Research map / Multi-agent systems and games / Human-AI and goal-conditioned
ProAgent: Building Proactive Cooperative Agents with Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ProAgent, a training-free LLM-based proactive cooperative agent that infers teammate intentions, corrects beliefs, and verifies skills to coordinate zero-shot in Overcooked-AI.
Problem
Zero-shot coordination in cooperative multi-agent games: learning-based cooperative agents generalize poorly to unfamiliar teammates, both AI and human proxies, in Overcooked-AI.
Approach
Use an LLM planner in a decentralized, training-free pipeline with memory, teammate-intention inference, belief correction, a verificator for skill failures, and a controller that maps skills to low-level actions, so the agent adapts to teammates without training with them.
Key results
- On Overcooked-AI, ProAgent as Player 0 achieves the highest average reward in all five layouts against SP, PBT, FCP, MEP, and COLE partners (e.g., 197.3 vs. best baseline 178.8 in Cramped Room first row).
- With human proxy BC partners, ProAgent outperforms the baseline in four of five layouts, with the paper reporting an average improvement above 10% over the state of the art.
- Ablation in Cramped Room: score 204 with analysis and belief, 184 without belief, and 100 with neither; removing the Verificator drops success rate to 20%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProAgent instead of the strongest baseline for Player 0 | AI-agent cooperation · Cramped Room; Asymmetric Advantages; Coordination Ring; Forced Coordination; Counter Circuit | average reward | Helped |
| Use ProAgent instead of the strongest baseline for Player 1 | AI-agent cooperation · Asymmetric Advantages; Coordination Ring; Counter Circuit | average reward | Helped |
| Use ProAgent instead of the strongest baseline for Player 1 | AI-agent cooperation · Cramped Room | average reward | Hurt |
| Use ProAgent instead of the strongest baseline for Player 1 | AI-agent cooperation · Forced Coordination | average reward | Hurt |
| Add teammate-intention belief to analysis-based planning (ablation) | ProAgent Planner · Cramped Room | average reward: 184 -> 204 | Helped |
| Add analysis and teammate-intention belief instead of directly outputting a skill plan (ablation) | ProAgent Planner · Cramped Room | average reward: 100 -> 204 | Helped |
| Add the Verificator module and feedback-based replanning (ablation) | ProAgent · Overcooked-AI, 100 steps | success rate: 20% | Helped |
| Use ProAgent with human-proxy partners instead of the baseline | Human-proxy cooperation · Overcooked-AI layouts | cumulative reward | Helped |
Surprising findings
- ProAgent as Player 1 loses to PBT in Forced Coordination and slightly underperforms FCP in Cramped Room.
- Baseline pairs show large performance differences depending on starting position in asymmetric layouts, while ProAgent is largely unaffected.
Limitations
- Controller uses the built-in Best-First Search rather than a learned controller, so low-level performance depends on the environment's planner.
- Verificator and planner rely on prompting; the paper notes a failure case as Player 1 that is only examined in the appendix.
- Evaluation limited to Overcooked-AI with five episodes per pair and five BC proxies.
Experimental setup
Datasets: Overcooked-AI
Benchmarks: Overcooked-AI
Metrics: average episode reward success rate
Baselines: SP PBT FCP MEP COLE
Builds on: chain-of-thought Overcooked-AI Population-Based Training Fictitious Co-Play
Closest papers by meaning
- Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance (ICLR 2025 · similarity 0.90)
- LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination (ICML 2025 · similarity 0.88)
- AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024 · similarity 0.88)
- Learning to Cooperate with Humans using Generative Agents (NeurIPS 2024 · similarity 0.87)
- Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · 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.