When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents

When Users Change Their Minds. Measuring and Repairing Intent Drift in LLM Agents. By Yanjie Zhang and three others, from the Hong Kong University of Science and Technology, the Chinese University of Hong Kong, and one other institution. Published on the twenty-sixth of September. No code repository is linked to it.

LLM agents usually work through conversations of several turns, and users change their minds along the way, before anything is executed. The paper names the resulting failure intent drift: parts of what the user wanted earlier, and has since replaced, keep shaping the final answer or the tool action. A task handed over cleanly in one turn does not test for this at all.

IntentFlux is the benchmark built to measure it, and it is executable. It takes tasks that can already be verified and turns them into dialogues with controlled changes of intent, keeping the original graders. The final task is the same one a grader already knows how to check; only the route to it now winds through superseded and withdrawn requests.

Because the graders are preserved, the comparison is clean. The same final task can be posed directly in a single turn, or recovered from an evolving dialogue, and scored the same way. A calibration of six hundred and twenty-seven cases varies how much superseded and withdrawn information the dialogues contain.

In that calibration, mean task score falls from zero point four seven six to zero point three eight four as the dialogues carry more outdated material. Across eight models, the rate of fully correct solutions is significantly lower when the task must be pulled out of an evolving dialogue than when it is given directly.

The authors then try a repair. StateForge explicitly maintains the active requirements before generation. On General-Test it lifts mean task score from zero point three six seven to zero point four six seven. A further experiment hands the model the ground-truth final state. That helps more, but it still does not recover single-turn performance, which suggests that errors in estimating the state account for only part of the gap.

Two results, then, of different strength. Intent drift is established as a failure mode that can be measured, and keeping the current requirements explicit is shown to reduce it. Yet even a perfect record of what the user now wants leaves a shortfall, so part of the problem lies somewhere other than tracking.