Discriminative World Models for Web Agents
The paper trains web-agent world models to make predicted outcomes distinguishable across competing actions, not just plausible next states.
Its “predicted-state matching” objective asks the model to identify the real resulting state against states produced by alternative actions at the same decision point. The authors build a branching dataset from WebArena Go-Browse trajectories to support that comparison. They report gains over supervised next-state prediction on a held-out matching benchmark, stronger PRM-style action ranking on WebPRMBench, and improved end-to-end success on WebArena-Lite. ArXiv · AI/CL/LG's note
Its “predicted-state matching” objective asks the model to identify the real resulting state against states produced by alternative actions at the same decision point. The authors build a branching dataset from WebArena Go-Browse trajectories to support that comparison. They report gains over supervised next-state prediction on a held-out matching benchmark, stronger PRM-style action ranking on WebPRMBench, and improved end-to-end success on WebArena-Lite. ArXiv · AI/CL/LG's note
score 5