Zhiqi Wang

Lost in Compaction: Evaluating Side-Constraint Loss under Context Compaction

This work identifies a class of user-issued instructions, Session Constraints, that are meant to constrain LLM's behavior for the remainder of a session but are silently dropped during compaction, and introduces COMPINT, an evaluation suite that evaluates compactors across three long-context scenarios: multi-turn chat, agentic trajectory, and long-horizon research.