Original title: Discriminative World Models for Web Agents
Kelvin Li, Dhruv Pendharkar, Anish Pahilajani, Chuyi Shang +5 · September 2026
Web agents are AI programs that browse the internet to solve tasks. To pick good moves, some agents imagine future web pages and rank them. Normally, these "world models" learn by trying to recreate the exact page, such as its raw code. But this method fails to highlight the small differences an agent needs to spot the best option. To fix this, researchers built a method called predicted-state matching. Instead of copying a whole page, the model learns to tell apart the real future page from pages created by other actions. They trained it on branching web paths with multiple choices at every step. This method beats standard models at predicting states, helps models rank actions better, and boosts overall task success on real web benchmarks.
Imagine packing for a trip where you can only carry one rain jacket, and you need to guess which one will keep you dry. Normally, you might check the forecast, but it's often unclear. Now suppose you had a diary from someone who took the same trip and wrote down exactly which jacket worked. In this paper, the system learns from similar past experiences to tell which action choice will actually lead to a good outcome.
This helps web agents pick the right actions more often, making automated tasks on websites more reliable for users and developers.
The work is tested on specific benchmarks like WebArena-Lite and WebPRMBench. It doesn't show performance on all types of web tasks or real-world deployment. The approach relies on having a branching dataset of actions and outcomes.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Discriminative World Models for Web Agents — Kelvin Li, Dhruv Pendharkar, Anish Pahilajani et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.