Plan-Then-Execute Paradigm Proposed for Web Agents
A recent paper on arXiv suggests that web agents should implement a plan-then-execute approach rather than the commonly utilized ReAct framework. The researchers assert that ReAct's integration of unverified web information into decision-making processes opens a vulnerability for prompt injections that can manipulate the agent's control flow. Conversely, the plan-then-execute method establishes a task-specific program prior to engaging with real-time web content, thereby safeguarding against untrusted information altering user tasks or prompting the model to create new actions. The study evaluates WebArena, a well-known benchmark for web agents, concluding that all tasks align with the proposed approach.
Key facts
- Paper argues web agents should adopt plan-then-execute over ReAct.
- ReAct allows untrusted web content to influence agent actions.
- Plan-then-execute commits to a program before observing runtime content.
- Untrusted data can only affect values or branches within a predefined graph.
- Analysis of WebArena shows all tasks are compatible with plan-then-execute.
- Paper is on arXiv with ID 2605.14290.
Entities
Institutions
- arXiv