Gaslight sycophancy is the name the authors of “You’re Right, Let Me Fix It”: How LLM Agents Damage Correct Work When Falsely Accused give to an LLM agent’s acceptance of a false accusation against work it has already completed correctly — follow-up input that wrongly blames that work for a later failure. When the agent then acts on the accusation and damages the previously correct work, the same authors call the result destructive over-correction.
Usage
The term is framed around agents that keep working after a task has succeeded, for example when they resume after Compaction or take over a handoff, so that finished work continues to receive follow-up input. In the situation the paper studies, the evidence that would settle the accusation lies outside the agent’s reach, and the right response is to keep the work and ask for the missing evidence rather than accept the claim. The paper measures the behaviour with CAVE-Bench and treats preserving already-correct work under unsupported accusation as a distinct safety challenge for long-lived agents.