arXiv cs.AIMonday · May 25, 2026FREE

When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions

A new paper from arXiv (cs.AI) titled 'When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions' reveals that chain-of-thought (CoT) reasoning is not universally beneficial. The authors show that CoT can provide marginal or negative gains on factual and open-ended tasks while multiplying token consumption. Through systematic analysis, they find that early-stage entropy dynamics reliably signal whether a task benefits from CoT: tasks that benefit exhibit consistent entropy reduction, while others show unstable or increasing patterns. This behavior is interpreted as a phase transition from a high-entropy exploratory regime to a low-entropy structured reasoning regime. Based on these insights, the authors propose EDRM (Entropy Dynamics-based Reasoning Manifold), a lightweight, training-free routing framework that leverages early decoding entropy to adaptively select inference strategies. EDRM embeds entropy trajectories to decide when to apply CoT, potentially reducing token usage on tasks where reasoning is unnecessary. The paper is available on arXiv under ID 2605.22873, published May 25, 2026.