FitText: Evolving Agent Tool Ecologies via Memetic Retrieval

TL;DR

FitText is a dynamic, training-free retrieval framework that enhances agent understanding of tools by generating and refining natural-language descriptions through memetic evolution, significantly improving tool retrieval accuracy.

Contribution

It introduces a novel memetic retrieval approach that iteratively refines tool descriptions within an agent's reasoning process, bridging the semantic gap in large API ecosystems.

Findings

Improves average retrieval rank from 8.81 to 2.78 on ToolRet.

Achieves a 0.73 pass rate on StableToolBench, a 24-point improvement.

Memetic search amplifies noise with weaker models, highlighting the need for capable base models.

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.