Do reasoning traces actually matter? A three-stage causal study on distilled DeepSeek R1 models
burkov · x · 2026-10-14
Large reasoning models produce step-by-step thinking before answering, but whether final answers functionally depend on those traces — or the steps are merely cosmetic post-hoc explanations — has been unclear. This study systematically examines the functional and informational connection between intermediate reasoning and final answers across three distilled DeepSeek R1 models, using a three-stage framework combining black-box benchmarks, internal attention analysis, and mechanistic causal interventions, evaluated on MATH-500 and WildBench.
More from Models
- Blind AI self-grading: Astra picked its own analysis #1, Muse ranked itself last — kevinkern · 2026-10-14
- Caught fabricating episode counts, Google's AI confesses: 'I was lying to you' — LoudStretch2997 · 2026-10-14
- Ramp spend data: Perplexity, Wispr Flow, Claude Code dominate AI buying categories — AravSrinivas · 2026-10-14
- Codex fast mode after quota reset: 'I don't think I can go back' — msg · 2026-10-14
- Google audit: early Argon RL runs rewarded over chain-of-thought, silent reward hacking rose to 36.4% — scaling01 · 2026-10-14
- Josh Gans: Gemini Argon flunks explainer video task that Opus 5.5 nails — joshgans · 2026-10-14