FitText: Evolving Agent Tool Ecologies via Memetic Retrieval

2026-05-04

AI Summary

This work addresses the semantic gap between user tasks and tool documentation, which static retrieval methods struggle to bridge due to their inability to adapt to the dynamically evolving tool requirements during agent execution. To overcome this limitation, the authors propose FitText, a novel framework that introduces, for the first time, a training-free dynamic retrieval mechanism embedded within the agent’s reasoning loop. FitText iteratively generates and refines natural language pseudo-tool descriptions, leveraging stochastic diversity exploration and evolution-based selection guided by tool memory to co-evolve tool understanding and selection. Experimental results demonstrate that FitText significantly improves retrieval effectiveness, reducing the average rank on ToolRet from 8.81 to 2.78, and achieves a 73% average pass rate on StableToolBench—an absolute improvement of 24 percentage points over static retrieval baselines.

Abstract

A semantic gap separates how users describe tasks from how tools are documented. As API ecosystems scale to tens of thousands of endpoints, static retrieval from the initial query alone cannot bridge this gap: the agent's understanding of what it needs evolves during execution, but its tool set does not. We introduce FitText, a training-free framework that makes retrieval dynamic by embedding it directly in the agent's reasoning loop. FitText generates natural-language pseudo-tool descriptions as retrieval probes, refines them iteratively using retrieval feedback, and explores diverse alternatives through stochastic generation. Memetic Retrieval adds evolutionary selection pressure over candidate descriptions, guided by a tool memory that avoids redundant search. On ToolRet (43k tools, 4 domains), FitText improves average retrieval rank from 8.81 to 2.78; on StableToolBench (16,464 APIs), it achieves a 0.73 average pass rate--a 24-point absolute gain over static query retrieval. The gains transfer across base models capable of acting as competent semantic operators; under weaker base models, Memetic's evolutionary search inverts--amplifying noise rather than refining signal--surfacing model capacity as a prerequisite for evolutionary tool exploration.

Problem

• semantic gap
• tool retrieval
• agent reasoning
• API ecosystems
• dynamic retrieval

Innovation

• dynamic retrieval
• memetic retrieval
• tool ecology
• agent reasoning loop
• pseudo-tool description