Skill^{infty}: Recursive Self-Improvement for Agent Skill Optimization

Skill^{infty}. Recursive Self-Improvement for Agent Skill Optimization. By Wangcheng Tao and three others, from the National University of Singapore and City University of Hong Kong. Published on the twenty-sixth of October. No code repository is linked to it.

Recursive self-improvement, where a system also improves the way it improves, is a long-standing aspiration on the path to AGI. Language agents now store more and more of their procedural knowledge in reusable skills, and that opens a direct route. If the optimization procedure can itself be written as a skill, then optimizing skills can include optimizing the optimizer. The obstacle is that existing skill optimizers follow a hand-designed procedure. Even methods that evolve a meta-skill keep it confined to a fixed pipeline, and relearn it for each task.

Skill^{infty} makes skill optimization recursively self-improving. A meta agent follows a meta-skill to optimize task skills, and then uses that same procedure to optimize the meta-skill itself. One procedure, applied at two levels, with the upper level able to rewrite its own rules.

The optimization is led by the agent. The meta agent diagnoses execution traces, proposes candidates, requests independent reviews, and allocates the search budget. Each of those steps is defined by the meta-skill, and that is what allows self-improvement to rewrite them.

The meta-skill is also decoupled from the task skills. It is refined on multiple skill optimization tasks, and then reused to optimize task skills across diverse domains, rather than relearned each time. As in deployment, skills at both levels are standard packages, loaded through native skill discovery.

The evaluation uses eighty-three SkillsBench tasks. Skill^{infty} reaches a pass rate of sixty point seven percent. Human-curated skills reach fifty-six point four percent on the same tasks. SkillEvolver reaches fifty-four point five percent.

What distinguishes the approach is where the procedure lives. Earlier optimizers keep their method outside the thing being optimized, as a pipeline someone designed. Here the method is a skill like any other, packaged and loaded the same way, which puts it within reach of the very process it describes. And because it is refined once across several tasks and then carried to new domains, the improvement to the improver is kept rather than thrown away with each new task.