Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsYes—but “without words” means exchanging numerical internal representations or signaling through actions, not reading one another’s minds. Recent research demonstrates two main agent-to-agent approaches: transmitting hidden states or latent embeddings directly, and inferring information from a partner’s behavior. These methods work only under defined model, alignment and benchmark conditions; they do not show that arbitrary AI systems can communicate perfectly or that AI has human-like telepathy.
What “talking” without words means for AI
Natural-language text is only one communication channel. A language model also produces numerical vectors at each layer as it processes an input. Those hidden states encode features of the model’s current computation—such as entities, relationships, plans or uncertainty—although they are not readable thoughts in the human sense.
As an Amazon Associate I earn from qualifying purchases.
In a latent-communication system, one agent passes some of those vectors, or a compressed version of them, to another agent. The receiving model uses the numerical signal as part of its own computation instead of waiting for a sentence. A system can also create a shared latent workspace in which agents write and read embeddings. A different research tradition uses actions as signals: an agent moves or acts in a way that reveals information about a private state to a partner.
Free tools Windows power users keep installed
One-click scans. No signup required.
These are technically different from a person using gestures with an AI. Human-facing systems may generate or interpret speech-linked body motion, facial movement or gesture, but that does not imply that two language models are exchanging hidden vectors.
#1 Best Overall
Three distinct forms of nonverbal AI communication
| Approach | Information passed | Connection between agents | What the evidence shows |
|---|---|---|---|
| Direct latent-state exchange | Continuous hidden states, often compressed | One model transmits representations to another | Interlat reports a feasibility study of entirely latent inter-agent communication and acceleration up to 24× in its experimental setup. ACL Anthology paper |
| Shared latent memory | Last-layer embeddings representing intermediate “thoughts” | Agents read and write a common latent working memory | LatentMAS is described as training-free and reports benchmark gains, lower token use and faster inference on nine benchmarks. PMLR paper |
| Behavioral signaling | Actions whose choice reveals a private state | A partner infers the message from observed behavior | Policy Belief Learning studies collaborative games where explicit communication is disabled. AAAI paper |
| Human-facing nonverbal interaction | Speech-linked motion, gesture or audiovisual cues | A model perceives or generates social behavior for a person | DyaPlex and Meta’s audiovisual work address speech-motion or dyadic interaction, not necessarily hidden-state exchange between LLM agents. DyaPlex research Seamless Interaction research |
How direct latent communication works
1. The first model computes an internal representation
An agent receives an observation, task or prompt and processes it through its network. Instead of converting its result into a text message, the system selects a hidden state—often from a late layer—or another latent embedding. This vector is a numerical object, not a sentence.
2. The representation is made transferable
Raw hidden states can be large and tied to a particular model. Interlat therefore describes learned compression for continuous last-hidden-state communication. Compression reduces the amount of data that must cross the agent boundary, but it also creates a learned interface whose quality depends on the models and training setup used in the experiment.
3. The receiving agent incorporates the signal
The second model receives the vector and combines it with its own computation. It may use the signal to continue reasoning, choose an answer or coordinate an action without reconstructing an English explanation. The receiving model must have a compatible way to interpret the representation; a vector from one architecture is not automatically meaningful to every other architecture.
4. The system is evaluated on a task
Researchers compare latent exchange with a text-based baseline using task quality, output length, latency or end-to-end inference time. A faster result can come from avoiding token generation, but the comparison is valid only for the reported models, compression method, hardware and benchmark conditions.
Rank #2
Interlat’s authors characterize their work as “a feasibility study of entirely latent space inter-agent communication.” That wording matters: it establishes that the channel can be built and tested, not that latent messaging is a solved, universal protocol. Read the full method and conditions in the ACL 2026 publication.
How a shared latent workspace differs
LatentMAS takes a collaboration-oriented approach. Its agents generate latent thoughts from last-layer embeddings and transfer them through shared latent working memory. Instead of repeatedly writing and reading natural-language messages, agents deposit numerical representations that other agents can access.
The paper calls the framework “an end-to-end training-free framework that enables pure latent collaboration among LLM agents.” Training-free does not mean setup-free or model-independent: the framework still relies on the representations and interfaces supported by the evaluated models. It means the reported collaboration method does not require an additional end-to-end training phase for the agents to use the latent workspace.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →On nine benchmarks, LatentMAS reports up to 14.6% higher accuracy, 70.8%–83.7% fewer output tokens and 4×–4.3× faster end-to-end inference. These are the paper’s maximum reported improvements under its experimental conditions, not guarantees for every task or model. The benchmark definitions, baselines and full results are in the Proceedings of Machine Learning Research paper.
Behavior can carry a message even when vectors are hidden
Not all nonverbal communication involves passing an LLM’s internal state. In Policy Belief Learning, an agent’s actions are selected so a partner can infer information about a private state. A route, movement, timing choice or other action becomes a signal because the partner has learned what that behavior usually means.
