Thought Communication in Multiagent Collaboration

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.

Natural language has enabled human collaboration at scale, but it also imposes fundamental limitations. While powerful, language is inherently sequential, ambiguous, and imprecise, offering only an indirect and fragmented reflection of thought [von Humboldt, 1988]. This constraint is deeply rooted in human cognition, which lacks direct channels for transmitting mental content. Machines, however, are not subject to the same physical constraints of speech or perception. This difference may be one of the central reasons why superhuman intelligence is possible. Every transformative achievement, from scientific discovery to societal progress, relies on collaboration. Likewise, superhuman intelligence will require not only individual reasoning beyond human capability but also collective reasoning beyond human coordination [Vinge, 1993]. This calls for a new form of communication that transcends the limits of language.

However, existing large language model (LLM)-based multi-agent systems (MAS) rely on natural language as the medium of communication, exchanging information via tokens or their embeddings [Du et al., 2023, Liang et al., 2023, Pham et al., 2023, Zhang et al., 2024a, Zeng et al., 2025, Wang et al., 2025b]. These systems typically assume that multiple LLM agents exchange natural language messages to convey internal ideas and coordinate toward a shared goal. However, natural language remains fundamentally limited in its ability to express the underlying latent thoughts that drive reasoning and decision making. As a result, current systems remain restricted by the bottlenecks of language, limiting their potential for superhuman collaboration. Indeed, recent empirical analyses [Cemri et al., 2025, Hu et al., 2025] highlight that many failures in inter-agent collaboration stem from vague message specification and inter-agent misalignment, both ultimately caused by the indirect nature of lossy language-based communication. Then, the core question reveals itself:

What form of communication goes beyond the limits of language?

To answer this, we turn to the idea of communication through latent thoughts. Nothing is more direct than transmitting what one truly thinks, i.e., telepathy. Just as human actions are guided by internal mental states, agents likely operate based on latent representations that encode goals, beliefs, and reasoning. If these could be identified, agents could share them directly, bypassing the ambiguity and distortion of language. This enables a fundamentally different mode of communication, based not on the exchange of surface tokens or their embeddings, but on the direct transfer of intent and understanding. Furthermore, in multi-agent settings, some thoughts are intended to be broadly shared, while others are inherently private or uniquely tailored to certain individual agents. Revealing both the latent thoughts and their structural organization allows agents to better detect alignment, resolve conflicts, and integrate diverse reasoning paths.

Contributions: We formalize this idea by introducing a latent generative model for inter-agent communication. Specifically, we assume that the model states Ht of all agents before communication round t are generated from a set of latent thoughts Zt through an unknown function f, such that Ht = f(Zt). We establish both a nonparametric identifiability result that guarantees recovery of latent thoughts, and a general framework that facilitates direct mind-to-mind communication. Theoretically, we prove that in a general nonparametric setting, both shared and private latent thoughts can be identified from hidden states under a sparsity regularization. Our identifiability result ensures that the recovered latent representations reflect the true internal structure of agent reasoning. Moreover, we show that the structures between thoughts and individual agents can be reliably recovered, enabling a provable correspondence between agents and their cognitive content. Experiments on various synthetic environments confirm the validity of the theory.

Practically, we develop a principled framework for latent communication among agents. Guided by the theory, we implement a sparsity-regularized autoencoder to extract latent thoughts from agent hidden states and infer the underlying mapping between agents and these thoughts. Each agent is equipped with a set of inferred thoughts, along with the structure of how each thought is shared. This allows agents not only to understand what others are thinking but also to reason about which thoughts are mutually held or privately maintained. Experiments across diverse models and scenarios demonstrate that communication beyond language directly benefits collaboration among LLM agents.