You Changed Your Mind, The Model Didn't: Demystifying Intent in Multi-Turn Dialogue
Intent-Eval shows rejected or replaced user changes can still derail multi-turn LLM task execution.
The paper finds that merely mentioning a change a user later rejects can derail multi-turn task execution even when the final intent is unchanged. Intent-Eval measures this across tool actions, code, databases, and mathematics, showing models treat mentioned content as active requirements after it is rejected or replaced. Accuracy degradation can deepen or persist as interaction continues. Intent-OPSD uses decision-conditioned on-policy self-distillation, with a frozen teacher supervising the student to follow only the active user decision.
- Rejected proposals and superseded requirements still alter model behavior.
- Intent-Eval covers tools, code, databases, and mathematics.
- Accuracy loss can deepen or persist across later turns.
- Intent-OPSD distills active-intent supervision from a frozen teacher.
- Products
- Intent-EvalIntent-OPSD
From the original article
1 min read at huggingface.co
When a large language model handles a multi-turn task and a user proposes a change but ultimately rejects it, the model should continue as if nothing changed. We find a surprising failure: merely mentioning a rejected change can derail task execution, even when the user's final intent remains unchanged.
By Junle Chen, Wei Chen, Zhengjun Huang · 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.