PaperlayerView exact revision on arXiv

Source-linked AI summary

Thought Communication in Multiagent Collaboration

Yujia Zheng, Zhuokai Zhao, Zijian Li, Yaqi Xie, Mingze Gao, Lizhu Zhang, Kun Zhang

arXiv:2510.20733v1cs.LGcs.AIcs.MA

TL;DR

LLM multi-agent systems typically communicate through natural language, which is lossy, ambiguous, and indirect. This paper formalizes thought communication through a latent-variable model, proves recovery of shared and private thoughts, and develops THOUGHTCOMM, which improves collaboration across synthetic and real-world evaluations.

  • Problem

    LLM multi-agent systems typically exchange natural-language tokens or embeddings, despite language’s lossy, ambiguous, and indirect reflection of thought.

  • Method

    The paper models agent states as generated from latent thoughts, establishes identifiability guarantees, and uses a sparsity-regularized autoencoder to extract and assign shared or private thoughts.

  • Results

    THOUGHTCOMM consistently outperforms baseline methods across MATH and GSM8K, achieving 93% MATH accuracy on Qwen 3-1.7B and a 67.23% relative improvement over single answer on average.

  • Takeaways & Limitations

    Thought communication provides a framework for leveraging latent agent information to improve multi-agent collaboration and potentially extend beyond language and LLMs.

  • Takeaways & Limitations

    The experiments use model states as observed variables, which may not be feasible for closed-source models; the theory also assumes an invertible, twice-differentiable generating function.

Abstract

from arXiv · show

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.

1 Introduction

The paper proposes thought communication to overcome the ambiguity and indirectness of language-based multi-agent collaboration. It formalizes latent thoughts and provides theory and a framework for sharing relevant thoughts among agents.

  • Motivation: Existing LLM multi-agent systems exchange natural-language tokens or embeddings, leaving collaboration constrained by language’s ambiguity, imprecision, and indirectness.The paper links vague message specification and inter-agent misalignment to these communication limits.
  • Thought Communication: Thought communication directly transfers latent thoughts underlying agents’ goals, beliefs, and reasoning instead of exchanging surface-level messages.The proposed paradigm is intended to bypass ambiguity and distortion in language-based communication.
  • Theory: The paper models agent states as outputs of an unknown function of latent thoughts and establishes nonparametric identifiability guarantees for recovering them.The formal model represents pre-communication states as Ht = f(Zt).
  • Theory: Shared and private latent thoughts, along with the structure linking thoughts to agents, can be recovered under sparsity regularization.Synthetic experiments confirm the validity of these theoretical results.
  • Framework: A framework extracts latent thoughts from agent states and assigns each agent relevant thoughts with their sharing structure, improving collaboration beyond language.Experiments across diverse models and scenarios support the collaborative benefits of this communication approach.

2 Problem Formulation

The paper formulates agent communication as a latent generative process in which an unknown function maps latent thoughts to model states. It defines agent-relevant thoughts through the function’s dependency structure and illustrates shared and private influences on decisions.

  • Data-generating process: The data-generating process maps latent thoughts Zt to all agents’ model states Ht through an unknown, invertible, twice-differentiable function f.The model states are the foundation-model hidden-layer representations before communication.
  • Example: In the airport example, thoughts about luggage, speed, and punctuality influence agents’ responses, with agents selecting different subsets of those thoughts.Agent 1 chooses a car based on luggage and punctuality, while Agent 2 selects a train under the influence of speed.
  • Thought structure: The thought-sharing structure determines which latent thoughts are common, private, irrelevant, or distracting for particular agents.The framework therefore aims to reveal only thoughts relevant to each agent’s goals and role.
  • Thought structure: A latent thought is relevant to an agent when it influences at least one component of that agent’s model state, identified by the non-zero pattern of the Jacobian of f.This dependency pattern is represented as a binary matrix linking thought components to state components.
  • Data-generating process: Each agent’s model state is a slice of the global state, corresponding to the components associated with that agent.The global state concatenates the model-state components of all agents.

3 Identifiability Theory

The paper develops identifiability guarantees for recovering shared, private, and structural relationships among latent thoughts from agent states under sparse, nonparametric assumptions. It deliberately targets pairwise, communication-relevant recovery rather than full latent-variable recovery, which is impossible without stronger assumptions or auxiliary information.

