Web agents built on vision-language models typically decide each action by generating text like click(12). A new arXiv paper points out that most of that text is redundant: it simply selects from a small set of action types and a few dozen on-page elements. The authors argue that generating full text for every step is an inefficient way to make such a limited choice.
Their proposed alternative is to attach plug-in decision heads that read the choice directly from the model's internal representations, bypassing text generation. The abstract is truncated, so the full method and experimental results are not yet visible, but the motivation is clear: if the decision is just a selection among a few options, a dedicated head should be faster and cheaper than writing out a command.
The paper is a single source, and because the abstract cuts off mid-sentence, the article cannot confirm any reported performance gains. What is available is a focused critique of current practice and a concrete architectural suggestion for making web agents more efficient.