CMU-Stanford paper finds reasoning models amplify behaviors weakly linked to correctness

A new paper from a CMU–Stanford team examines how the reasoning behaviors of "thinking models" (reasoning models) relate to final correctness, reaching a striking conclusion: most behaviors amplified in reasoning models have little to do with task success—their advantages come from elsewhere. As relayed by @Jeanded, models show massive amplification of behaviors like self-correction, hypothesis testing, and acknowledging uncertainty, yet these correlate weakly (sometimes negatively) with answering correctly; what actually predicts correctness are behaviors such as confidence calibration, knowledge alignment, and self-awareness, which are barely amplified at all. From an RL perspective, this means the strategies reinforcement learning teaches models are not always the ones most tied to accuracy gains, and high accuracy may mask inefficiencies in the reasoning path.

Confirmed

• Thinking models amplify self-correction, hypothesis testing, and uncertainty acknowledgment, but these behaviors correlate weakly or even negatively with success
• Behaviors genuinely tied to success—confidence calibration, knowledge alignment, and self-awareness—are hardly amplified
• Even so, RL chain-of-thought models still typically outperform instruction-tuned models in accuracy
• The core advantage of reasoning models lies in "recovery after failure": on extended reasoning tasks requiring step-by-step computation, reasoning models recover from failures 2-3 times faster than non-reasoning models

Why it matters

The study challenges the intuition that visible chain-of-thought behaviors are what cause success, suggesting that high accuracy may rest on inefficient or even ineffective reasoning paths. The authors hope these findings help understand and improve how reasoning models are trained—if RL doesn't necessarily teach the most effective reasoning strategies, aligning RL rewards with genuinely effective behaviors may become key to improving reasoning efficiency.

2026-08-26 ~ 2026-08-26 · 5 related posts

• [source] Study Finds Reasoning Models' Amplified Behaviors Weakly Linked to Correctness — Jeande_d · 2026-08-26
• [source] Reasoning Models Outperform via Higher Recovery Rates, Not Just "More Thinking" — Jeande_d · 2026-08-26
• Study Finds Reasoning Models Amplify Ineffective Behaviors Over Success-Linked Ones — Jeande_d · 2026-08-26
• [source] CMU/Stanford paper: thinking models amplify self-correction but barely amplify success-linked behaviors — Jeande_d · 2026-08-26
• Paper: RL Does Not Always Teach Models the Most Effective Reasoning Strategies — Jeande_d · 2026-08-26