What they did
The authors treat agent skills—documents containing domain knowledge and procedural guidance—as discrete, non-differentiable objects to be optimized from execution feedback. SkillAdam applies functional analogues of the Adam optimizer’s two running statistics rather than attempting gradient-based updates.
An optimization memory records previously identified problems and the outcomes of attempted fixes, helping useful corrections persist across iterations. A volatility-driven edit budget tracks variation in recent case-level improvements and adjusts how extensively the skill document is revised. The method was evaluated on seven benchmarks spanning short- and long-horizon tasks.
Key findings
- SkillAdam achieved state-of-the-art performance across the seven evaluated benchmarks, according to the authors.
- Its optimization trajectories were more stable than those of prior heuristic skill self-evolution methods, with successful corrections less likely to be overwritten by feedback from a single iteration.
- The adaptive edit budget tied revision scope to the consistency of recent improvements rather than using a fixed update size.
- SkillAdam produced stronger skills with substantially fewer optimization iterations and lower cost than prior methods, although the abstract does not provide exact reductions.
Why it matters
Skill documents offer a way to improve specialized agent behavior without changing the underlying language model. A more stable and iteration-efficient method for revising those documents could reduce the human and computational cost of adapting frozen agents to new domains, especially where tasks require repeated execution and feedback.
Caveats
The supplied abstract does not identify the seven benchmarks, underlying models, baselines, absolute performance gains, or cost reductions, so the breadth and practical size of the reported advantage cannot be assessed here. SkillAdam is also an analogy to Adam for document editing, not a gradient optimizer, and its reliability outside the tested tasks remains an open question.