Tree Search for Language Model Agents
Jing Yu Koh, Stephen McAleer, Daniel Fried, Ruslan Salakhutdinov
- 🏛 Institutions
- CMU
- 📅 Date
- July 1, 2024
- 📑 Publisher
- TMLR 2025
- 💻 Env
- Web
- 🔑 Keywords
TLDR
This paper adds inference-time best-first tree search to language-model web agents by searching directly in the environment and guiding expansion with a model-based value function. On top of a GPT-4o baseline it reports a 39.7% relative gain on VisualWebArena and a 28.0% relative gain on WebArena, showing that web-agent performance scales with additional test-time search.
Related papers
- Agent Alpha: Tree Search Unifying Generation, Exploration and Evaluation for Computer-Use AgentsFebruary 3, 2026 · arXiv
- WebOperator: Action-Aware Tree Search for Autonomous Agents in Web EnvironmentDecember 14, 2025 · arXiv
- WALT: Web Agents that Learn ToolsOctober 1, 2025 · ICLR 2026 (Poster)
- LiteWebAgent: The Open-Source Suite for VLM-Based Web-Agent ApplicationsMarch 4, 2025 · NAACL 2025 System Demonstrations
- Attacking Vision-Language Computer Agents via Pop-upsNovember 4, 2024 · ACL 2025
- ExACT: Teaching AI Agents to Explore with Reflective-MCTS and Exploratory LearningOctober 2, 2024 · ICLR 2025 (Poster)