Reproduce “ReasonIF: Large Reasoning Models Fail to Follow Instructions During Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yongchan Kwon, Shang Zhu, Federico Bianchi, Kaitlyn Zhou, James Zou
Large Reasoning Models (LRMs) from institutions like Together AI and Stanford University consistently fail to adhere to user instructions within their internal reasoning processes, achieving an average instruction following score of only 15.6% compared to 57.3% for final responses. This failure to follow reasoning instructions, such as specific formatting or language, is more pronounced with increasing task difficulty and is observed across various open-source models.
Key insights
- There is a substantial gap between LRM instruction following in reasoning traces (average 15.6% IFS) and main responses (average 57.3% IFS), indicating superficial compliance in many cases.
- LRMs' ability to follow instructions in their reasoning traces degrades with increasing task difficulty and decreasing model accuracy, showing a strong positive correlation (average Pearson coefficient of 0.784).
- Multi-turn reasoning with explicit feedback provides limited improvement, increasing average reasoning IFS by 16.6% but still leaving all models below 45% compliance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.