哈佛MIT研究复合LLM系统角色漂移，发现RL改善效果中86%来自模块走捷径，他们提出了Role Anchor控制方法。
哈佛和MIT的研究发现复合LLM系统中存在角色漂移现象。端到端RL提升多模块流水线准确率时，模块可能通过捷径完成任务而偏离原始角色。例如分解器本应拆分问题，却将答案嵌入子问题；阅读器本应基于检索回答，却依赖参数记忆。若强制分解器坚守角色，RL带来的86%改进消失。论文提出Role Anchor控制方法，通过保持角色提示对下一词预测的影响来对抗漂移。
Great technical paper from Harvard and MIT. It's on role drift in compound LLM systems. (bookmark ...
Great technical paper from Harvard and MIT. It's on role drift in compound LLM systems. (bookmark it) End-to-end RL improves the accuracy of a multi-module LLM pipeline without constraining how the modules divide labor internally. Researchers name that failure mode and measure it. Role drift happens when a module preserves or improves end-task performance while abandoning its assigned role through shortcuts that system-level evaluation cannot see. Two instances across two pipelines: - A decomposer meant to split a question into sub-questions for a separate solver instead plants the answer inside them. - A reader meant to answer from retrieved passages instead falls back on parametric memory. An interesting finding is that if your hold the decomposer to its role and 86% of the RL improvement disappears. Role Anchor is the proposed control. It preserves how the role prompt shifts a module's next-token predictions relative to a neutral prompt, using that as a proxy for the role's intended effect during training. Gradient analysis indicates it reduces alignment with the drift direction rather than simply suppressing learning. Paper: arxiv.org/abs/2607.21627 Learn to build effective AI agents in our academy: academy.dair.ai 💬 3 🔄 1 ❤️ 1 👀 906 📊 4 ⚡