Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Multi-Expert Distillation for Few-Shot Coordination (Student Abstract)
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 Multi-Expert Distillation (MED), which trains a transformer generalist via supervised learning on advised actions from a population of expert teammates to achieve few-shot ad hoc coordination.
Problem
Ad hoc teamwork / few-shot coordination: population-based training learns a generalist agent from scratch with RL against diverse unseen teammates, which is inefficient and unstable, and does not adapt across episodes.
Approach
Treat the population's per-joint-policy partner as an expert and distill its advised actions into a transformer generalist using cross-entropy on across-episodic trajectories, so the policy conditions on interaction history and can adapt over episodes instead of learning by RL from scratch.
Key results
- On MoveBox and Overcooked, MED reaches comparable performance with far fewer sampling timesteps than DDQN, RL2 and PEARL trained from scratch.
- MED's training curves show significantly greater stability across random seeds than the RL baselines.
- In Overcooked, MED's second-episode performance surges above its first episode, which is comparable to all baselines, indicating use of interaction history.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| distill population experts’ advised actions with supervised learning instead of learning from scratch with RL | MED · Gridworld MoveBox | sampling timesteps to learn | Helped |
| distill population experts’ advised actions with supervised learning instead of learning from scratch with RL | MED · Overcooked | sampling timesteps to learn | Helped |
| use MED rather than the baseline methods | MED · Gridworld MoveBox and Overcooked | training stability | Helped |
| condition actions on historical interaction contexts across episodes | MED · One-Step Cooperative Matrix Game | expected return | Helped |
| condition actions on historical interaction contexts across episodes | MED · Overcooked | return | Helped |
| use MED in the first episode rather than the baselines | MED · Overcooked | return | No clear effect |
Limitations
- Evaluated only on small environments (a one-step matrix game, MoveBox, Overcooked).
- Only two-player settings considered.
- Relies on a pre-generated population of expert teammates.
Experimental setup
Datasets: One-Step Cooperative Matrix Game Gridworld MoveBox Overcooked
Benchmarks: One-Step Cooperative Matrix Game Gridworld MoveBox Overcooked
Metrics: return mean training timesteps efficiency training stability across seeds
Baselines: DDQN RL2 PEARL
Builds on: Population-Based Training LIPO DDQN RL2 PEARL
Closest papers by meaning
- LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination (ICML 2025 · similarity 0.86)
- Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation (ICLR 2025 · similarity 0.85)
- Multi-Agent Coordination via Support-Preserving Distillation (NeurIPS 2026 · similarity 0.84)
- Progress-Aware Distillation for Mitigating Stagnation in Small Language Model Agents (NeurIPS 2026 · similarity 0.84)
- N-agent Ad Hoc Teamwork (NeurIPS 2024 · similarity 0.84)
- Efficient Active Imitation Learning with Random Network Distillation (ICLR 2025 · similarity 0.84)
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.