Reasoning models override explicit instructions and default to familiar solution templates
math_rachel · x · 2026-10-19
From Rachel Thomas's thread, the paper (arXiv:2505.17225) finds reasoning models exhibit "reasoning rigidity": despite explicit user instructions, they override clearly stated conditions and default to habitual reasoning trajectories, leading to wrong conclusions. The authors built an expert-curated diagnostic set from modified AIME, MATH500, and famous puzzles, categorizing contamination into three modes: Interpretation Overload, Input Distrust, and Partial Instruction Attention. The figure shows Qwen3-32B and OpenAI o3 failing modified variants by forcing classic templates. The dataset is publicly released.
Related event: Study identifies "reasoning rigidity" as a distinct LLM failure mode(2 posts)→
More from Research
- Embedding arithmetic demo: cartoon − animation − comic = diagram — mayfer · 2026-10-23
- Omnii Genome: Single Generalist Model Outperforms AlphaGenome Across All Benchmarks — JosephJacks_ · 2026-10-23
- Domingos Cites New Result Claiming RL Is 'Basically Useless' vs Supervised Learning — pmddomingos · 2026-10-23
- Word Arithmetic on Embeddings: "Bathroom Minus Toilet Equals Kitchen" Beats SAE Probing — mayfer · 2026-10-23
- Wolfram Opens Rule 30 Prizes to AIs: $30,000 for Three Open Problems — cephaloform · 2026-10-23
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23