Know When to Stop: Segment-Level Credit Assignment for Reducing Overthinking

• 类型：arxiv
• 标识：2607.00482
• 链接：https://arxiv.org/abs/2607.00482
• 主分类：llm-infra
• 形态：method
• 被引：1
• 被引来源：Semantic Scholar
• S2被引：1
• 影响力被引：0
• TLDR：DASH (Drift Aware advantage SHaping), which assigns segment-level credit based on whether each reasoning segment leads toward or away from correctness, achieves the highest accuracy where overthinking is prevalent while reducing overthinking behaviors and achieving more productive self-correction than baselines.
• 待LLM分类：否
• 标题中文：学会适时停止：面向减少过度思考的片段级信用分配
• TLDR中文：DASH (Drift Aware advantage SHaping) 根据每个推理片段是朝向还是偏离正确性来分配片段级 credit，在 overthinking 普遍的场景下取得最高准确率，同时减少 overthinking 行为，比基线实现更高效的自纠错。
• 成熟度：research
• 场景：推理效率优化、思维链压缩、信用分配
• 来源文件：
• /inbox/tom/_candidates/2026-08-06-agent-rag-longcontext-candidates.json
• [S2 enrich]