28 Nov 2026

PG Seminar (CSE-BUET): Action-Aware Tree Search for Autonomous Agents in Web Environment

Abstract: Large language model (LLM) based web agents usually act greedily, choosing one action at a time without considering alternatives. On the web this is risky: the agent sees only part of the website, and a single wrong action can be hard to undo. Tree search lets an agent explore alternatives and backtrack, but returning to an earlier page is not straightforward, because websites change between visits and replaying earlier actions can repeat updates to stored data, such as placing an order twice. This thesis introduces WebOperator, a best-first tree-search framework that ranks candidate actions by predicted progress while considering whether they can be undone. It explores reversible actions first, postpones actions that may change stored data, re-roots the search after a detected change, and checks that a rebuilt page still supports the intended next action before replaying it. With gpt-4o, WebOperator achieves 54.6% task success on WebArena, outperforming all prior tree-search methods using the same backbone, and is also evaluated on live websites from WebVoyager. An ablation shows that naive tree search can lower performance, whereas the full action-aware design improves it. The thesis further introduces WebReplayBench, a web environment that records which actions actually change the website's state, to evaluate whether WebOperator detects irreversible actions and avoids repeating them during backtracking.

 

Presenter: Mahir Labib Dihan (Std No. 1024052050)

Venue: Graduate Seminar Room