Researchers propose Decision Language Model (DLM) for multi-agent offline decision-making, formulating the task as dialogue-style sequence prediction

DLM trains in two stages: supervised fine-tuning on dialogue-style datasets for centralized training with inter-agent context, followed by group relative policy optimization to enhance robustness to out-of-distribution actions through lightweight reward functions.

Unlike existing offline multi-agent reinforcement learning methods that rely on fixed observation formats and action spaces, DLM leverages the flexible modeling interface of large language models (LLMs) to accommodate heterogeneous observations and actions across agents.

Experiments show DLM outperforms strong offline MARL baselines and LLM-based conversational decision-making methods, while demonstrating strong zero-shot generalization to unseen scenarios across tasks.