ReasonIF
收藏资源简介:
ReasonIF是一个系统基准数据集,用于评估大型推理模型在推理过程中遵循指令的能力。它包含300个样本,涵盖数学、科学和常识问题,要求模型在解决这些问题时遵循明确的指令约束,如多语言推理、格式控制和长度限制。该数据集旨在揭示模型在推理过程中指令遵循的不足,并支持相关改进策略的研究。
ReasonIF is a systematic benchmark dataset for evaluating the instruction-following capabilities of large reasoning models during their inference processes. It comprises 300 samples spanning mathematical, scientific, and commonsense problems, where models are required to comply with explicit instruction constraints such as multilingual reasoning, format control, and length limits when solving these tasks. This dataset aims to uncover the deficiencies in models' instruction-following performance during inference, and support research on relevant improvement strategies.
ReasonIF 数据集概述
数据集简介
ReasonIF 是一个系统性基准测试,用于评估大型推理模型在推理过程中遵循指令的能力。该基准涵盖多语言推理、格式控制和长度控制等维度。
核心发现
- 前沿大型推理模型(包括 GPT-OSS-120B、Qwen3-235B 和 DeepSeek-R1)在超过75%的情况下无法遵循推理指令
- 随着任务难度增加,推理指令遵循能力进一步下降
- 推理指令遵循得分显著低于主要响应指令遵循得分
数据集文件
data/reasonIF_dataset.json:核心基准数据集,包含300个样本,要求模型在解决数学、科学和常识问题时遵循显式指令data/number_of_words_reference.json:词数限制指令的参考文件,提供论文结果复现所需的参考计数
使用方法
环境设置
bash uv sync source .venv/bin/activate
模型推理
bash python -m src.main --model_name "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"
结果评估
bash python -m src.eval_core --model_name "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"
输出结果
结果保存在 outputs/[model-name]/ 目录:
model_outputs_reasonIF.json:包含推理内容的原始模型输出eval_results.json:评估指标,包括指令遵循准确率
改进策略
- 多轮推理
- 使用合成数据进行推理指令微调(RIF)
相关资源
- 数据集地址:https://huggingface.co/datasets/ykwon-hf/reasonIF
- 论文地址:https://arxiv.org/pdf/2510.15211.pdf
- 博客地址:https://www.together.ai/blog/large-reasoning-models-fail-to-follow-instructions-during-reasoning-a-benchmark-study
引用信息
bibtex @article{kwon2025reasonif, title = {ReasonIF: Large Reasoning Models Fail to Follow Instructions During Reasoning}, author = {Yongchan Kwon and Shang Zhu and Federico Bianchi and Kaitlyn Zhou and James Zou}, year = {2025}, journal = {arXiv preprint arXiv:2510.15211}, archivePrefix= {arXiv}, eprint = {2510.15211}, primaryClass = {cs.LG}, note = {Preprint — submitted October 17, 2025} }




