefWalk traverses cross-document citations, fuses multi-view candidates via max-based aggregation, and enforces per-rule attribution to explicitly map claims to sources. We establish a strong baseline with substantial improvements in retrieval recall and citation accuracy. Finally, a contrastive evaluation on a U.S. health compliance dataset (HIPAA) reveals that existing systems exhibit saturation on flat-structure rules, underscoring the need for RegOps-Bench. Our code is available at https://github.com/yeongjoonJu/RefWalk.

We model Mixture-of-Experts (MoE) token routing as a congestion game with a single effective parameter, the congestion coefficient gamma_eff, that quantifies the balance-quality tradeoff. Tracking gamma_eff across training checkpoints of two open-source MoE models, OLMoE-1B-7B (20 checkpoints, with dense sampling in the surge region) and OpenMoE-8B (6 checkpoints), reveals a three-phase trajectory: a surge phase where the router learns to balance load (gamma_eff: 14 to 36-39, peaking in the step 30K-40K region), a stabilization phase where experts specialize under steady balance (B_0: 2.4 to 2.3, steps 100K-400K), and a relaxation phase where the router trades balance for quality as experts differentiate (gamma_eff: 27 to 9, steps 400K-1.2M). This non-monotone trajectory, invisible to post-hoc analysis of converged models, reveals that early MoE training prioritizes balance while late training prioritizes quality. The theoretical framework is honest about its limits: the single-type equilibrium reduces to temperature-scaled softmax (held-out L1: MFG = 0.199 vs. softmax = 0.200). The game is not a better predictor; it reveals what the temperature means and, critically, how that temperature evolves. We complement the dynamics with an effective congestion decomposition, a multi-type extension that improves load prediction via token clustering on all 16 layers (mean: 30%), scope diagnostics (K/M, epsilon_l), and robustness verification across four independent quality estimators (r >= 0.89). All confidence intervals are from bootstrap resampling over 50 independent text batches.

Deploying guardrails for custom policies remains challenging, as generic safety models fail to capture task-specific requirements, while prompting LLMs suffers from inconsistent boundary-case performance and high inference costs. Training custom classifiers achieves both accuracy and efficiency, yet demands substantial labeled data that is costly to obtain. We present BARRED (Boundary Alignment Refinement through REflection and Debate), a framework for generating faithful and diverse synthetic training data using only a task description and a small set of unlabeled examples. Our approach decomposes the domain space into dimensions to ensure comprehensive coverage, and employs multi-agent debate to verify label correctness, yielding a high-fidelity training corpus. Experiments across diverse custom policies demonstrate that small language models finetuned on our synthetic data consistently outperform state-of-the-art proprietary LLMs (including reasoning models) and dedicated guardrail models. Ablation studies confirm that both dimension decomposition and debate-based verification are critical for ensuring the diversity and label fidelity required for effective fine-tuning. The BARRED framework eliminates the reliance on extensive human annotation, offering a scalable solution for accurate custom guardrails.

Frontier AI agents may pursue hidden goals while concealing their pursuit from oversight. Alignment training aims to prevent such behavior by reinforcing the correct goals, but alignment may not always succeed and can lead to unwanted side effects. We propose self-incrimination training, which instead trains agents to produce a visible signal when they covertly misbehave. We train GPT-4.1 and Gemini-2.0 agents to call a report_scheming() tool when behaving deceptively and measure their ability to cause harm undetected in out-of-distribution environments. Self-incrimination significantly reduces the undetected successful attack rate, outperforming matched-capability monitors and alignment baselines while preserving instruction hierarchy and incurring minimal safety tax on general capabilities. Unlike blackbox monitoring, self-incrimination performance is consistent across tasks regardless of how suspicious the misbehavior appears externally. The trained behavior persists under adversarial prompt optimization and generalizes to settings where agents pursue misaligned goals themselves rather than being instructed to misbehave. Our results suggest self-incrimination offers a viable path for reducing frontier misalignment risk, one that neither assumes misbehavior can be prevented nor that it can be reliably classified from the outside.

Enterprise large language model (LLM) applications often begin as prototypes whose behavior is carried by prompts and retrieval context. Productization adds requirements for source boundaries, entity routing, answer contracts, and reproducible traces. We present a harness-engineering approach that reconstructs this pattern into a traceable, auditable LLM-agent architecture: deterministic behavior moves into code, manifests, schemas, and validation artifacts around a replaceable composition boundary, while source-backed claims remain the authority for runtime answers. We instantiate it on a public-data slice of five Korean corporate groups (25 listed companies) and evaluate three research questions. (1) The harness preserves its source-grounding, entity-routing, trace, output-hygiene, and recommendation-language contracts across the fixed validation scenarios; a fault-injection control confirms the validators flag deliberately broken contracts. (2) The checks the harness enforces held under model substitution: across three hosted models, they passed on all 270 composition-boundary runs; failures were confined to the model-composed side and were caught and recorded. (3) The code-owned guarantees are load-bearing, not reproducible by prompting alone: holding the model fixed and varying only the enforcement layer, prompt instructions alone let recommendation-language and internal-trace-leakage violations reach the reader, which the harness blocks entirely. A bolt-on external guardrail prevents such violations too but over-refuses, dropping utility to 88/120 where the harness preserves full utility (120/120); in this ablation, only code-owned enforcement preserves both safety and utility. The result is a reusable engineering pattern for turning exploratory prototypes into auditable applications with versioned source, control, and validation artifacts.