COLM paper: thinking models amplify the wrong reasoning behaviors, 15,282 traces show
Jeande_d · x · 2026-10-04
A COLM 2026 paper, 'Amplified Does Not Mean Predictive,' asks which reasoning behaviors actually correlate with correct answers and whether reasoning training amplifies them. Introducing Behavioral Lift — how much correctness changes when a behavior appears in a trace — the authors annotate 15,282 traces across 15 models and 6 benchmarks (text and vision-language). They find an Amplification-Lift Gap: thinking models strongly amplify self-correction, hypothesis testing and uncertainty acknowledgment (the latter 3-7x amplified yet weakly or negatively tied to correctness), while the highest-lift behaviors — confidence calibration, knowledge alignment, self-awareness — are barely amplified. Confidence calibration is among the strongest positive correctness signals in both modalities. The takeaway: process-level objectives should reward calibrated, grounded reasoning rather than surface deliberation. Poster session Oct 8 at COLM.
More from Models
- Rumor: Anthropic's Opus 5.5 was a massive leap that blindsided OpenAI's plans — haider1 · 2026-10-04
- Theo's verdict: Sol 6.1 replaces Sonnet for review, but Opus 5.5 stays his daily driver — airesearch12 · 2026-10-04
- Anthropic reportedly saved its strongest model, Fable 5.5, for just before its IPO — thesaraharminta · 2026-10-04
- GPT-6 guardrails quietly swap tasks instead of refusing, researcher warns — moyix · 2026-10-04
- PewDiePie Says OpenAI Banned Him Twice While Distilling GPT Into a Local Model — ccerrato147 · 2026-10-04
- Pretraining Still Matters: RL Is Just More Compute-Intensive, Not More Important — akbirthko · 2026-10-04