Researchers identify critical information flow failures in large reasoning models and propose StepFlow to fix them

Large reasoning models (LRMs) that generate long chains of thought excel at multi-step math, science, and coding tasks but suffer from unstable and hard-to-interpret behavior

Step-Saliency analysis tool reveals two recurring information-flow failures: Shallow Lock-in (shallow layers over-focus on current steps and ignore earlier context) and Deep Decay (deep layers lose saliency on thinking segments)

StepFlow, a test-time intervention method, adjusts shallow saliency patterns using Odds-Equal Bridge to address the identified reasoning failures

The research demonstrates that understanding and fixing step-to-step information flow is crucial for improving the reliability and interpretability of long-chain reasoning in AI models