Discriminative World Models for web agents

A paper by Kelvin Li and colleagues, posted as an arXiv preprint on 2 September, tackles a weakness in web agents: predicting the next state is not enough to rank alternative actions.

The contribution is predicted-state matching over branching WebArena trajectories, a step toward better test-time action selection for web agents. No institutions are listed on the abstract page.

The limitations are significant: the results are author-reported, and the work has had no peer review, independent reproduction or production-safety assessment.