Thought Communication in Multiagent Collaboration (2510.20733v1)

Abstract: Natural language has long enabled human cooperation, but its lossy, ambiguous, and indirect nature limits the potential of collective intelligence. While machines are not subject to these constraints, most LLM-based multi-agent systems still rely solely on natural language, exchanging tokens or their embeddings. To go beyond language, we introduce a new paradigm, thought communication, which enables agents to interact directly mind-to-mind, akin to telepathy. To uncover these latent thoughts in a principled way, we formalize the process as a general latent variable model, where agent states are generated by an unknown function of underlying thoughts. We prove that, in a nonparametric setting without auxiliary information, both shared and private latent thoughts between any pair of agents can be identified. Moreover, the global structure of thought sharing, including which agents share which thoughts and how these relationships are structured, can also be recovered with theoretical guarantees. Guided by the established theory, we develop a framework that extracts latent thoughts from all agents prior to communication and assigns each agent the relevant thoughts, along with their sharing patterns. This paradigm naturally extends beyond LLMs to all modalities, as most observational data arise from hidden generative processes. Experiments on both synthetic and real-world benchmarks validate the theory and demonstrate the collaborative advantages of thought communication. We hope this work illuminates the potential of leveraging the hidden world, as many challenges remain unsolvable through surface-level observation alone, regardless of compute or data scale.

• The paper introduces ThoughtComm, a method that enables direct exchange of latent thoughts among agents, overcoming the limitations of token-based communication.
• It employs a sparsity-regularized autoencoder to extract and structure latent thoughts, integrating them as prefixes for enhanced response generation.
• Experiments on math reasoning benchmarks demonstrate improved accuracy and consensus among agents, validating the framework’s scalability and efficiency.

"Thought Communication in Multiagent Collaboration" Summary

Introduction

This paper challenges traditional communication paradigms in multi-agent systems (MAS), which primarily rely on natural language communication through tokens or embeddings exchanged among LLMs. The established norm of using natural language presents limitations: sequential nature, ambiguity, and indirect expression of latent thoughts. The authors propose "thought communication," a novel paradigm where agents share latent thoughts directly, bypassing the constraints inherent in language-based exchange. This innovation is rooted in a formalization of inter-agent communication as a latent variable model, where agent states stem from an unknown function of underlying thoughts. The authors lay out theoretical guarantees for recovering shared and private latent thoughts and their structures among agents.

ThoughtComm Framework

Latent Thought Extraction:

The ThoughtComm framework starts by encoding agents' model states into a common latent space using a sparsity-regularized autoencoder. This process reveals latent thoughts from the agents' model states (a concatenated form of these states per communication round). The autoencoder enforces sparsity, reflecting the minimal shared cognitive basis among agents and enabling the extraction of latent thoughts with identified structures.

Structural Information Integration:

Latent thoughts, once extracted, are redistributed among agents based on structured dependency. This involves:

• Identifying relevant thoughts for each agent from the latent space,
• Grouping thoughts by how widely they apply across agent interactions (shared vs. private),
• Reweighting these thoughts to reflect their consensus among agents before reintroducing them into their models as prefixes in subsequent communication rounds.

Prefix Adaptation:

Reintegration into agent operations is handled through “prefix adaptation.” An adapter transforms selected latent thoughts into a prefix vector, which is prepended to an agent’s input context, guiding the generation of the agent's subsequent responses. This approach supports continuous latent guidance in multi-agent reasoning, extending beyond fixed natural language exchanges.

Theoretical Underpinnings

The authors provide clear identifiability results ensuring that the recovered latent thoughts accurately reflect the underlying cognitive processes of agents:

• Shared Thoughts: The framework guarantees that thoughts shared between any two agents are recoverable, disentangled from private thoughts.
• Private Thoughts: Theories also support recovering private thoughts, which are uniquely held by individual agents.
• Thought Structure: The structure linking latent thoughts to specific agents, indicating which agents share which thoughts, is identifiable up to permutation.

These identifiability results are grounded in classical theory but extend into the nonlinear, high-dimensional interaction spaces typical of LLMs.

Experiments

Experiments spanning synthetic and real-world benchmarks validate the advantages thought communication offers over previous approaches. Tests on math reasoning datasets (e.g., MATH and GSM8K) demonstrate significant improvements in accuracy and consensus among agents, across diverse LLM models and sizes. These experiments establish that ThoughtComm not only increases collaborative effectiveness but also scales efficiently even with growing numbers of agents or debate rounds.

Conclusion

By formalizing thought as a direct communication medium among agents, ThoughtComm overcomes the inherent limitations of language-based MAS systems. Though practical limitations around accessing model states exist, the paper proposes embedding alternatives to extend applicability. Future work could explore non-textual modalities using this paradigm, implying vast potential in areas beyond text and natural language, significantly enriching multi-agent collaborative capabilities. Overall, the paper provides a strong theoretical and practical foundation for the proposed communication paradigm, offering a scalable, efficient, and comprehensive solution for enhancing collective intelligence in MAS.