Skip to main content
Back

LMAC Lets an LLM Design Multi-Agent Communication so Robots Stop Flying Blind

LMAC uses an LLM to design communication protocols for multi-agent RL, letting agents reconstruct shared world state and dramatically outperform prior baselines

Paper, Full Access
  • LMAC uses an LLM to design structured communication protocols for cooperative multi-agent RL, replacing end-to-end learned messaging.
  • The core objective is state reconstruction: every agent should be able to recover the global state as accurately and uniformly as possible.
  • A two-step refinement loop improves per-agent accuracy (Step 1) and reduces knowledge gaps between agents (Step 2) using only RL rollout data.
  • Evaluated on SMAC, Level-Based Foraging, and Google Research Football, outperforming TarMAC, MASIA, DSMC, SMS, and FinMAC baselines.
  • The resulting protocol is human-readable code -- interpretable and sparse by design -- integrated into standard CTDE training via a meta-cognitive latent module.
  • Code is publicly available at the LMAC project page.

In cooperative multi-agent reinforcement learning, agents face a fundamental handicap: they can only see a slice of the world. A drone in a swarm, a unit in a StarCraft battle, a robot on a warehouse floor -- each agent operates with partial observability, meaning it only perceives what is within its local sensor range. Communication between agents is the standard fix, but designing what to say, to whom, and when has remained a stubborn open problem. A new paper from researchers introduces LMAC (LLM-driven Multi-Agent Communication), a framework that hands the protocol-design job to a large language model.

The gap that keeps tripping up MARL

The challenges in MARL are fundamentally rooted in partial observability, where agents cannot access the complete system state and thus struggle to determine which information is task-relevant or beneficial to other agents. The traditional fix is to let agents learn what to communicate end-to-end alongside their policies -- methods like TarMAC, CommNet, and MASIA all take this route. In MARL, communication enables coordination but can lead to inefficient information exchange, since agents may generate redundant or non-essential messages.

The deeper issue is asymmetry. Even when agents do communicate, some end up with much better situational awareness than others. An overseer drone knows where the enemy is; the ten banelings it commands do not. If the protocol does not explicitly close that gap, the team's weakest-informed agent becomes the bottleneck. Achieving efficient state reconstruction requires exchanging only the essential messages needed for recovery, yet identifying them is challenging because it demands understanding the task objective and the relationship between the state and observations.

Let the LLM write the protocol

Given natural-language descriptions of the task goal and the state and observation dimensions, the LLM designs an initial communication protocol to support accurate state reconstruction. This is the core insight of LMAC: instead of learning communication from scratch via gradient descent, use an LLM's pre-trained reasoning to bootstrap a structured, interpretable protocol before any RL training begins.