CMU–Stanford Paper: Reasoning Models Amplify Behaviors Largely Unrelated to Correctness
A new paper from CMU and Stanford examines whether the behaviors displayed by reasoning models actually relate to producing correct answers, analyzing 15 models across 15,282 reasoning traces. The striking conclusion: behaviors amplified by reasoning training are mostly unrelated to task success, and the models' advantage comes from elsewhere.
Confirmed
- Reasoning models show self-correction, hypothesis testing, and uncertainty acknowledgment far more often than instruction-tuned models, yet these behaviors correlate weakly with success—some even negatively. Per @zainhas, this suggests surface-level signals are unreliable for judging output quality
- Confidence calibration and knowledge alignment—the strongest predictors of correctness—along with self-awareness were barely amplified
- RL chain-of-thought models still typically outperform their instruction-tuned counterparts in accuracy
- The models' core advantage is recovery from failure: in step-by-step extended reasoning tasks, they recover 2-3x faster than non-reasoning models
- Per @Jeanded, RL-learned behavioral strategies are not always the ones most correlated with accuracy, so high accuracy may mask inefficient reasoning paths
Why it matters
The study challenges the intuition that visible chain-of-thought behaviors cause success, suggesting high accuracy can rest on inefficient or ineffective reasoning paths. Aligning RL reward signals with genuinely effective behaviors may be key to improving reasoning efficiency and training.
2026-08-26 ~ 2026-08-27 · 8 related posts
Primary sources
- CMU/Stanford paper: thinking models amplify self-correction but barely amplify success-linked behaviors — Jeande_d ·
- Paper: Self-correction and uncertainty aren't signals of correct answers in reasoning models — zainhas ·
- Reasoning Models Outperform via Higher Recovery Rates, Not Just "More Thinking" — Jeande_d ·
- 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
- Study: Reasoning models' self-correction does not guarantee correctness — zainhas · 2026-08-26
- Paper reveals reasoning models' 'thinking' behaviors are often uncorrelated with correct answers — Jeande_d · 2026-08-27
1 near-duplicate retellings: zainhas