{"id":44068,"date":"2025-10-13T07:30:53","date_gmt":"2025-10-13T07:30:53","guid":{"rendered":"https:\/\/youzum.net\/swireasoning-entropy-driven-alternation-of-latent-and-explicit-chain-of-thought-for-reasoning-llms\/"},"modified":"2025-10-13T07:30:53","modified_gmt":"2025-10-13T07:30:53","slug":"swireasoning-entropy-driven-alternation-of-latent-and-explicit-chain-of-thought-for-reasoning-llms","status":"publish","type":"post","link":"https:\/\/youzum.net\/ja\/swireasoning-entropy-driven-alternation-of-latent-and-explicit-chain-of-thought-for-reasoning-llms\/","title":{"rendered":"SwiReasoning: Entropy-Driven Alternation of Latent and Explicit Chain-of-Thought for Reasoning LLMs"},"content":{"rendered":"

SwiReasoning is a decoding-time framework that lets a reasoning LLM decide when to think in latent space<\/strong> and when to write explicit chain-of-thought<\/strong>, using block-wise confidence estimated from entropy trends in next-token distributions<\/strong>. The method is training-free<\/strong>, model-agnostic<\/strong>, and targets Pareto-superior<\/strong> accuracy\/efficiency trade-offs on mathematics and STEM benchmarks. Reported results show +1.5%\u20132.8%<\/strong> average accuracy improvements with unlimited tokens and +56%\u201379%<\/strong> average token-efficiency gains under constrained budgets; on AIME\u201924\/\u201925, it reaches maximum reasoning accuracy earlier<\/strong> than standard CoT.<\/p>\n

What SwiReasoning changes at inference time?<\/strong><\/h3>\n

The controller monitors the decoder\u2019s next-token entropy<\/strong> to form a block-wise confidence<\/strong> signal. When confidence is low (entropy trending upward), it enters latent reasoning<\/strong>\u2014the model continues to reason without emitting tokens<\/strong>. When confidence recovers (entropy trending down), it switches back to explicit reasoning<\/strong>, emitting CoT tokens to consolidate and commit to a single path. A switch count control<\/strong> limits the maximum number of thinking-block transitions<\/strong> to suppress overthinking before finalizing the answer. This dynamic alternation is the core mechanism behind the reported accuracy-per-token gains. <\/p>\n

\n
\"\"
https:\/\/arxiv.org\/pdf\/2510.05069<\/figcaption><\/figure>\n<\/div>\n

Results: accuracy and efficiency on standard suites<\/strong><\/h3>\n

It reports improvements across mathematics and STEM reasoning tasks:<\/p>\n