End-to-End Optimization of Multi-Agent Communication via Latent Representations Outperforms Natural Language

arXiv·medium signal

Yu et al. propose differentiable end-to-end optimization of multi-agent language systems, replacing natural-language inter-agent communication with learned latent representations. While most multi-agent research focuses on orchestration and agent roles, this work treats the communication channel itself as a learnable parameter. Results show latent communication significantly outperforms natural language on complex reasoning tasks, suggesting the verbosity of natural language is a bottleneck in multi-agent pipelines.