LLM agents frequently work across multi-turn conversations in which a user's intent shifts before the agent acts. A new Hugging Face research paper names this failure mode intent drift: the tendency of superseded or withdrawn parts of a user's intent to keep influencing the final answer or tool action.
To measure the problem, the authors built IntentFlux, an executable benchmark that turns verifiable tasks into dialogues with controlled intent changes while keeping the original task graders intact. In a 627-case calibration, the mean task score dropped from 0.476 to 0.384 as dialogues accumulated more superseded and withdrawn information.
Testing across eight models, the researchers found that the rate of fully correct solutions was significantly lower when the same final task had to be recovered from an evolving dialogue rather than supplied directly in a single turn.
As a mitigation, the paper proposes StateForge, which explicitly maintains the active requirements before generation. On the General-Test set, it raised the mean task score from 0.367 to 0.467.
Even when the ground-truth final state was provided, performance improved further but still fell short of single-turn results — suggesting that state-estimation errors account for only part of the gap. The authors conclude that intent drift is a measurable multi-turn failure mode and that explicit state maintenance offers a partial fix.