Proposer-Agent-Evaluator (PAE): Autonomous Skill Discovery For Foundation Model Internet Agents
Yifei Zhou, Qianlan Yang, Kaixiang Lin, Min Bai, Xiong Zhou, Yu-Xiong Wang, Sergey Levine, Li Erran Li
- 🏛 Institutions
- UC Berkeley, UIUC, Amazon
- 📅 Date
- December 17, 2024
- 📑 Publisher
- ICML 2025 (Poster)
- 💻 Env
- Web
- 🔑 Keywords
TLDR
PAE is a web-agent learning system that lets foundation-model agents autonomously propose tasks, attempt them, and score the resulting trajectories with a VLM-based evaluator. By turning those evaluations into RL signals, it improves zero-shot generalization on unseen websites and tasks for vision-based internet agents.
Related papers
- SkillWeaver: Web Agents can Self-Improve by Discovering and Honing SkillsApril 9, 2025 · arXiv
- Reinforcement Learning on Web Interfaces Using Workflow-Guided ExplorationFebruary 24, 2018 · ICLR 2018 (Poster)
- ClawGUI: A Unified Framework for Training, Evaluating, and Deploying GUI AgentsApril 13, 2026 · arXiv
- DistRL: An Asynchronous Distributed Reinforcement Learning Framework for On-Device Control AgentsOctober 18, 2024 · ICLR 2025 (Poster)
- WebArena-Infinity: Generating Browser Environments with Verifiable Tasks at ScaleMarch 2026 · Blog Post
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web AgentsMarch 5, 2026 · arXiv