Research map / Multi-agent systems and games / Cooperative multi-agent RL
AgentPO: Enhancing Multi-Agent Collaboration via Reinforcement Learning
Lin Sun, Chuang Liu, Can Zhang, Yubin Wu, Weijia Lu, Ning Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train a lightweight hint collaborator with GRPO in front of a frozen actor | Llama-3.1-8B-Instruct actor · Average over 5 math benchmarks | Pass@1: 19.4 (CoT) -> 31.5 | Improved |
| train collaborator with GRPO (with AgentPO) vs untrained hint model (ablation) | Qwen-Plus actor, Hint-Actor · Average over 5 math benchmarks | Pass@1: 55.1 -> 58.8 | Improved |
| train critic with AgentPO vs untrained critic (ablation) | Qwen-Plus actor, Critic-Actor · Average over 5 math benchmarks | Pass@1: 57.9 -> 58.9 | Improved |
| fine-tune only the 3B hint collaborator instead of the 7B actor | Qwen2.5-Math-7B · Average over 5 math benchmarks | Pass@1: 48.0 (Prime-Zero-7B) -> 49.4 | Improved |
| add AgentPO collaborator to multi-agent debate | Llama-3.2-3B-Instruct debate · OlympiadBench | Pass@1: 15.3 -> 21.1 | Improved |
| add AgentPO collaborator to role assignment | Llama-3.2-3B-Instruct role assignment · MATH500 | Pass@1: 48.3 -> 52.2 | Improved |
| increase training samples from 500 to 700 and 1,000 (ablation) | AgentPO hint model · Average over 5 math benchmarks (Qwen2.5-Math-7B actor) | Pass@1: 49.4 at 500 samples, slight dip at 700 and 1000 | Hurt |
| use hint model Qwen2.5-3B-Instruct instead of larger Qwen2.5-7B-Instruct (ablation) | AgentPO with Llama-3.1-8B-Instruct actor · Average Pass@1 | Pass@1: 29.1 -> 31.5 | Improved |
| add AgentPO hint model to API-based Qwen-Plus actor | Qwen-Plus · OlympiadBench | Pass@1: 49.6 -> 49.6 | No clear effect |
Closest papers by meaning
- Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.89)
- MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.89)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.88)
- InfoPO: Information-Driven Policy Optimization for User-Centric Agents (ICML 2026 · similarity 0.88)
- Learning to Deliberate: Meta-policy Collaboration for Agentic LLMs with Multi-agent Reinforcement Learning (AAAI 2026 · similarity 0.87)
- Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
