EvoFlux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents
EvoFlux is an inference-time search method that repairs flawed tool plans for small language models by treating them as executable workflows. It uses an evolutionary loop to iteratively edit plans—swapping tools, correcting arguments, or adding steps—based on execution feedback from live servers. This execution-guided repair significantly boosts the success rate of agents using limited supervision compared to standard fine-tuning methods.