When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents
Researchers measure intent drift in LLM agents and show explicit state tracking only partly closes the gap.
The paper studies intent drift, where superseded user requests still influence an LLM agent's final answer or tool action. IntentFlux converts verifiable tasks into multi-turn dialogues with controlled intent changes while keeping the original graders. Across 627 calibration cases, mean task score fell from 0.476 to 0.384 as superseded information increased, and eight models were less often fully correct than on single-turn prompts. StateForge, which maintains active requirements before generation, raised General-Test mean score from 0.367 to 0.467, but even a ground-truth final state did not match single-turn performance.
- IntentFlux turns graded tasks into dialogues with controlled intent changes.
- Mean score fell from 0.476 to 0.384 as withdrawn intent increased.
- Eight models scored lower on evolving dialogues than single-turn prompts.
- StateForge lifted General-Test mean score from 0.367 to 0.467.
- Ground-truth state still failed to match single-turn performance.
From the original article
1 min read at huggingface.co
LLM agents often operate over multi-turn interactions in which user intent changes before execution. We study intent drift: the failure mode in which superseded parts of the user's intent continue to influence the final answer or tool action. We introduce IntentFlux, an executable benchmark that converts verifiable tasks into dialogues with controlled intent changes while preserving their original graders.
By Yanjie Zhang, Bowen Cao, Zixin Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.