jaival17/AIME2025

AIME 2025 (RLVR format)

This dataset contains the 30 problems of the 2025 American Invitational Mathematics Examination (AIME I and AIME II), each with its integer answer, in the verl-style RLVR (reinforcement learning with verifiable rewards) format. In our paper Adaptive Negative Reinforcement for LLM Reasoning: Dynamically Balancing Correction and Diversity in RLVR, it is the hardest of our three evaluation benchmarks and is out of distribution relative to the MATH training set.

The file is an unmodified copy of the AIME 2025 evaluation data released by Zhu et al. with The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (NeurIPS 2025). The data_source field is TianHongZXY/AIME2025. We re-host the data so that our experiments can be reproduced with exactly the same file.

Evaluation only. This dataset is intended for evaluation. Do not train on it.

Dataset summary

Split	Problems	Use in our paper
test	30	Evaluation (Pass@k, k = 1 … 256, 256 samples per problem)

AIME answers are integers from 0 to 999. One stored target, 336^\circ, keeps a degree sign from the source data, so strip ^\circ before an exact-match comparison.

Because there are only 30 problems, one problem corresponds to 3.33 percentage points of Pass@k. Small differences between methods should be read with this in mind.

Data fields

Field	Type	Description
data_source	string	Always "TianHongZXY/AIME2025"
prompt	list of {role, content}	Chat-format prompt: one user message containing the problem statement
ability	string	Always "math"
reward_model.style	string	Always "rule" (rule-based verifiable reward)
reward_model.ground_truth.question	string	Problem statement
reward_model.ground_truth.target	string	Final answer
extra_info.index	int	Problem index (0–29)
extra_info.split	string	Always "test"

No reference solutions are included.

Example

{
  "data_source": "TianHongZXY/AIME2025",
  "prompt": [{"role": "user", "content": "Find the sum of all integer bases $b>9$ for which $17_{b}$ is a divisor of $97_{b}$."}],
  "ability": "math",
  "reward_model": {
    "style": "rule",
    "ground_truth": {
      "question": "Find the sum of all integer bases $b>9$ for which $17_{b}$ is a divisor of $97_{b}$.",
      "target": "70"
    }
  },
  "extra_info": {"index": 0, "split": "test"}
}

Usage

from datasets import load_dataset

ds = load_dataset("<your-username>/AIME2025-RLVR", split="test")
for row in ds:
    question = row["prompt"][0]["content"]
    answer = row["reward_model"]["ground_truth"]["target"]

In our experiments with Qwen2.5-Math-1.5B, the problem was wrapped in this chat template:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
{input}
Please reason step by step, and put your final answer
within \boxed{}.<|im_end|>
<|im_start|>assistant

Generation settings: temperature 0.6, top-p 0.95, 256 samples per problem. Pass@k was computed with the unbiased estimator of Chen et al. (2021).

Source and license

• Problems: 2025 AIME I and AIME II, © Mathematical Association of America (MAA). The problems are reproduced here for non-commercial research and evaluation. Copyright remains with the MAA.
• Preprocessing: Zhu et al. (2025), https://github.com/TianHongZXY/RLVR-Decomposed.

Citation

@inproceedings{zhu2025nsr,
  title     = {The Surprising Effectiveness of Negative Reinforcement in {LLM} Reasoning},
  author    = {Zhu, Xinyu and Xia, Mengzhou and Wei, Zhepei and Chen, Wei-Lin and Chen, Danqi and Meng, Yu},
  booktitle = {NeurIPS},
  year      = {2025}
}

Our paper:

@article{ingle2026adaptive,
  title  = {Adaptive Negative Reinforcement for {LLM} Reasoning: Dynamically Balancing Correction and Diversity in {RLVR}},
  author = {Ingle, Yash and Chauhan, Jaival and Yadav, Ankit and Mishra, Sudhakar},
  year   = {2026}
}

Code: A-NSR · CW-NSR