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

Evoflux tackles a practical bottleneck for small (1.5B–4B) tool-using agents: with only a few hundred teacher traces available, should that scarce supervision go into fine-tuning the model's weights, or into search at inference time? The paper reframes compact MCP-style tool use as executable workflow repair rather than one-shot function calling—the agent has to resolve tools from live catalogs, satisfy schemas, preserve cross-step dependencies, and ground its final answer in real execution—and introduces Evoflux, an inference-time evolutionary loop that evolves typed workflow graphs through structured edits, execution feedback, adaptive intensity, meta-guided redesign, and diversity pruning, all without updating weights. On held-out MCP-Bench tasks spanning live servers and 250 tools, Evoflux raises execution feasibility from roughly 3% to 17–24% across small planners, while SFT and SFT+DPO trained on the same search-mined data merely match, underperform, or even collapse below the zero-shot baseline—a nice quantification of an underreported risk in small-corpus fine-tuning of compact agents (ReAct hits higher peaks but with far higher variance and token cost). The takeaway is clean: under tight teacher-trace budgets, execution-grounded search transfers more reliably than weight updates.