Effect of grounding scheme on observation representation trends

Ascertain whether the same dependence of web-agent performance on observation representation (HTML versus accessibility tree) and thinking token budget, observed under id-based DOM element grounding, also holds under coordinate-based and other grounding schemes.

Background

All experiments in the paper employ id-based grounding, where actions reference DOM elements by identifiers for execution through Playwright.

Many web-agent systems use alternative grounding schemes (e.g., coordinate-based). It is explicitly stated that it is an open question whether the observed performance trends with HTML versus accessibility tree and thinking token budgets persist under different grounding schemes.