Large Reasoning Models Fail to Follow Instructions During Reasoning: A Benchmark Study
Research paper
ReasonIF, a benchmark for evaluating instruction-following in reasoning traces of large reasoning models (LRMs).
It finds that frontier LRMs such as GPT-OSS-120B, Qwen3-235B, and 
DeepSeek-R1 fail to follow reasoning instructions more than 75% of the time, and performance degrades with task difficulty.
