DLM: Unified Decision Language Models for Offline Multi-Agent Sequential Decision Making

Researchers from an undisclosed institution have introduced the Decision Language Model (DLM), a novel approach that unifies offline multi-agent sequential decision making using large language models (LLMs). The key insight is to reframe multi-agent coordination as a dialogue-style sequence prediction problem under the centralized training with decentralized execution (CTDE) paradigm. DLM is trained in two stages: first, supervised fine-tuning on dialogue-style datasets that incorporate inter-agent context and execute actions from offline trajectories; second, group relative policy optimization (GRPO) to improve robustness against out-of-distribution actions via lightweight reward functions.

On multiple benchmarks, a single unified DLM outperforms strong offline MARL baselines and existing LLM-based conversational decision-making methods. Notably, the model demonstrates strong zero-shot generalization to unseen scenarios across tasks, suggesting it can adapt to new environments without additional training. The work, detailed in a 22-page paper with 11 figures, addresses a key limitation of traditional MARL approaches that rely on fixed observation formats and action spaces. By leveraging the flexible modeling interface of LLMs, DLM can naturally handle heterogeneous observations and actions, making it more scalable and reusable for real-world multi-agent systems like autonomous driving fleets, warehouse robotics, or coordinated drone swarms.