FitText: Dynamic Tool Retrieval for AI Agents via Memetic Evolution
FitText, a novel framework, tackles the disconnect between user task descriptions and tool documentation within expansive API ecosystems. Created by researchers and published on arXiv, it integrates dynamic retrieval into an agent's reasoning process, eliminating the necessity for training. FitText produces natural-language pseudo-tool descriptions as probes, refining them through iterative retrieval feedback and exploring various alternatives via stochastic generation. Its primary innovation, Memetic Retrieval, employs evolutionary selection pressure on candidate descriptions, utilizing a tool memory to prevent redundant searches. In testing on the ToolRet benchmark (43,000 tools across 4 domains), FitText enhanced the average retrieval rank from 8.81 to 2.78. Additionally, on StableToolBench (16,464 APIs), it recorded a 0.73 average pass rate, marking a 24-point improvement over baseline methods. The framework is designed to be training-free and scalable with expanding API ecosystems.
Key facts
- FitText is a training-free framework for dynamic tool retrieval in AI agents.
- It generates and refines natural-language pseudo-tool descriptions as retrieval probes.
- Memetic Retrieval adds evolutionary selection pressure over candidate descriptions.
- Tool memory avoids redundant searches during retrieval.
- On ToolRet (43k tools, 4 domains), average retrieval rank improved from 8.81 to 2.78.
- On StableToolBench (16,464 APIs), average pass rate reached 0.73, a 24-point improvement.
- The framework addresses the semantic gap between user tasks and tool documentation.
- Published on arXiv with ID 2605.02411.
Entities
Institutions
- arXiv