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-agentsystems still rely solely on natural language, exchanging tokens or theirembeddings. To go beyond language, we introduce a new paradigm, thoughtcommunication, which enables agents to interact directly mind-to-mind, akin totelepathy. To uncover these latent thoughts in a principled way, we formalizethe process as a general latent variable model, where agent states aregenerated by an unknown function of underlying thoughts. We prove that, in anonparametric setting without auxiliary information, both shared and privatelatent thoughts between any pair of agents can be identified. Moreover, theglobal structure of thought sharing, including which agents share whichthoughts and how these relationships are structured, can also be recovered withtheoretical guarantees. Guided by the established theory, we develop aframework that extracts latent thoughts from all agents prior to communicationand assigns each agent the relevant thoughts, along with their sharingpatterns. This paradigm naturally extends beyond LLMs to all modalities, asmost observational data arise from hidden generative processes. Experiments onboth synthetic and real-world benchmarks validate the theory and demonstratethe collaborative advantages of thought communication. We hope this workilluminates the potential of leveraging the hidden world, as many challengesremain unsolvable through surface-level observation alone, regardless ofcompute or data scale.