Thought Communication in AI

• Thought communication is the externalization and transfer of internal cognitive states, enabling AI transparency and collective intelligence.
• It encompasses techniques such as visible thinking, inner monologues, and latent variable transfer to improve reasoning and multiagent collaboration.
• Current frameworks show measurable gains in efficiency and user engagement across dialogue systems, multiagent tasks, and human-AI interfaces.

Thought communication encompasses the processes and mechanisms by which internal cognitive states—such as reasoning steps, intentions, values, or emotions—are externalized, shared, or transferred between agents, whether human, artificial, or hybrid. In contemporary AI and cognitive modeling, this concept spans explicit reasoning traces in LLMs, latent variable transfers in multiagent systems, inner monologues in dialogue agents, and protocol learning in emergent communication. Empirical and theoretical advances have reframed thought communication as a multiscale, modality-independent substrate for both human-AI interaction and collective intelligence, with applications ranging from transparency in LLMs to efficient multiagent collaboration, as well as foundational implications for the language of thought hypothesis and interface design.

1. Foundations and Definitions

Thought communication diverges from conventional information exchange by focusing on the transfer or externalization of internal cognitive structures. In LLMs and multiagent systems, thoughts may be instantiated as:

• Visible thinking: Short, value- or intention-revealing statements exposed to users as part of an interaction, but distinct from chains of logical inference (Cox et al., 23 Jan 2026).
• Internal monologues: Private, skill-annotated reasoning steps not exposed to users but shaping downstream communication (Zhou et al., 2023).
• Structured reasoning traces: Chains or matrices of intermediate computations (chain-of-thought, matrix-of-thought, etc.) interleaving subproblem solutions and context (Tang et al., 4 Sep 2025, Kudo et al., 2024).
• Latent variable exchange: Direct sharing of internal hidden states between agents, potentially bypassing natural language (“mind-to-mind” communication) (Zheng et al., 23 Oct 2025).
• Emergent protocols: Learned, non-symbolic codes for intention transmission in multiagent reinforcement learning (Zhang, 19 Mar 2026).

Definitions vary by context. For example, in multiagent latent variable models, “thoughts” are unobserved components $Z_{t,j}$ in a generative process mapping to agents’ observable states $H_t^{(k)}$ (Zheng et al., 23 Oct 2025). In LLM-based dialog systems, visible thinking comprises one- or two-sentence utterances using first-person language to disclose values or ethical stances, not inference chains (Cox et al., 23 Jan 2026). The domain thus includes both explicit, natural-language “thoughts” and sub-symbolic, protocol-level transmissions.

2. Canonical Frameworks and Algorithms

Several distinct frameworks operationalize thought communication:

2.1 Explicit Reasoning and Inner Monologue

• CSIM (Cultivating Skills via Inner Monologue): Interposes a latent monologue $M$ (“think” stage) before generating each response $R$ (“speak” stage), thereby enabling explicit reasoning about empathy, topic transition, summarization, and other communication skills. No additional parameters or losses are introduced; the same backbone model generates both $M$ and $R$ with prompt conditioning (Zhou et al., 2023).
• Matrix of Thought (MoT): Embeds reasoning in a 2D array $M\in\mathbb{R}^{m\times n}$ of thought nodes, allowing parallel multi-strategy search via “column-cell communication” with explicit context transfer governed by a weight matrix $\alpha_{i}^{\,j}$. Inter-unit context exchange is finely controlled, enabling both diversity and information consolidation (Tang et al., 4 Sep 2025).
• Exchange-of-Thought (EoT): Generalizes to multi-LLM networks, modeling cross-model communication via well-defined topologies (bus, star, ring, debate tree) and explicit protocols for reasoning and confidence tracking (Yin et al., 2023).

2.2 Latent and Non-Linguistic Thought Transfer

• ThoughtComm (Latent Variable Model): Employs a sparsity-regularized autoencoder to extract latent “thoughts” $Z_t$ from agent states, identifies shared/private subspaces, and uses prefix adapters to inject disentangled latent variables directly into agent generation pipelines, enabling “mind-to-mind” interaction without text (Zheng et al., 23 Oct 2025).
• Efficiency Attenuation Phenomenon (EAP): Demonstrates that agents using emergent, inscrutable codes for agent-agent interaction can achieve 50.5% higher efficiency than their counterparts using human-defined, symbolic protocols in navigation tasks (Zhang, 19 Mar 2026).

