Research map / Multi-agent systems and games / Cooperative multi-agent RL

Generalized Intention Modeling in Multi-Agent Reinforcement Learning

NeurIPS 2026 · Multi-agent systems and games › Cooperative multi-agent RL · approach: MARL benchmarks and grouping · 0 citations (OpenAlex)

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 MIX, a mixture-of-intention-experts opponent modeling framework that learns performance-driven weights over four intent embeddings (opponent actions, opponent observations, future states, future ego rewards), with the reward expert trained via InfoNCE.

Problem

Opponent modeling in MARL typically encodes opponent intent using a single fixed episode component (e.g., next opponent action or future state), assuming it is universally representative, although the best representation is likely task- and environment-dependent.

Approach

A soft mixture-of-experts gate over four encoder-decoder intention embeddings trained end-to-end with the PPO policy, plus a new future-reward embedding trained with an InfoNCE objective to maximize mutual information with the ego-agent's future returns.

Key results

• MIX matches or outperforms LIAM, MeLIBA, OMG and NOM across Kuhn Poker, POPP, LBF in seen and unseen opponent settings, and is the only method not catastrophically degrading in high-diversity unseen POPP.
• On GRF with 6 opponents, MIX noticeably outperforms all baselines in episode reward and sample efficiency (20 seeds).
• MIX trains in 4.32 hours on POPP versus 8.49 h for OMG and 25.58 h for MeLIBA, at 5.18 ms inference time versus 1.83 ms for NOM.

Surprising findings

• OMG underperforms even the no-opponent-modeling baseline in no-diversity POPP, attributed to reliance on future global state reconstruction under partial observability.
• Opponent-action embeddings generally perform poorly in most environments but are highly competitive in GRF.

Limitations

• Assumes ego-agent faces randomly sampled opponent policies per episode, not cross-episode evolving opponents.
• Assumes a single ego-agent modeling all opponents jointly; team-based settings not addressed.
• Uses roughly 4x more parameters in the opponent modeling component, increasing inference latency.

Experimental setup

Datasets: Kuhn Poker (OpenSpiel) Partially Observable Predator-Prey Level-Based Foraging Google Research Football

Benchmarks: Kuhn Poker POPP LBF GRF

Metrics: episode return sample efficiency expert relative importance (Gradient x Input)

Baselines: LIAM MeLIBA OMG NOM (PPO without opponent modeling)

Builds on: LIAM OMG MeLIBA PPO MAPPO InfoNCE

Closest papers by meaning

• Opponent Modeling based on Subgoal Inference (NeurIPS 2024 · similarity 0.89)
• An Open-Ended Learning Framework for Opponent Modeling (AAAI 2025 · similarity 0.88)
• Efficient Adaptation in Mixed-Motive Environments via Hierarchical Opponent Modeling and Planning (ICML 2024 · similarity 0.87)
• Mimicking To Dominate: Imitation Learning Strategies for Success in Multiagent Games (NeurIPS 2024 · similarity 0.87)
• Opponent Modeling with In-context Search (NeurIPS 2024 · similarity 0.87)
• Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)