IntentScore: Intent-Conditioned Action Evaluation for Computer-Use Agents

TL;DR

IntentScore is a plan-aware reward model that scores candidate GUI actions before they are executed — catching the irreversible mistakes that computer-use agents would otherwise cascade through subsequent steps. Trained on 398K offline GUI interaction steps across three operating systems.

Key contributions:

• A plan-aware architecture that embeds each candidate’s planning intent in the action encoder, discriminating between candidates with similar actions but different rationales.
• Two complementary training objectives — contrastive alignment for state-action relevance and margin ranking for action correctness — reaching 97.5% pairwise discrimination accuracy.
• Generalization to unseen agents and tasks: deployed as a re-ranker for Agent S3 on OSWorld (entirely unseen during training), improving task success rate by 6.9 points.

BibTeX

@misc{chen2026intentscore,
      title={IntentScore: Intent-Conditioned Action Evaluation for Computer-Use Agents},
      author={Rongqian Chen and Yu Li and Zeyu Fang and Sizhe Tang and Weidong Cao and Tian Lan},
      year={2026},
      eprint={2604.05157},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2604.05157},
}