「We explicitly define the Reasoning Completion Point and propose a clear analytical framework for determining the optimal termination point in reasoning. Leveraging feature importance analysis via CatBoost models, we further mine common signals indicative of reasoning completion, from which we distill a concise and efficient set of heuristic rules named Reasoning Completion Point Detection. This rule set precisely identifies RCP with minimal computational over- head, effectively mitigating overthinking in LLMs.」とのこと。トークン数削減だけなく、推論性能上も有利なのが興味深い。