COBRA-Skills formulates skill optimization as budgeted sequential optimization, using a contextual bandit algorithm to prioritize which candidate skills to evaluate and evidence-grounded evolution to refine skills from execution feedback. Across six benchmarks and three target models, it achieved the strongest average performance while reducing optimization cost by 55-58% relative to SkillOpt, using only 50 unique optimization examples per benchmark. This makes skill optimization more efficient, reducing the need for costly evaluations and large datasets, and the method is robust to changes in the agent harness and works when the target model itself generates and refines skills.
Bandit-guided skill evolution cuts LLM agent optimization costs
COBRA-Skills uses a contextual bandit to prioritize candidate evaluations, then refines reusable agent skills using execution feedback.
Top University
Pingchen Lu · Xiangyi Wang · Xiang Li · Jie Mao · Zikun Qu · Junfeng Luo · +3 more
The Chinese University of Hong Kong, Shenzhen · Tianjin University · The Hong Kong University of Science and Technology (Guangzhou) · National University of Singapore
Research Digest··2 min read
Thread:Skill Selection for Agents
In one line
COBRA-Skills uses contextual bandits to optimize LLM agent skills, achieving top performance with 55-58% less cost than SkillOpt.
Before you read this
Two ideas this paper builds onRead it as
Lu et al. frame agent-skill optimization as a budget-constrained search over a changing pool of candidates.
The authors developed COBRA-Skills, a framework for optimizing reusable instructions or procedures distilled from an agent’s previous task experience.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code