Search papers, labs, and topics across Lattice.
DiffMAS is introduced as a training framework that jointly optimizes latent communication and multi-agent reasoning by treating inter-agent communication as a learnable component. It uses parameter-efficient supervised training over multi-agent latent trajectories, allowing agents to learn optimal information encoding and interpretation. Experiments across mathematical reasoning, scientific QA, code generation, and commonsense benchmarks demonstrate that DiffMAS improves reasoning accuracy and decoding stability compared to single-agent inference, text-based multi-agent systems, and existing latent communication methods.
Ditch the fixed interface: DiffMAS unlocks surprisingly large gains in multi-agent reasoning by jointly optimizing latent communication, outperforming text-based and prior latent methods by a wide margin.
Multi-agent systems built on large language models have shown strong performance on complex reasoning tasks, yet most work focuses on agent roles and orchestration while treating inter-agent communication as a fixed interface. Latent communication through internal representations such as key-value caches offers a promising alternative to text-based protocols, but existing approaches do not jointly optimize communication with multi-agent reasoning. Therefore we propose DiffMAS, a training framework that treats latent communication as a learnable component of multi-agent systems. DiffMAS performs parameter-efficient supervised training over multi-agent latent trajectories, enabling agents to jointly learn how information should be encoded and interpreted across interactions. Experiments on mathematical reasoning, scientific QA, code generation, and commonsense benchmarks show that DiffMAS consistently improves reasoning accuracy and decoding stability over single-agent inference, text-based multi-agent systems, and prior latent communication methods, achieving 26.7% on AIME24, 20.2% on GPQA-Diamond, and consistent gains across reasoning benchmarks.
