COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization

简体中文

L1 core-mechanism diagnostic. Runs contextual UCB allocation and feedback admission without external LLMs or six real environments.

Paper information

Field	Details
Paper	COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
Organization	First-author affiliation: The Chinese University of Hong Kong, Shenzhen
First public release	2026-09-10
Authors' code	https://github.com/Jerry-LuP/COBRA-Skills
Adapter / method	cobra-skills
Local reproduction code	src/auto_research/agent_research/latest_20260914.py

Original paper

Motivation and method

Treat skill optimization as a budgeted contextual bandit over a dynamic candidate space. Evaluate high-reward or informative skills first and evolve them using execution feedback.

flowchart LR
  I[Input and public state] --> M[COBRA-Skills mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or update]

Key figure from the paper

Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core computation or operation

The local deterministic reference kernel implements the decisive operation and reports gate, selection, projection, graph-edge, or state-update statistics. Tests cover formula invariants and boundaries.

Paper results

Across six benchmarks and three models, cost falls 55%–58%, and performance improves by 13.1/26.9/22.5 percentage points against agents without skills. These paper results are not directly comparable with the scaled-down local run. No online gain is inferred without an online A/B test.

Local reproduction

See the three-seed metrics. All candidates use identical public fixtures, budgets, and seeds. These results establish only execution of the core state transitions.

Reproduction boundaries

The local run exercises contextual UCB allocation and feedback admission, not external LLMs or six real environments.