Explore Briefly, Then Decide: Mitigating LLM Overthinking via Cumulative Entropy Regulation
This research introduces TECA and CER, novel metrics and methods to regulate entropy in LLM reasoning, significantly reducing response lengths while optimizi...