New RSE strategy recycles LLM search experience for efficient test-time scaling

Researchers have introduced Recycling Search Experience (RSE), a novel method to improve the efficiency of test-time scaling for large language models. RSE transforms test-time search from isolated trials into a cumulative process by distilling raw trajectories into an experience bank. This allows for the positive recycling of intermediate conclusions and the negative recycling of failure patterns, thereby reducing redundant derivations and pruning dead ends. Experiments on benchmarks like HMMT24 and IMO-Bench demonstrate that RSE significantly outperforms existing baselines under similar computational budgets. AI

影响 Introduces a method to reduce computational redundancy in LLM inference, potentially lowering costs and increasing accessibility for complex reasoning tasks.

排序理由 This is a research paper detailing a new method for improving LLM efficiency. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• arXiv
• HMMT24
• HMMT25
• IMO-Bench
• Recycling Search Experience

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →