Discriminative World Models for Web Agents
Recent web agents often use "world models" to plan their next moves. By predicting what a website will look like after a specific action, these agents can compare different choices before committing to one. Traditionally, these models are trained to recreate a specific, fixed format—like a snapshot of the website’s code or a text summary. However, this paper argues that this approach is flawed because it focuses on reproducing a format rather than capturing the specific details that actually matter for making a decision. The authors introduce a new training objective called "predicted-state matching," which shifts the focus toward creating representations that help an agent distinguish between the outcomes of different possible actions.
A New Way to Train World Models
Instead of forcing a model to generate a rigid, predefined string, the authors train it to produce a representation that acts as a unique identifier for a specific outcome. To do this, they created a "branching" dataset from existing web-agent trajectories. In this dataset, every decision point includes multiple possible actions and the different resulting states they lead to. During training, the model is rewarded when its predicted representation allows a "judge" to correctly identify the true outcome of an action when presented with an alternative, incorrect outcome. This makes the model's output flexible, allowing it to prioritize the most relevant information needed to differentiate between choices.
Improving Action Selection
The authors tested whether these discriminative representations actually help agents perform better. They integrated their world model into a Process Reward Model (PRM), which is a system used to rank candidate actions. By providing the PRM with these more informative, discriminative representations, the agent could better evaluate the consequences of its potential moves. The results showed that this approach outperformed traditional models that were trained to simply predict fixed-format snapshots, as those snapshots often contained too much irrelevant information or missed the subtle changes that distinguish one action from another.
Better Performance on Web Tasks
The researchers evaluated their method on WebArena-Lite, a benchmark for web-based tasks. They found that using their discriminative world model for test-time action selection led to higher task success rates. By focusing on the differences between potential future states rather than just trying to "draw" a picture of the next page, the agent gained a clearer understanding of how its actions would impact the web environment. This suggests that for complex, multi-step navigation problems, it is more effective for an agent to learn what makes an outcome unique rather than trying to perfectly replicate the entire state of a webpage.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!