FeasiGen: Automatically Creating Infeasible Tasks for Tool-Using Agents
A recent study published on arXiv (2605.28532) presents FeasiGen, an automated system aimed at creating infeasible tasks for tool-utilizing agents. This initiative seeks to enhance efficiency by allowing agents to recognize when a task cannot be accomplished with the tools at their disposal, thereby reducing unnecessary computational expenses. FeasiGen gathers tool-calling traces from successful runs across various agent systems, pinpoints essential tools frequently employed in different strategies, and conceals these tools to convert solvable tasks into infeasible ones. Human validation has demonstrated an accuracy rate exceeding 94% for the infeasibility annotations. This method tackles the real-world issue of agents squandering resources on unattainable tasks in limited tool settings.
Key facts
- FeasiGen is an automatic pipeline for constructing infeasible agent tasks.
- It identifies critical tools required for task completion.
- It extracts tool-calling traces from successful executions across multiple agent systems.
- It masks critical tools to transform solvable tasks into infeasible ones.
- Human verification confirms over 94% accuracy for infeasibility annotations.
- The approach aims to reduce computational cost by early detection of infeasible tasks.
- The research is published on arXiv with ID 2605.28532.
- The paper focuses on tool-using agents in constrained tool environments.
Entities
Institutions
- arXiv