Read More, Think More: Stronger Models Perform Better with Full HTML in Web Agents
A study submitted on April 2, 2026 found that low-capability models benefit from a simplified accessibility tree, while strong models benefit from full HTML layout; increasing thinking tokens further amplifies the advantage of HTML, and diff history can balance information and tokens.
Web agents have long treated HTML redundancy as a problem that must be compressed, but observations indicate there is no one-size-fits-all optimal solution. Model capability, thinking budget, and page history together determine how much raw structure should be provided.
Experiments compared full HTML, accessibility tree, observation history, and diff representations across multiple models and thinking budgets. Error analysis revealed that strong models can leverage layout for action localization, while weak models are more prone to hallucination with long inputs. The conclusion supports dynamically selecting observation format based on model and budget.
Browser agent infrastructure needs to shift from fixed DOM cleaners to configurable observation layers; context compression, action localization, and model routing will become a single performance optimization problem.
Teams should A/B test page representations based on model tier and task difficulty, using completion rate and unit success cost to decide how much DOM to retain, rather than applying uniform maximum compression.
Needs to extend to dynamic pages, visual information, prompt injection, and different website structures, and measure the latency, cost, and security exposure introduced by full HTML.