Inducing Programmatic Skills for Agentic Tasks
Zora Zhiruo Wang, Apurva Gandhi, Graham Neubig, Daniel Fried
- 🏛 Institutions
- CMU, Microsoft
- 📅 Date
- April 9, 2025
- 📑 Publisher
- COLM 2025
- 💻 Env
- Web
- 🔑 Keywords
TLDR
This paper proposes Agent Skill Induction (ASI), which learns executable program-based skills online from web interaction experience and reuses them as tasks evolve. The use of programs makes skill induction verifiable, improving both WebArena success rate and step efficiency over static agents and text-skill baselines while also supporting cross-website transfer.
Related papers
- ColorBrowserAgent: Complex Long-Horizon Browser Agent with Adaptive Knowledge EvolutionJanuary 12, 2026 · arXiv
- SteP: Stacked LLM Policies for Web ActionsOctober 5, 2023 · COLM 2024
- The Tool Illusion: Rethinking Tool Use in Web AgentsApril 3, 2026 · arXiv
- When Users Change Their Mind: Evaluating Interruptible Agents in Long-Horizon Web NavigationApril 1, 2026 · arXiv
- WebArena-Infinity: Generating Browser Environments with Verifiable Tasks at ScaleMarch 2026 · Blog Post
- AI Planning Framework for LLM-Based Web AgentsMarch 13, 2026 · arXiv