Research map / Multi-agent systems and games / LLM agents
WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WebPilot, a multi-agent MCTS-based web agent with global plan refinement and local tailored MCTS, reaching 37.2% success on WebArena with GPT-4o.
Problem
LLM web agents struggle with complex, dynamic web tasks due to rigid expert-designed policies; classical MCTS struggles with vast action spaces, unpredictable transitions, and incomplete information in web environments.
Approach
A dual optimization strategy: Global Optimization decomposes tasks into subtasks and refines the plan via reflection (Planner, Controller, Extractor); Local Optimization runs a tailored MCTS per subtask with goal-oriented PUCT selection, reflection-enhanced one-action expansion, granular bifaceted 0-10 LLM self-reward, and maximal value backpropagation.
Key results
- WebPilot with GPT-4o achieves 37.2% success rate on WebArena, versus 19.2% for LM-Tree Search (93% relative increase) and 33.5% for SteP.
- With GPT-3.5, WebPilot reaches 29.1% on WebArena, versus 8.9% for the WebArena baseline.
- On MiniWoB++ (43 text-only tasks), WebPilot achieves 95.6% SR, slightly below SteP's 96.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WebPilot instead of LM-TS | WebArena with GPT-4o · WebArena average | success rate: 19.2 -> 37.2; 93% relative increase | Helped |
| Use WebPilot instead of SteP | WebArena with GPT-4o · WebArena average | success rate: 33.5 -> 37.2 | Helped |
| Use WebPilot instead of SteP | WebArena · WebArena GitLab | success rate: 31.7 -> 39.4; 7.7% | Helped |
| Use WebPilot instead of SteP | MiniWoB++ · MiniWoB++ | success rate: 96.0% -> 95.6% | Hurt |
| Use GPT-4o instead of GPT-3.5 | WebPilot · WebArena Shopping | success rate: 11.8% | Helped |
| Use GPT-4o instead of GPT-3.5 | WebPilot · WebArena Reddit | success rate: 6.6% | Helped |
| Use GPT-4o instead of GPT-3.5 | WebPilot · WebArena GitLab | success rate: 9.4% | Helped |
| Add Child Reflection (ablation) | WebPilot · WebArena IS tasks | success rate: 74% -> 100% | Helped |
| Add Child Reflection (ablation) | WebPilot · WebArena WI tasks | success rate: 70% -> 100% | Helped |
| Add Sibling Reflection (ablation) | WebPilot · WebArena IS tasks | success rate: 72% -> 100% | Helped |
| Add Sibling Reflection (ablation) | WebPilot · WebArena WI tasks | success rate: 60% -> 100% | Helped |
| Add Controller (ablation) | WebPilot · WebArena IS tasks | success rate: 86% -> 100% | Helped |
Surprising findings
- On MiniWoB++, WebPilot does not beat SteP, which the authors attribute to task simplicity limiting the benefit of exploration.
Limitations
- Performance bounded by LLM capabilities in understanding and interacting with web environments via text.
- Text-only actree observations lack visual cues.
- Evaluated on 43 of MiniWoB++ tasks and WebArena only; gains on MiniWoB++ are small.
Experimental setup
Datasets: WebArena MiniWoB++
Benchmarks: WebArena MiniWoB++
Metrics: success rate
Baselines: WebArena baseline SteP LM-Tree Search RCI AdaPlanner zero-shot agent
Builds on: Monte Carlo Tree Search UCT PUCT (AlphaGo) Reflexion Chain-of-thought
Closest papers by meaning
- REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.89)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.89)
- WebArena: A Realistic Web Environment for Building Autonomous Agents (ICLR 2024 · similarity 0.88)
- A Subgoal-driven RL Framework for Improving Long-Horizon Web Agents (NeurIPS 2026 · similarity 0.88)
- Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration (CVPR 2026 · similarity 0.88)
- WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.