The week's most disruptive science, explained for humans.

Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought Reasoning

Readable chain-of-thought traces look like a window into why a model answered as it did — but this paper shows that even strong LLM judges cannot reliably tell which steps actually change the chance of a correct answer. A caution for process rewards and critic models, not a product timeline.

• ▸What: The authors measure a reasoning step’s real importance as its advantage — how much including it changes expected reward, estimated with Monte Carlo rollouts — then test whether LLM judges can spot high-advantage steps from the text alone.
• ▸Why it matters (abundance angle): Safer, cheaper cognitive tools need supervision that actually tracks what the model is doing. If critics reward fluent text instead of causal steps, we waste compute and trust on a scarce expert-judgment problem that looks solved.
• ▸Who should care: Alignment researchers, process-reward and critic-model builders, eval teams, and anyone treating CoT traces as interpretability for high-stakes assistance.