FlowEvo lets AI agents self-evolve by saving reusable skills
New framework achieves 82.8% on ALFWorld with half the tokens of baselines.
FlowEvo, developed by Zeyu Ren and colleagues, tackles a key limitation in LLM agents: useful procedures discovered during execution are often transient and lost after the task. The framework creates a closed loop where successful reasoning, tool use, and code execution traces are compiled into callable skill records. These records include structured guidance and pass interface, replay, and safety checks before being stored in an inference-time skill bank. When facing new tasks, agents can retrieve and execute relevant skills directly or inject them as structured context, allowing them to build on past successes without retraining.
On three benchmarks, FlowEvo delivered strong results. In the ALFWorld interactive environment, it achieved 82.8% success rate—23.6 percentage points above the strongest baseline—while using less than half the average tokens per episode. On code generation (HumanEval) and math reasoning (GSM8K), it matched or exceeded baseline performance at lower token cost. Controlled ablations confirmed that all three mechanisms—compilation, feedback, and curation—contribute to the gains. FlowEvo is open-source, and its self-evolving approach could make long-running agents more efficient and capable over time.
- FlowEvo compiles successful agent traces into reusable skill records stored in a skill bank, without any model fine-tuning.
- On ALFWorld, it achieved 82.8% success rate (+23.6 points over baseline) while using less than half the tokens per episode.
- Three mechanisms drive the loop: workflow-to-skill compilation, skill-to-workflow feedback, and skill curation to prevent negative transfer.
Why It Matters
FlowEvo enables LLM agents to continuously improve from experience, slashing costs and boosting performance without retraining.