2.3 Thought Communication in Dialogue and Speech

• Visible Thinking in Chatbots: Brief value- or expertise-oriented reflections (not chains of reasoning) interposed before responses affect user perceptions of empathy, trust, and warmth, functioning as a social cue rather than as an explanation of computation (Cox et al., 23 Jan 2026).
• DiffuSpeech: Speech LLMs generate internal text “thought traces” along with speech output in a unified masked diffusion process, allowing model reasoning to be inspected (and potentially edited) before speech synthesis (Lou et al., 30 Jan 2026).
• Orality: Transforms spoken input into a semantically organized node-link canvas, supporting explicit structuring, “thinking aloud,” and real-time AI scaffolding for idea clarification (Li et al., 3 Mar 2026).

3. Empirical Evaluations and Comparative Results

Empirical studies across diverse modalities substantiate both the utility and modality dependence of thought communication:

System	Domain	Quantitative Gains	Key Mechanism
MTQA/MoT	QA reasoning	F1 +3–4.2 pts, EM +3.9–4.5 pts over ToT	2D matrix reasoning with α-control
CSIM	Dialogue	Humanness +1.0–1.2 pts, Proactivity +1.03 pts	Inner monologue conditioning
EoT	Reasoning	+3–3.3 pts over CoT and PHP, same cost as SC(3–5)	Model-to-model rationale exchange
ThoughtComm	Math (multiagent)	Qwen-1.7B/MATH: 93% vs. 75.8% (Multi-FT)	Latent variable, prefix injection
DiffuSpeech	Speech QA	S→S: +5.5 pts, TTS WER 6.2% best among models	Text reasoning + speech diffusion
TBS	Social Simulation	Internal-state trace coherence across simulations	Private-to-public, stepwise reasoning

In user studies, visible thinking boosts chatbot-perceived empathy (≈55.6 vs. 45.8 control) and warmth but can cause negative expectancy violations if the advice remains neutral (Cox et al., 23 Jan 2026). Orality increases post-task thought-clarity and supports divergent, self-guided sensemaking over transcript-based baselines (Li et al., 3 Mar 2026). CSIM, by adding invisible inner-monologue steps, increases multi-turn engagement and goal success by over 400% (0.18→0.93) (Zhou et al., 2023). DiffuSpeech demonstrates that explicit thinking traces in speech-to-speech QA lead to a +13.4 point accuracy gain (Lou et al., 30 Jan 2026).

4. Thought Communication in Multiagent and Multimodal Systems

Recent work has emphasized the advantages of direct thought transfer in collective settings:

• Identifiability and Sharing Structure: Under a nonparametric, invertible mixing model, both shared and private latent thoughts can be recovered with theoretical guarantees, and the full incidence matrix $B(J_f)$ detailing which agents share which thoughts is identifiable up to permutation under mild assumptions (Zheng et al., 23 Oct 2025).
• Protocol Emergence and Efficiency: MARL experiments show that private, symbolically inscrutable emergent codes are not only viable for high-efficiency cooperation but also empirically superior to hand-designed symbolic protocols, challenging the universality of the language-of-thought hypothesis (Zhang, 19 Mar 2026).
• Causal Probing in LLMs: Activation patching and probe-based analyses reveal that LLMs instantiate a stratified reasoning regime: trivial subproblems are solved pre-chain-of-thought (“Think-to-Talk”), while complex, multi-step answers are computed during chain-of-thought emission (“Talk-to-Think”) (Kudo et al., 2024).
• Social Simulation (TBS): Separates agent-internal evaluation from public utterance generation, with structured private states (opinion, dissonance, motivation, isolation risk) and orchestrated public floor allocation, thus making pathways from thought to communication observable (Yang et al., 2 Jun 2026).

5. User Thought Capture and HCI Implications

Beyond model-to-model protocols, human thought communication in human-AI interaction is now a tractable empirical domain:

• ThoughtTrace Dataset: Large-scale collection (10,174 annotations) of self-reported user “reasons” and “reactions” paired with real-world LLM conversations reveals that thoughts are semantically distinct, hard to infer from context, and provide significant advances in downstream behavior prediction (+41.7% relative gain) and personalized alignment (+3.9 points in Arena-Hard alignment task) (Jin et al., 19 May 2026).
• Orality and Thought Externalization: By mapping spoken input to a persistent canvas of interlinked thoughts, systems can scaffold sensemaking, enable error/conflict detection, and support mixed-initiative, metacognitive workflows (Li et al., 3 Mar 2026).

