Research map / Multi-agent systems and games / LLM agents
LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination
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 SEMDIV, which uses GPT-4o to iteratively generate natural-language coordination behaviors, turn them into reward functions that train verified diverse teammates, and train multi-head agents that LLM-select a policy head for unseen teammates.
Problem
Teammate diversity in cooperative MARL is typically policy-level and lacks semantic grounding, making teammate generation inefficient and agents poorly adaptable to unseen teammates with specific coordination behaviors.
Approach
Generate diverse coordination behaviors in natural language with an LLM, translate each into a reward function to train a teammate policy verified for validity, novelty, and behavior alignment, then train multi-head agents with continual learning and select heads via LLM reasoning over behavior descriptions.
Key results
- SEMDIV averages R1 0.74 and R2 0.68 across four environments with unseen teammates.
- SEMDIV outperforms the best baseline Macop by 19% on task success rate and 39% on behavior satisfaction R2 (GRF case in paper summary).
- SEMDIV-PBT and multi-head SEMDIV outperform FCP, MEP, and LIPO; FCP with 48 teammates is matched by SEMDIV-PBT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select policy heads using behavior-description embedding distance instead of the LLM selector (ablation) | SEMDIV · Average | R1 (task success rate): 0.65 ±0.02 -> 0.42 | Hurt |
| Select policy heads using behavior-description embedding distance instead of the LLM selector (ablation) | SEMDIV · Average | R2 (behavior satisfaction rate): 0.64 ±0.02 -> 0.33 | Hurt |
| Use multi-head policies instead of a single policy network (ablation) | SEMDIV · Four MARL environments | Coordination performance | Helped |
| Use SEMDIV-generated teammates with population-based training instead of SEMDIV's multi-head training (ablation) | SEMDIV · Four MARL environments | Coordination performance | Hurt |
| Use an LLM to select the policy head rather than embedding-distance selection (ablation) | SEMDIV · Four MARL environments | Coordination performance | Helped |
Surprising findings
- LLM-Agent, using only the LLM without MARL training, performs comparably in simple LBF and PP but degrades severely in SMACv2 and GRF (GRF R1 0.14).
Limitations
- SEMDIV still falls short of the Oracle upper bound in all environments.
- Relies on proprietary GPT-4o and requires per-iteration LLM reward generation and verification.
- Evaluated with five hand-designed unseen teammates per environment.
Experimental setup
Datasets: Level-Based Foraging Predator-Prey SMACv2 Google Research Football
Benchmarks: Level-Based Foraging Predator-Prey SMACv2 Google Research Football
Metrics: R1 task success rate R2 satisfaction of teammate preferred coordination behavior
Baselines: FCP MEP LIPO Macop SEMDIV-Dist LLM-Agent SEMDIV-PBT Macop-PBT BRDiv
Builds on: FCP MEP LIPO Macop MAPPO VDN
Closest papers by meaning
- NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming (NeurIPS 2026 · similarity 0.89)
- Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.88)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
- TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (ICML 2026 · similarity 0.88)
- ProAgent: Building Proactive Cooperative Agents with Large Language Models (AAAI 2024 · similarity 0.88)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.88)
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.