Daily AI Paper

Web agents are getting better at navigating complex sites, but they still struggle with one very practical problem: deciding which action is actually best before they click. This paper looks at the world models used inside those agents, which predict what the page will look like after a candidate action and then help rank the options. The authors argue that standard training is misaligned, because predicting the next webpage accurately is not the same as making the useful differences between actions easy to tell apart. Their fix is a new training objective called predicted-state matching, which teaches the model to represent the true outcome of an action in a way that clearly distinguishes it from the outcomes of alternative actions. That makes the world model more helpful for ranking, improves action selection in benchmarks, and boosts end-to-end success on real web tasks. In short, it is a simple idea with broad impact: if an agent can better imagine the consequences of its choices, it can make smarter decisions online.