Skim: A Speculative Execution Framework for Efficient Web Agents
Skim serves as a speculative execution framework designed for web agents, leveraging the consistent architecture of specialized websites. The primary insight is that websites maintain uniform URL patterns, response formats, and mappings from tasks to trajectories for similar queries. An offline profiler identifies these patterns for each site just once. During execution, Skim aligns each query with a template, generates the corresponding URL, and retrieves the answer using a compact model. A minimal verifier oversees each rapid output; infrequent mispredictions escalate to the complete agent, which is initiated by the final URL from the fast path.
Key facts
- Skim is a speculative execution framework for web agents.
- It exploits predictable structure of purpose-built websites.
- Websites enforce stable URL patterns, answer formats, and task-to-trajectory mappings.
- An offline profiler captures patterns once per site.
- At runtime, Skim matches queries to templates, synthesizes URLs, and extracts answers with a small model.
- A lightweight verifier gates fast-path outputs.
- Misspeculations cascade to the full agent, warm-started by the fast path's final URL.
- The framework aims to reduce expense of web-agent tasks.
Entities
—