  • 3.1 Identifiability of Shared Thoughts: Theorem 1 guarantees that shared thoughts between any agent pair can be disentangled from other latent variables, up to permutation.The guarantee relies on the stated Jacobian support and sparsity conditions.
  • 3.4 Discussion on Theoretical Contribution: The theory recovers true shared components under invertibility and avoids misattributing decisions to unrelated latent causes when private thoughts are disentangled.The discussion connects these guarantees to faithful common ground and preservation of meaningful, less-common thoughts.
  • 3.2 Identifiability of Private Thoughts: Theorem 2 extends identifiability to private thoughts, separating an agent-specific component from the remaining latent variables for each agent pair.Pairwise decompositions can then reveal thoughts that are unique or rare within larger groups.
  • 3.3 The Structure of Thoughts: Theorem 3 identifies the nonzero thought-agent dependency pattern up to relabeling, revealing which agents hold which thoughts and which thoughts are shared.This structure-level recovery supports reconstruction of larger thought-agent incidence patterns.
  • 3.4 Discussion on Theoretical Contribution: Rather than requiring auxiliary signals or global recovery, the approach uses basic assumptions to obtain coarser shared/private guarantees that remain meaningful for communication.The paper contrasts this pairwise target with prior identifiability settings using weak supervision, restricted function classes, or structural graph conditions.

4 THOUGHTCOMM: Multiagent Communication via Thought

THOUGHTCOMM extracts latent thoughts from concatenated agent states, recovers their dependency structure, and selectively routes personalized thought representations to agents. It injects those representations through prefix adaptation so agents can use latent communication without explicit message passing.

  • 4.1 Uncovering the Latent Thoughts: THOUGHTCOMM encodes all agents’ model states into a shared latent space, then uses the recovered dependency structure to route relevant thought dimensions to each agent.The process distinguishes shared and private thoughts before reintegrating them into agent contexts.
  • 4.1 Uncovering the Latent Thoughts: The concatenated state Ht is mapped through a sparsity-regularized autoencoder to recover latent thoughts ˆZt and their agent-relevance structure.The encoder is trained to reconstruct the full state vector while enforcing Jacobian sparsity, then extracts thoughts at test time.
  • 4.2 Leveraging the Structure of Thoughts: Agreement-based reweighting groups each agent’s relevant thoughts by how many agents’ hidden states depend on each latent dimension.Each agreement group receives a distinct weight reflecting the thought’s relevance or generality across agents.
  • 4.2 Leveraging the Structure of Thoughts: A structural mask filters latent dimensions to those relevant for each agent, shaping the personalized representation used for subsequent generation.This routing step directly determines which thought information is injected into each agent.
  • 4.3 Latent Injection via Prefix Adaptation: Prefix adaptation converts each personalized latent representation into a prefix vector that is prepended to token embeddings, guiding generation without explicit message passing.The adapter is trained with short continuations, semantic similarity, and regularization intended to keep injected effects linguistically natural.

5 Experiments

Experiments test latent-thought identifiability synthetically and THOUGHTCOMM’s collaborative effectiveness on math benchmarks. Results show recovery of shared/private thoughts and consistent gains over single-agent and multiagent-finetuning baselines.

  • 5.1 Synthetic Evaluation: The sparsity-regularized model identifies shared and private latent regions, whereas the non-sparse baseline fails to disentangle them.R2 measures correspondence between estimated latent variables and matching ground-truth components.
  • 5.2 Real-World Evaluation: THOUGHTCOMM consistently outperforms single answer and Multiagent Finetuning on MATH and GSM8K across five LLMs.On average, it improves relative performance by 67.23% over single answer and 19.06% over the current state of the art.
  • 5.2 Real-World Evaluation: 93% MATH accuracy on Qwen 3-1.7B represents a 17.2% absolute gain over Multiagent Finetuning and a 113.3% relative gain over single answer.
  • 5.2 Real-World Evaluation: THOUGHTCOMM’s training overhead depends on embedding dimension rather than LLM parameter count, making it smaller and model-agnostic than full-model finetuning.For Llama-3-70B and 405B, the shared 16,384-dimensional embedding size keeps THOUGHTCOMM’s overhead unchanged.
  • 5.4 Varying the Debate Rounds: Across increasing debate rounds, Multiagent Finetuning loses accuracy while THOUGHTCOMM gains in both accuracy and consensus.The comparison uses two agents on MATH with debate rounds varying from 2 to 6.
  • 5.3 Varying the Prefix Lengths: Accuracy and consensus remain stable as prefix length increases from 1 to 16, with fluctuations under five percent across four models.The results indicate robustness to prefix-length tuning and near-optimal performance with a single injected vector.

6 Related Works

Related work covers communication strategies for LLM-based multi-agent systems and identifiability theory for latent-variable models. The paper positions its approach as combining latent-thought communication with nonparametric structural recovery.

  • Multiagent LLMs communication: LLM multi-agent research explores communication efficiency, flexible workflows, error propagation, token-level collaboration, and alternatives to text-token exchange.
  • Identifiability of latent variable models: The paper differs by targeting shared and private latent thoughts from agent pairs without the auxiliary signals commonly used for full latent recovery.
  • Identifiability of latent variable models: Classical latent-variable identifiability studies emphasize linear models, while later work relaxes linearity using auxiliary variables and related constraints.

