AI Agents Self-Evolve Skills with High Reliability

This research presents reSolve, a novel framework designed to enable AI agents to self-evolve their skills more effectively and reliably. Current methods for skill evolution often result in agents underperforming human-curated skills or failing to reproduce learned trajectories consistently at deployment. reSolve addresses these issues through three core components: a "solve-and-reproduce" protocol that decouples interactive problem-solving from the creation of a self-contained, independently re-executable skill package; the enhancement of sparse reward signals with a surrogate verifier that operates without access to hidden tests or reference answers; and the application of verifier-guided beam search over a solution-construction graph. Within a fixed testing harness, a computationally inexpensive model using reSolve was able to self-evolve skills that achieved a 74.9% mean-of-3 success rate. This represents a substantial 14.8 percentage point improvement over the 60.1% human-curated baseline and even surpasses the strongest official curated-skill result of 67.3% achieved by more powerful models. The paper also details observed failure cases and domain-specific results, including performance on Natural Science tasks, to clarify the conditions under which this approach is most beneficial.