Steve-Evolving: Open-World Embodied Agents Achieve Autonomous Self-Evolution via Fine-Grained Failure Diagnosis and Dual-Track Knowledge Distillation

arXiv·medium signalXBlueskyLinkedInCopy link

Researchers published Steve-Evolving on arXiv (March 16), demonstrating that the primary bottleneck for long-horizon open-world agent tasks is not single-step planning quality but the agent's ability to distill interaction experience into reusable knowledge. The system uses fine-grained failure diagnosis to identify exactly where a multi-step trajectory broke down, then applies dual-track knowledge distillation — one track for successful strategies, one for failure modes — to update agent behavior without retraining from scratch. Evaluated in Minecraft-style open-world tasks, Steve-Evolving significantly outperforms static agents on novel task sequences by turning each failed episode into a structured learning signal.