7 Conclusion

The conclusion frames THOUGHTCOMM as a latent-variable approach to thought-based multi-agent communication while acknowledging limits of relying on model states. It notes that the theory can extend to nontextual observational data, though this extension is not empirically tested.

  • 7 Conclusion: THOUGHTCOMM formulates multi-agent communication as a latent-variable model and establishes identifiability results for recovering latent thoughts and their structures.
  • 7 Conclusion: Using model states as observed variables may be infeasible for closed-source models, motivating context-aware embeddings as an alternative.
  • 7 Conclusion: The framework can accommodate observational data from any modality, but this broader extension has not been explored empirically.

A.1 Proof of Theorem 1

The proof establishes that sparsity and Jacobian support structure identify latent thoughts up to permutation. It uses invertibility, support-subspace assumptions, and a bipartite-graph argument to recover shared and private structure.

  • A.1 Proof of Theorem 1: Jacobian support-subspace spanning assumptions let the proof construct vectors and compare the nonzero patterns of the true and estimated mappings.
  • A.1 Proof of Theorem 1: Invertibility permits an invertible change of variables between true and estimated latent representations.The proof defines h = ˆf^-1 ◦ f and differentiates the relation ˆZt = h(Zt).
  • A.1 Proof of Theorem 1: A bipartite graph built from the nonzero pattern of an invertible matrix enables a permutation matching through Hall’s marriage theorem.
  • A.1 Proof of Theorem 1: The resulting permutation preserves which latent components influence observed states, yielding correspondence between true and estimated latent coordinates.
  • A.1 Proof of Theorem 1: The theorem concludes that the Jacobian’s nonzero pattern is identifiable up to relabelling under invertibility and ℓ0 regularization.Formally, B(J ˆ f) = B(Jf)P for a permutation matrix P.

A.2 Proof of Theorem 2

The proof establishes that private latent thoughts are identifiable under the theorem’s assumptions, using invertibility and sparsity to align recovered representations.

  • A.2 Proof of Theorem 2: Theorem 2 identifies private thoughts up to a permutation of latent variables under the stated assumptions.The proof derives this correspondence through an invertible change of variables and Jacobian support relationships.
  • A.2 Proof of Theorem 2: The recovered Jacobian has the same nonzero pattern as the original after a latent-variable permutation.This support equivalence is stated explicitly for every pair of latent coordinates.

A.3 Proof of Theorem 3

Theorem 3 shows that the structure linking latent thoughts to agents is identifiable up to relabelling, and the framework can substitute response embeddings when model states are inaccessible.

  • A.3 Proof of Theorem 3: Theorem 3 identifies the thought-sharing structure up to a permutation of latent variables.The nonzero pattern of the recovered Jacobian equals the original pattern multiplied by a permutation matrix.
  • A.3 Proof of Theorem 3: The proof establishes structural equivalence by preserving Jacobian support under an invertible change of variables and ℓ0 regularization.The resulting support equivalence holds coordinatewise after relabelling.
  • Alternative to model states: When model states are unavailable, compact embeddings of textual responses can serve as surrogate inputs to the same latent-thought pipeline.The framework then continues with latent recovery, thought routing, and prefix adaptation.
  • Alternative to model states: The embedding-based replacement supports deployment across open- and closed-source agents and can extend to other modalities.The proposed encoders need only summarize response content at the linguistic level.

C.2 Additional Results on Scaling Debate Rounds

Across additional Qwen-3-1.7B experiments, THOUGHTCOMM remains stable or improves as debate rounds increase, whereas Multiagent Finetune declines beyond four rounds.

  • C.2 Additional Results on Scaling Debate Rounds: THOUGHTCOMM accuracy and consensus remain stable or improve through six debate rounds, while Multiagent Finetune declines beyond four.The comparison is reported for additional Qwen-3-1.7B results.
  • C.2 Additional Results on Scaling Debate Rounds: High consensus can coexist with stagnant or degraded accuracy, producing premature agreement rather than genuine reasoning improvement.This failure mode is reported for Multiagent Finetune as debate rounds increase from two to six.
  • C.2 Additional Results on Scaling Debate Rounds: THOUGHTCOMM aligns increased agreement with improved accuracy and avoids degradation from additional communication rounds.The reported result supports using more rounds without the decline observed in prior debate-style frameworks.

C.3 Additional Results on Varying Latent Dimensions

Increasing latent dimensionality improves MATH accuracy until gains saturate, while adding agents helps initially and THOUGHTCOMM remains stable as agent count grows.

  • C.4 Additional Results on Varying Number of Agents: Both models gain notably when increasing from two to three agents, but additional agents yield plateaus or slight declines.The comparison uses two rounds, latent dimension 1024, and one prefix token on MATH.
  • C.4 Additional Results on Varying Number of Agents: THOUGHTCOMM maintains stable accuracy as the number of agents grows, indicating robustness to redundant or conflicting signals.The baseline is especially prone to plateauing or declining beyond three agents.
Loading 2510.20733v1…