Language models struggle to track changing user intents accurately
IntentDelta: A Human-Constructed Benchmark for Evolving User Intent
Computation and Language
Summary
People often change their minds when asking computers to help with tasks like shopping or travel planning. The authors created a dataset called IntentDelta that shows how user intentions evolve through conversations. They tested various computer programs to see how well they could understand these shifting goals and act correctly. The results show that these programs often miss important changes in what users want, especially when preferences are soft or hard constraints. Even when the programs understand the user's current intent, they might still fail to take the right actions.
What this means in practice
- •For conversational ai developers: Improve testing processes by using IntentDelta to evaluate how well agents adapt to evolving user intents during multi-turn tasks like shopping and travel planning.
- •For customer service platform teams: Incorporate IntentDelta data to identify gaps in automated assistants’ abilities to track changing customer needs and improve response accuracy.
Authors
Jieyu Zhou, Zijian Wang, Yun Wang
Abstract
Language-model agents must adapt as users revise their requirements and priorities during complex tasks. We introduce IntentDelta, a human-constructed multi-turn benchmark with 638 trajectories and 3,766 turns across WebShop and TravelPlanner. Each turn pairs a user utterance with a cumulative intention state and human reference actions. The trajectories are cross-verified and cover feasibility- and preference-driven changes in constraints and priorities. We evaluate eleven agents on intention recovery and action quality. Average Intention Recovery is 33.1\% in TravelPlanner and 44.7\% in WebShop, while average Action Score is 31.6% and 41.0%, respectively. Agents struggle to preserve earlier requirements and distinguish soft preferences from must-have constraints. Correct intention recovery also does not guarantee successful execution, highlighting challenges in translating evolving user intentions into actions.