We propose a generalizable multi-agent reinforcement learning framework exploiting causal structures, enabling efficient adaptation across networked systems under domain shifts, validated by theoretical guarantees and numerical experiments

This paper is to jointly optimize mutiple modules in complex RAG pipeline using multi-agent reinforcement learning.

We propose FCGrad, a conflict-aware gradient adjustment method that improves social welfare while ensuring fairness in multi-agent mixed-motive settings by resolving conflicts between individual and collective objectives.

Recurrent neural networks spontaneously model partners during collaboration — without specialised architectures — but only when partner-specific adaptation improves task performance.

We develop and analyze a scalable algorithm for multi-agent RL by sampling from the mean-field distribution of the agents to overcome the curse of dimensionality.

We propose an agent-conditioned hypernetwork (HyperMARL) that enables adaptive cooperation in MARL by decoupling observation- and agent-conditioned gradients, resolving a key interference problem caused by coupling agent IDs with observations.

Large fixed context length limits exploration and introduces redundancy in MARL. We propose an adaptive context length optimization method with Fourier-based low-frequency truncation to improve long-term decision-making.

We propose a novel framework employs a MoE-based multi-agent world model for learning and planning in multi-task scenarios.

A novel value decomposition framework of Continued Fraction Q-Learning (QCoFr) is proposed to model rich cooperation for multi-agent reinforcement learning without combinatorial explosion.

We propose BayesG, a decentralized actor–critic framework that jointly learns context-aware interaction graphs and policies via variational Bayesian inference over ego-graphs in large-scale networked MARL.

We introduce OrbitZoo, a multi-agent RL environment designed for orbital dynamics, enabling scalable and flexible simulations for autonomous satellite operations with robust and accurate dynamics.

This paper investigates a principle and develops a method for intervening on a single, targeted agent in a multi-agent reinforcement learning to reach a collective objective consisting of the primary task goal and an additional desired outcome.

This paper establishes the mathematical foundation of value decomposition in MARL.

We extend autoregressive multi-agent sequence models, including Sable and MAT, to the Offline MARL setting and demonstrate that they significanlty outperform current state-of-the-art methods across a diverse set of benchmarks with up to 50 agents.

MACS: a new multi-agent reinforcement learning method for faster geometry optimization of crystal structures.