Latent Collaboration in Multi-Agent Systems

Researchers introduce LatentMAS, a framework enabling LLM agents to collaborate directly in latent space rather than through text, achieving up to 14.6% higher accuracy while reducing token usage by 70.8%-83.7% and improving inference speed 4× faster than text-based multi-agent systems.

LatentMAS represents a fundamental shift in how language model agents can interact with one another. Rather than converting internal representations to text for communication—a process that introduces information loss and computational overhead—agents now exchange continuous latent embeddings directly through a shared working memory. This approach preserves the full richness of model reasoning without the bottleneck of discrete tokenization.

The innovation builds on established trends in model interpretability and efficient reasoning. Previous multi-agent systems relied on natural language as an interface between agents, mirroring human collaboration but inheriting its inefficiencies. By operating in the continuous latent space where models naturally reason, LatentMAS bypasses this constraint. The framework requires no additional training, making it immediately deployable across existing LLM architectures.

For developers and researchers, this work carries significant implications. The 70% reduction in output tokens directly translates to lower inference costs—a critical metric for production systems. The 4× speedup enables real-time collaborative reasoning at scale. The theoretical analysis demonstrating higher expressiveness suggests latent collaboration captures nuances that text-based systems inherently lose through discretization. Performance gains across nine diverse benchmarks spanning mathematics, science, code generation, and commonsense reasoning indicate broad applicability rather than narrow optimization.

The open-sourced nature of the work accelerates adoption. Teams building reasoning systems—whether for scientific discovery, complex problem-solving, or autonomous agents—now have access to a production-ready framework. Future research likely builds on this foundation, exploring optimal latent communication protocols and scaling to larger agent collectives. The work fundamentally challenges assumptions about information exchange in multi-agent systems.

• →LatentMAS enables agents to communicate through continuous embeddings instead of text, eliminating re-encoding overhead and information loss.
• →The framework achieves up to 14.6% accuracy improvement while reducing token usage by 70.8%-83.7% and providing 4× faster inference.
• →Direct latent collaboration requires no additional training and works with existing LLM architectures, enabling immediate deployment.
• →Performance gains across math, science, code generation, and commonsense tasks demonstrate broad applicability beyond narrow problem domains.
• →Open-sourced code and theoretical analysis provide foundation for future research in efficient multi-agent reasoning systems.