Harness-RL: Black-Box Reinforcement Learning with Action-Args Decoupling for Central-Agent Multi-Agent Harnesses

TL;DR

Harness-RL is a reinforcement learning framework that addresses challenges in training central policies in multi-agent harnesses. It combines Conflict-Aware Policy Optimization with black-box trajectory construction to improve performance. This approach enables more effective optimization of action and argument tokens.

Why It Matters

Harness-RL has practical significance in large language model agents and multi-agent systems, where it can improve task performance and efficiency. Its ability to handle complex, dynamic environments makes it a valuable tool for real-world applications. This can lead to breakthroughs in areas like question answering and agentic retrieval.

Key Findings

• average F1 scores of 42.93 and 47.79 on benchmarks
• CAPO contributes to improved performance
• central-only optimization is favored in the evaluated setting