Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 26·5 min read
DRIVE: Modeling Skills at the Reasoning and Interaction Levels for Web Agents under Continual Learning
TL;DR
Web agents struggle to balance high-level reasoning, which is transferable across tasks, and low-level interactions, which are specific to web pages. DRIVE is a dual-level skill modeling framework that separates reasoning skills from interaction skills, allowing agents to adapt to new websites more effectively.
✦ Why It Matters
Engineers can leverage DRIVE to improve web agent adaptability and task success across diverse online environments.
Key Takeaways
- Implement DRIVE to enhance web agent performance.
- Utilize distinct reasoning and interaction skills for better adaptability.
- Leverage scene-aware coordination for task-specific skill retrieval.
- Focus on skill-level reflection to refine agent capabilities.
How It Works
DRIVE separates skills into two categories: reasoning skills, which capture abstract task logic, and interaction skills, which are specific to web page elements. This separation allows agents to apply transferable reasoning across different contexts while executing precise actions based on the current page structure.
A coordination mechanism dynamically selects the appropriate skills based on the task's requirements, enhancing overall task execution.