EN Submit a tool
Paper

Harvard and MIT Reveal "Role Drift" Problem in Compound LLM Systems

Published: Source: X: Elvis Saravia (@omarsar0, DAIR.AI)

ShareXFacebookTelegramWhatsApp

Harvard and MIT Reveal "Role Drift" Problem in Compound LLM Systems

A study by Harvard and MIT found that modules in compound LLM systems can experience "role drift" under end-to-end reinforcement learning: the decomposer embeds answers directly into sub-questions to improve task accuracy, while the reader relies on its own parametric memory rather than retrieved content. If the decomposer is forced to stick to its original role, 86% of the RL performance gains disappear. The paper proposes a "role anchoring" method that suppresses drift by constraining the module's prediction distribution.

Read the original (opens in a new tab)

News stream data aggregated by AI HOT

Related newsLatest in this category
· Hacker News Hot (buzzing.cc Chinese translation)
· The Decoder: AI News (RSS)
· X: Rohan Paul (@rohanpaul_ai)
· X: Rohan Paul (@rohanpaul_ai)
· X: Xiaohu (@xiaohu)