Search papers, labs, and topics across Lattice.
SkillAdam adapts the first- and second-moment mechanics of the Adam optimizer to discrete prompt optimization, pairing an optimization memory to stabilize revision directions with a volatility-driven edit budget to dynamically scale update magnitudes. This formulation resolves key failure modes in agent self-evolution, where noisy, iteration-local execution feedback frequently overwrites effective procedural behaviors. Evaluated across seven short- and long-horizon benchmarks, the approach establishes new state-of-the-art skill evolution performance while substantially decreasing optimization iterations and inference costs compared to heuristic baselines.
Translating Adam's momentum and second-moment adaptivity into discrete prompt memories and dynamic edit budgets fixes the chronic instability and catastrophic overwriting that plague agent skill self-evolution.
Agent skills provide a lightweight way to equip frozen language-model agents with domain knowledge and procedural guidance, yet obtaining high-quality skills remains costly and difficult to scale. Expert-written skills require substantial human effort. Recent skill self-evolution methods automate an iterative loop that uses execution feedback to revise skills, but their heuristic update strategies often yield unstable optimization and low iteration efficiency. We identify two challenges in realizing stable and efficient skill self-evolution. Direction Stability requires effective corrections to accumulate rather than be overwritten by iteration-local feedback. Update Adaptivity requires the scope of each revision to reflect the consistency of recent case-level improvements. We introduce SkillAdam, an Adam-inspired framework for optimizing discrete and non-differentiable skill documents. As a functional analogue of Adam's first moment, an optimization memory records identified problems and the outcomes of prior solution attempts to stabilize the update direction. As a functional analogue of Adam's second moment, a volatility-driven edit budget tracks the history-weighted variation of recent case-level improvements and adaptively controls the update magnitude. Across seven benchmarks that span short- and long-horizon tasks, SkillAdam achieves state-of-the-art performance with more stable optimization dynamics. It also obtains stronger skills with substantially fewer optimization iterations and lower cost than prior methods. Code repository: https://github.com/ruc-datalab/SkillAdam
