SkillFlow: Flow-Driven Recursive Skill Evolution for Agentic Orchestration

2026-05-13

AI Summary

This work addresses key challenges in task orchestration with large language model agents, including policy collapse, opaque credit assignment, and the absence of training signals for skill evolution. To overcome these limitations, the authors propose SkillFlow, a framework that introduces a trainable supervisor and structured environment interactions. It employs a tempered trajectory balance loss based on flow matching to enable multi-turn task orchestration while preserving policy diversity and enabling transparent credit assignment at zero additional inference cost. Furthermore, SkillFlow incorporates a recursive skill evolution mechanism coupled with a dynamic skill repository, establishing a closed-loop pathway from training signals to capability enhancement. Experimental results demonstrate that the proposed method significantly outperforms baseline approaches across 14 diverse datasets spanning question answering, mathematical reasoning, code generation, and real-world decision-making tasks.

Abstract

In recent years, a variety of powerful LLM-based agentic systems have been applied to automate complex tasks through task orchestration. However, existing orchestration methods still face key challenges, including strategy collapse under reward maximization, high gradient variance with opaque credit assignment, and unguided skill evolution whose decisions are typically made by directly prompting an LLM to judge rather than derived from principled training signals. To address these challenges, we propose SkillFlow, a flow-based framework that takes a trainable Supervisor as the agent and a structured environment with dynamic skill library and frozen executor, automating task orchestration through multi-turn interaction. SkillFlow employs Tempered Trajectory Balance (TTB), a regression-based flow-matching loss that samples trajectories proportional to reward, preserving diverse orchestration strategies rather than collapsing to a single mode. The same flow objective yields a jointly learned backward policy that provides transparent per-step credit assignment at zero additional inference cost. Building on these flow diagnostics, a recursive skill evolution mechanism determines when to evolve, what skills to create or prune, and where decision gaps lie -- closing the loop from training signal to autonomous capability growth. Experimental results on 14 datasets show that SkillFlow significantly outperforms baselines across question answering, mathematical reasoning, code generation, and real-world interactive decision making tasks. Our code is available at https://anonymous.4open.science/r/SkillFlow-E850.

Problem

• task orchestration
• strategy collapse
• credit assignment
• skill evolution
• LLM-based agentic systems

Innovation

• flow-based learning
• Tempered Trajectory Balance
• recursive skill evolution
• credit assignment
• agentic orchestration