摘要
大型推理模型（LRM）虽在复杂任务中表现优异，但自回归推理导致高延迟。本文系统刻画了小型推理模型（SRM）的能力边界，识别出路径发散、认知过载及恢复无能三类推理风险。为此，提出 TrigReason，一种基于触发器的协作推理框架。该框架通过选择性干预替代连续轮询，仅在战略规划、检测到过度自信或陷入无效循环时激活 LRM。实验表明，TrigReason 在保持精度的同时，显著提升了 SRM 的任务卸载比例，有效降低了延迟与 API 成本。
AI 推荐理由
论文核心研究大小模型协同推理机制，解决路径发散等推理风险。
研究机构
AGI Institute, School of Computer Science, Shanghai Jiao Tong University, Shanghai, China
State Key Laboratory for Novel Software Technology, Nanjing University
Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University, Shanghai, China
Shanghai Key Laboratory for Intelligent Information Processing, Fudan University
School of Artificial Intelligence, Wuhan University
Institute of Computer Science, University of Göttingen, Germany
论文信息