Evoflux boosts small AI agents' tool use from 3% to 24%

Compact language models (LMs) promise lower cost and latency for tool agents, but they struggle when navigating live tool catalogs, satisfying schemas, and preserving dependencies. Standard fine-tuning from teacher traces often fails because it cannot cover the recovery behaviors needed to repair broken plans. To address this, researchers from IBM Research and Rensselaer Polytechnic Institute developed Evoflux, a novel inference-time method that reimagines tool use as the evolution of executable workflow graphs.

Evoflux starts with a candidate workflow, then applies structured edits—adding, removing, or rerouting tool calls—guided by execution feedback and adaptive intensity. It uses meta-guided redesign and diversity pruning to avoid local optima. Tested on the MCP-Bench suite with 250 real tools, Evoflux boosted success rates from roughly 3% to 17–24% across three small planners (e.g., Llama 3.2 3B). In contrast, supervised fine-tuning (SFT) and SFT with direct preference optimization (DPO) on the same search-mined data either matched or underperformed zero-shot baselines.

The key insight is that execution-grounded search is more sample-efficient than distillation when teacher traces are scarce. While ReAct-style agents can reach higher peaks, they do so with much higher variance and token cost. Evoflux offers a practical, resource-efficient path to enabling small models to use complex tool ecosystems reliably.