These findings establish that elicited or modeled user thoughts enable more responsive and adaptable AI assistants and make latent cognition a measurable, actionable modality.

6. Open Problems and Theoretical Controversies

Foundational research exposes critical tensions and future directions:

• Symbolic vs. Sub-Symbolic Format: The efficiency attenuation phenomenon directly challenges the necessity of language-like structure for cognition; emergent protocols in deep MARL are argued to act as sub-symbolic “thought-communication” more effective than compositional, explicit code (Zhang, 19 Mar 2026). This raises questions about interpretability, alignment, and cognitive pluralism.
• Transparency vs. Expectation: While value-focused visible thinking boosts rapport, it may foster miscalibrated expectations (“over-anthropomorphization”, “negative expectancy violation”), suggesting a need for adaptive, context-sensitive communication strategies (Cox et al., 23 Jan 2026).
• Latent Variable Identifiability: Guarantees currently rely on the existence of invertible, sufficiently sparse or structured mixing functions; practical identification in deep, noisy, or only partially observable systems remains a challenge (Zheng et al., 23 Oct 2025).
• Faithfulness of Explanations: Results on chain-of-thought show intricacy: LLM explanations may be post-hoc or genuinely stepwise depending on subproblem complexity (Kudo et al., 2024).
• Modality-Independent Protocols: Extending latent thought communication to multimodal, sensorimotor, or robotic agents is an active area, motivated by the desire to transcend language bottlenecks and tackle tasks inherently ill-suited to natural language.

A plausible implication is that future intelligent systems will instantiate hybrid regimes, combining directly shared latent thought subspaces with user-facing externalizations structured to optimize transparency, trust, and control across modalities and agent types.

7. Summary Table of Representative Frameworks

Framework / Paper	Modality	Nature of Thought Comm.	Main Evaluation Result
MTQA / MoT (Tang et al., 4 Sep 2025)	Textual QA	Multi-strategy explicit	F1/EM +4 pts, 14.4% reasoning time
CSIM (Zhou et al., 2023)	Dialogue	Inner monologue (latent)	+1.0 Humanness, 4.63 → 2.13 turns
ThoughtComm (Zheng et al., 23 Oct 2025)	Multiagent	Latent variable, prefix	93.0% accuracy (Qwen-1.7B)
DiffuSpeech (Lou et al., 30 Jan 2026)	Speech	Reasoning-speech diffusion	+5.5 QA pts, TTS WER 6.2%
Visible Thinking (Cox et al., 23 Jan 2026)	HCI	Value/stance utterances	Empathy +10 pts, Warmth +0.7
TBS (Yang et al., 2 Jun 2026)	Soc. sim	Private-public separation	Coherent internal-state traces
EoT (Yin et al., 2023)	Model ensemble	Cross-model rationale	+3.3% over PHP, 20% less cost
ThoughtTrace (Jin et al., 19 May 2026)	HCI	User thought annotation	+41.7% behavior prediction gain

• “Watching AI Think: User Perceptions of Visible Thinking in Chatbots” (Cox et al., 23 Jan 2026)
• “MTQA: Matrix of Thought for Enhanced Reasoning in Complex Question Answering” (Tang et al., 4 Sep 2025)
• “Think Before You Speak: Cultivating Communication Skills via Inner Monologue” (Zhou et al., 2023)
• “DiffuSpeech: Silent Thought, Spoken Answer via Unified Speech-Text Diffusion” (Lou et al., 30 Jan 2026)
• “The Efficiency Attenuation Phenomenon: A Computational Challenge to the Language of Thought Hypothesis” (Zhang, 19 Mar 2026)
• “ThoughtTrace: Understanding User Thoughts in Real-World LLM Interactions” (Jin et al., 19 May 2026)
• “Think-Before-Speak: From Internal Evaluation to Public Expression in Multi-Agent Social Simulation” (Yang et al., 2 Jun 2026)
• “Exchange-of-Thought: Enhancing LLM Capabilities through Cross-Model Communication” (Yin et al., 2023)
• “Thought Communication in Multiagent Collaboration” (Zheng et al., 23 Oct 2025)
• “Think-to-Talk or Talk-to-Think? When LLMs Come Up with an Answer in Multi-Step Arithmetic Reasoning” (Kudo et al., 2024)
• “Orality: A Semantic Canvas for Externalizing and Clarifying Thoughts with Speech” (Li et al., 3 Mar 2026)
• “Thinking LLMs: General Instruction Following with Thought Generation” (Wu et al., 2024)