SkillFlow: Flow-Driven Recursive Skill Evolution for Agentic Orchestration
SkillFlow, a newly introduced framework, tackles significant issues in LLM-driven agentic systems for managing tasks, such as strategy failure, excessive gradient variance, and unregulated skill development. It incorporates a trainable Supervisor agent, a structured setting featuring a dynamic skill library, and a fixed executor to facilitate automated multi-turn engagements. The framework utilizes Tempered Trajectory Balance (TTB), a regression-focused flow-matching loss that selects trajectories based on reward, ensuring a variety of orchestration strategies are maintained. This research is available on arXiv with the ID 2605.14089.
Key facts
- SkillFlow is a flow-based framework for LLM-based agentic systems.
- It addresses strategy collapse under reward maximization.
- It addresses high gradient variance with opaque credit assignment.
- It addresses unguided skill evolution.
- SkillFlow uses a trainable Supervisor agent.
- It employs a structured environment with dynamic skill library and frozen executor.
- It uses Tempered Trajectory Balance (TTB) as a regression-based flow-matching loss.
- The paper is available on arXiv with ID 2605.14089.
Entities
Institutions
- arXiv