Discriminative World Models for Web Agents

What it claims

The abstract adds two specifics the catalog summary omits: the exact baselines the method beats, and a third evaluation. Beyond the WebPRMBench and WebArena-Lite results the summary names, the paper reports a held-out predicted-state matching benchmark where discriminatively-trained world models outperform ones trained with supervised next-state prediction.

Evidence

The abstract states three comparisons, all qualitative (no numbers or sizes are given on the page):

• Held-out predicted-state matching benchmark: the discriminative objective beats supervised next-state prediction.
• WebPRMBench (PRM-style action ranking): the method is compared against two baselines — action-only PRMs (no world model at all) and PRMs augmented with supervised-next-state world models (the naive world-model approach). This two-baseline framing is what distinguishes the claim: the gain is over both no world model and a conventionally-trained one.
• WebArena-Lite (end-to-end): using the world model for test-time action selection improves task success.

How it works

The training data is a branching dataset derived from WebArena Go-Browse trajectories, where every decision point holds multiple alternative actions and their resulting states. That branching structure is what makes predicted-state matching trainable: the objective needs contrasting alternative-action outcomes at each node to learn to distinguish the true next state from the counterfactual ones.

Limits

The page is abstract-only. No metric values, dataset sizes, model sizes, compute figures, or the numeric margins on any of the three benchmarks are given — the improvements are asserted, not quantified. A project page is linked but its contents are not on this page. Check that page (or the PDF) for the actual numbers before deciding whether the gains justify building the branching dataset and swapping objectives.