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

In three linesStudy showing explicit reasoning (CoT) benefits only specific tasks. Authors propose EDRM, a training-free framework using early-stage entropy dynamics to adaptively route to CoT or direct inference. Across 15 benchmarks and 4 LLMs, EDRM achieves 41–55% token reduction while improving accuracy up to 4.7%.

Summary generated by Claude — human-verified