WebOperator: Action-Aware Tree Search for Autonomous Agents in Web Environment

WebOperator: Action-Aware Tree Search for Autonomous Agents in Web Environment WEBOPERATOR: ACTION-AWARE TREE SEARCH FOR AUTONOMOUS AGENTS IN WEB ENVIRONMENT Mahir Labib Dihan∗1 Tanzima Hashem1 Mohammed Eunus Ali2 Md Rizwan Parvez3 1Department of Computer Science and Engineering Bangladesh University of Engineering and Technology (BUET) 2Faculty of Information Technology, Monash University 3Qatar Computing Research Institute (QCRI) {dihan, tanzimahashem}@cse.buet.ac.bd, eunus.ali@monash.edu, mparvez@hbku.edu.qa ∗Work done when working as a remote RA at QCRI. ABSTRACT LLM-based agents often operate in a greedy, step-by-step manner, selecting ac- tions solely based on the current observation without considering long-term con- sequences or alternative paths. This lack of foresight is particularly problematic in web environments, which are only partially observable—limited to browser- visible content (e.g., DOM and UI elements)—where a single misstep often re- quires complex and brittle navigation to undo. Without an explicit backtracking mechanism, agents struggle to correct errors or systematically explore alternative paths. Tree-search methods provide a principled framework for such structured exploration, but existing approaches lack mechanisms for safe backtracking, mak- ing them prone to unintended side effects. They also assume that all actions are reversible, ignoring the presence of irreversible actions—limitations that reduce their effectiveness in realistic web tasks. To address these challenges, we intro- duce WebOperator, a tree-search framework that enables reliable backtracking and strategic exploration. Our method incorporates a best-first search strategy that ranks actions by both reward estimates and safety considerations, along with a robust backtracking mechanism that verifies the feasibility of previously vis- ited paths before replaying them, preventing unintended side effects. To further guide exploration, WebOperator generates action candidates from multiple, var- ied reasoning contexts to ensure diverse and robust exploration, and subsequently curates a high-quality action set by filtering out invalid actions pre-execution and merging semantically equivalent ones. Experimental results on WebArena and WebVoyager demonstrate the effectiveness of WebOperator. On WebArena, Web- Operator achieves a state-of-the-art 54.6% success rate with gpt-4o, underscoring the critical advantage of integrating strategic foresight with safe execution. All our resources are open-sourced at https://kagnlp.github.io/WebOperator/. 1 INTRODUCTION LLM-based WebAgents are increasingly applied to automate complex web interactions, ranging from form filling and content retrieval to multi-step workflows over dynamic pages (Deng et al., 2024). However, planning and executing such tasks remains challenging due to unique characteris- tics of web environments, such as being partially observable: the agent can access the current page’s DOM, UI elements, and visible content, but has no direct access to hidden server-side state or the broader global context. Despite these challenges, conventional WebAgents operate in a greedy, step-by-step manner, select- ing actions based solely on the current observation, without accounting for long-term consequences or alternative strategies (Ning et al., 2025). While off-the-shelf models for estimating action use- fulness are available, they are inherently short-sighted and imperfect (Chae et al., 2025) and this myopic approach is particularly fragile in web-like, partially observable environments, where a sin- 1 arXiv:2512.12692v1 [cs.AI] 14 Dec 2025 WebOperator: Action-Aware Tree Search for Autonomous Agents in Web Environment Web Browser Set my gitlab status as "Resting due to leg injury". 1 Navigate to Start Page Observe  Environment 2.1 AxTree Agent Reward Model Risk-Aware Selection Status has been set to 'Resting due to leg injury' Reward  Actions 4 Execute Selected action 9 Merge Actions 5 Select Best Action from Search tree 7 Generate Candidate  Actions 3 Thought 1: To set a status, we must first navigate to the user's profile area where the status widget is available.  Action 1: click('201') Thought 2: To set the GitLab status we must open the status editor.  Action 2: click('215') Thought 3: To change the status, we need to replace the textbox content with “Resting due to leg injury” and submit the form. Action 3: fill('1067', 'Resting due to leg injury', True) ...... Thought N: The desired status text is already shown in the UI, confirming it has been applied. No further interaction is required. Action N: stop("Status has been set to 'Resting due to leg injury'.") Solution Trajectory Add Actions to Search Tree 2.2 Add Observation  to Search Tree Action Space Observation Modify Context Variation N Open New Tab Rollback to  Nearest Checkpoint Observation Mismatch? Y Commit Close Old Tabs Close New  Tabs Backtrack Complete? Y Discard N Observation Replay Action Speculative Ba

This content is AI-processed based on open access ArXiv data.