sxiong/MLR_concat10

MLR-Concat10 Benchmark

MLR-Concat10 is a compact long-horizon reasoning benchmark for evaluating whether language models can sustain reasoning across a sequence of independent problems. Each example concatenates 10 randomly sampled questions into one prompt and asks the model to solve them sequentially.

The benchmark follows the long-horizon reasoning evaluation protocol in Enhancing Language Model Reasoning with Structured Multi-Level Modeling (ICLR 2026).

Configurations

Config	Source questions	Long-horizon examples	Questions per example
math500	MATH500	10	10
aime24	AIME24	10	10
mixture	MATH500, AIME24, GPQA-diamond, and BoardGameQA-Hard	10	10

Every configuration contains 100 source questions in total, arranged into 10 prompts. Questions in a prompt are independent; their ordering defines the required solution order.

Data format

Each row is a JSON object with three fields:

Field	Type	Description
ids	list[string]	Identifiers for the 10 component questions, in prompt order.
problem	string	The concatenated instruction and 10 numbered questions.
answers	list[string]	Gold answers for the component questions, aligned with ids and prompt order.

Example:

from datasets import load_dataset

ds = load_dataset("sxiong/MLR_concat10", "math500", split="train")
example = ds[0]

prompt = example["problem"]
gold_answers = example["answers"]  # 10 answers, in question order

For evaluation, parse one answer for each numbered question and compare it against the correspondingly positioned entry in answers. The dataset stores answers as strings to preserve mathematical notation and multiple-choice labels.

Citation

@inproceedings{xiong2026enhancing,
  title={Enhancing language model reasoning with structured multi-level modeling},
  author={Xiong, Siheng and Payani, Ali and Fekri, Faramarz},
  booktitle={International Conference on Learning Representations},
  volume={2026},
  pages={36557--36610},
  year={2026}
}