Analysis

Reasoning Fails in Large Models

Large reasoning models perform well on multi-step tasks but have unstable behavior. Step-Saliency analysis reveals information-flow failures.