Overview
Research prototype asking whether a browser agent uses fewer tokens when it receives only the page information it needs. Keeps full browser state in the controller, shows the model a compact working set, and reveals more on demand. Tested on 20 paired Korean Online-Mind2Web tasks with GPT-5.4-mini.
Technologies
Features
- Adaptive Disclosure Policy - Starts from a small grounded candidate set; budget grows on undisclosed-reference errors and shrinks after grounded runs.
- Externalized Observation Memory - Full DOM/ARIA state lives in the controller; the model sees a compact summary plus the newest disclosure delta.
- Finding: Element Disclosure ≈ No Gain - +5,522 tokens/task, 95% CI [−43,034, +56,650], p = 0.815, with 4.9 reveal round-trips per task.
- Finding: URL-Attribute Pruning −22% - Dropping per-link
/url:lines cut 36,090 tokens/task, 95% CI [−65,130, −9,150], p = 0.006, with no success drop (11/20 → 12/20 strict).