CM-USR: Critical-Missing Calibrated User-State Regeneration for Lost-in-Conversation

Abstract

Large language models (LLMs) often solve a fully specified instruction but fail when the same evidence arrives over multiple turns, a failure known as Lost-in-Conversation (LiC). We present CM-USR, a multi-agent inference-time systems that combines Critical-Missing (CM) calibration with User State and Regeneration (USR). A User State Agent records user-grounded facts, a Critical-Missing Gate Agent controls answer readiness through a Wait Policy, and a Regeneration Agent constructs a clean context for the Task Execution Agent. Across instructions in mathematics, API actions, and database querying and three LLMs, CM-USR improves mean performance over sharded dialogue by 19.47 points, reduces unreliability by 18.78 points. Ablations show that clean regeneration is especially important for action tasks, while the gate principally regulates answer timing. These results position explicit state, calibrated waiting, and contamination-free regeneration as complementary inference-time controls for reliable multi-turn reasoning.