Evoflux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents

Evoflux is introduced as an inference-time evolutionary search method designed to improve the execution of compact language models (LMs) in tool use scenarios. It enhances the feasibility of executing workflows by evolving typed workflow graphs through structured edits and feedback, achieving execution feasibility improvements from approximately 3% to between 17% and 24% on MCP-Bench tasks with 250 tools. This method demonstrates that execution-grounded search can outperform traditional fine-tuning methods, such as SFT and ReAct, particularly in environments with limited teacher traces, making it significant for practitioners working with compact agents.