The paper presents SkillAdam, a framework for automatically improving skill documents used by frozen LLM agents. It addresses two problems in prior self-evolution methods: updates often overwrite each other instead of accumulating improvements, and the size of each revision does not adapt to how consistently recent changes help. SkillAdam mimics Adam optimization by keeping a memory of past problems and solutions to stabilize the direction of updates, and by using a volatility-driven edit budget to control how much to change each time. On seven benchmarks covering short and long horizon tasks, SkillAdam achieves better performance with fewer iterations and lower cost than existing approaches.
Adam-inspired updates make agent skill evolution steadier and more efficient
The authors show that retaining optimization history and adapting edit size improves feedback-driven skill documents across seven agent benchmarks.
Chinese Tech
Gaoyuan Li · Meihao Fan · Yizhe Liu · Shaolei Zhang · Ju Fan · Siyi Wang · +4 more
Renmin University of China · Tencent
Research Digest··2 min read
Thread:Skill Selection for Agents
In one line
SkillAdam uses an Adam-inspired memory and adaptive budget to stabilize and accelerate skill evolution for LLM agents.
Before you read this
Two ideas this paper builds onRead it as
Li et al. introduce SkillAdam, a method for iteratively improving the natural-language instructions and procedural knowledge used by frozen language-model agents.
The authors treat —documents containing domain knowledge and procedural guidance—as discrete, non-differentiable objects to be optimized from execution feedback.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.