ReasonIF: Large Reasoning Models Fail to Follow Instructions During Reasoning

TL;DR

This paper addresses how large reasoning models (LRMs) often fail to follow user instructions within their reasoning traces, potentially hindering reliability and safety. It introduces ReasonIF, a benchmark that probes instruction adherence inside reasoning processes across six prompt types and five data sources, using the Instruction Following Score (IFS) as the evaluation metric. Experiments across multiple open-source LRMs reveal a substantial gap between adherence in reasoning traces versus main outputs, with stronger misalignment on harder problems. The authors explore mitigation strategies—multi-turn reasoning and Reasoning Instruction Finetuning (RIF)—which yield measurable improvements in reasoning IFS, suggesting promising directions for making LRMs more controllable and auditable in their reasoning workflows. Overall, ReasonIF highlights a critical, underexplored aspect of trustworthy AI and sets the stage for future work on controlling and validating reasoning paths in LRMs.

Abstract

The ability of large language models (LLMs) to follow user instructions is central to their reliability, safety, and usefulness. While prior studies assess instruction adherence in the model's main responses, we argue that it is also critical for large reasoning models (LRMs) to follow user instructions throughout their reasoning process. Reasoning instruction following makes LRMs more controllable and transparent, while reducing risks of undesirable shortcuts, hallucinations, or reward hacking within reasoning traces. To evaluate this dimension, we introduce ReasonIF, a systematic benchmark for assessing reasoning instruction following. ReasonIF includes six categories of instruction prompts, spanning multilingual reasoning, formatting and length control. Across many open-source LRMs including GPT-OSS, Qwen3, and DeepSeek-R1, we find substantial failures in reasoning instruction adherence: the highest instruction following score (IFS) remains below 0.25, meaning that fewer than $25\%$ of reasoning traces comply with the given instructions. Notably, as task difficulty increases, reasoning instruction following degrades further. We also explore two strategies to enhance reasoning instruction fidelity. (1) multi-turn reasoning and (2) Reasoning Instruction Finetuning (RIF) using synthetic data. RIF improves the IFS of $GPT-OSS-20B$ from 0.11 to 0.27, indicating measurable progress but leaving ample room for improvement.

Figures (11)

• Figure 1: LRMs do not follow instructions in their reasoning traces. (left) A real output from DeepSeek-R1 that fails to follow the 'Word limit' instruction in reasoning while producing the correct answer and (right) an ideal model output. We focus on a model's reasoning trace and investigate how well an LRM follows instructions during reasoning.
• Figure 2: IFS of state-of-the-art LRMs when the instruction’s constraint target is the reasoning trace versus the main response. We evaluate six state-of-the-art LRMs with the same set of questions and instructions for all models, differing only in the constraint target. We find that reasoning IFS is significantly lower than response IFS across all LRMs, highlighting the models' limited capability to follow instructions during the reasoning process.
• Figure 3: Instruction‑type‑wise comparison of IFS when the instruction’s constraint target is (left) the reasoning trace versus (right) the main response. Considering real-world applications, we focus on six instruction types and measure IFS for each instruction. The numbers represent the IFS values, and both heatmaps share the same color scale—dark shades indicate low IFS, while light shades indicate high IFS. Across all six instruction types, reasoning IFS is consistently lower than response IFS. This demonstrates that the key trend in Figure \ref{['fig:overall_comparison']} is consistently observed even at a more granular level.
• Figure 4: Data-source-wise comparison of IFS when the instruction’s constraint target is the reasoning trace (y‑axis) versus the main response (x‑axis) across four LRMs. We consider five different data sources in our dataset, and each point represents a data source. All points lie below the $y=x$ line, indicating that reasoning IFS is lower than response IFS for every dataset. Additional results for other two models are available in Appendix \ref{['app:additional_experimental_results']}.
• Figure 5: Correlation between problem difficulty and reasoning IFS across four LRMs. The black dotted line corresponds to a linear regression fit. For every LRM, we observe a positive correlation, implying that the harder the benchmark dataset, the less faithfully instructions are followed during reasoning. Additional figures for DeepSeek-R1-Distill-Qwen-14B and GPT-OSS-20B are available in Appendix \ref{['app:additional_experimental_results']}.

Theorems & Definitions (3)

• Remark 1: Justification for the Word Limit Choice
• Remark 2: IFS metric
• Remark 3