This is closer to signaling in a cooperative game than to a private data channel. The receiver observes an external action and updates its belief about the sender. The AAAI paper describes “collaborative games where explicit communication is disabled.” That setting shows that agents can coordinate through behavior when a message field is unavailable, but it does not show that they can exchange arbitrary hidden thoughts.
Related work on latent theory of mind studies latent communication in cooperative robot manipulation. Its focus is how one agent can represent or infer another agent’s internal state during joint physical tasks, rather than proving a general-purpose language-model protocol. See the PMLR paper for that research direction.
Why latent messages can be faster and shorter
- No token-by-token wording: A vector can represent intermediate computation without generating a sequence of vocabulary tokens.
- Less serialization: The sender does not need to turn every useful detail into grammar, and the receiver does not need to tokenize and parse a prose response.
- Compressed transfer: Methods such as Interlat learn to reduce the size of hidden-state messages before transmission.
- Shared intermediate results: A latent workspace can let several agents reuse an internal representation instead of restating it in separate messages.
Those advantages come with a trade-off. Natural language is inspectable and relatively portable; a latent vector is opaque and usually meaningful only within a compatible model-and-interface combination. Compression can lower bandwidth while discarding information that a different task would have needed.
Compatibility is the central engineering constraint
A successful experiment with two selected models does not establish interoperability among arbitrary models. Hidden states depend on architecture, layer placement, dimensionality, training distribution and alignment. Even models that accept the same text may organize their internal representations differently.
Before connecting agents, verify:
- Backbone and layer compatibility: Identify which model produced the representation and which layer or embedding type is transmitted.
- Interface alignment: Confirm that the receiver was designed, adapted or evaluated to consume that representation rather than ordinary token embeddings.
- Dimensionality and compression: Check vector size, compression method and any decoder or projection required between models.
- Task match: Test on the same task distribution used for the claimed result; a channel that works for one benchmark may fail on another.
- Fallback behavior: Decide what happens when a message is malformed, out of distribution or ignored by the receiver.
The available studies report particular experimental systems and alignment methods. They do not establish that a general latent protocol exists for heterogeneous commercial or open models.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the speed and accuracy numbers do—and do not—prove
LatentMAS reports its largest gains as “up to” values. The 14.6% accuracy improvement is a maximum over its nine-benchmark evaluation, while the 70.8%–83.7% reduction concerns output tokens and the 4×–4.3× figure concerns end-to-end inference. None of those numbers should be read as an average improvement for every deployment.
Interlat reports acceleration of up to 24× through additional compression and frames the work as a feasibility study. A speedup measured in one pipeline can shrink when projection, memory access, network transfer, safety checks or model loading are added. Compare like with like: the same task, hardware, number of agents, quality target and accounting for all communication overhead.
Best Value
“Without words” does not mean “without information loss”
Natural language forces an agent to make information explicit. That can remove detail, but it also creates an audit trail that a person can inspect. Latent communication preserves richer intermediate signals in numerical form, yet those signals are difficult to interpret and may contain details the system designer did not intend to share.
A latent channel also is not automatically safer. An opaque representation can make monitoring, red-team analysis and policy enforcement harder. It may carry sensitive information, encode a sender’s unintended biases or fail silently when the receiver’s representation drifts. Security and privacy claims require dedicated testing; the cited papers do not establish that latent communication is inherently secure.
How to evaluate a proposed latent-communication system
- Define the channel: State whether the system transmits hidden states, compressed embeddings, shared-memory entries or behavioral actions.
- Specify the participants: Record model versions, architectures, layers, projection modules and any alignment or calibration procedure.
- Choose a text baseline: Compare against a conventional natural-language message with an equivalent task budget.
- Measure quality and cost separately: Report task accuracy or success, output tokens, latency, bandwidth and end-to-end compute.
- Probe generalization: Test new tasks, changed prompts, altered partners and distribution shifts instead of relying on one benchmark.
- Inspect failures: Log malformed vectors, incompatible dimensions, receiver misinterpretations and cases where a text fallback would have prevented an error.
- Assess observability: Determine what humans can audit, what information can leak through the channel and how communication can be disabled or revoked.
Human-AI motion is a related but separate problem
NVIDIA’s DyaPlex project describes continuous speech and motion interaction, where a system handles speech and body movement as part of a full-duplex exchange. Meta’s audiovisual interaction research likewise models social cues between people. These projects matter for robots, avatars and assistive interfaces, but they should not be cited as proof that two LLMs exchanged private hidden states.
The distinction is simple: human-facing nonverbal interaction uses observable behavior such as gesture or motion, while latent inter-agent communication passes internal numerical representations or uses actions as deliberately designed signals.
Where the field stands
Current evidence supports a narrower statement than the headline might suggest. Selected research systems can coordinate through latent representations, shared latent memory or action-based signals without composing ordinary natural-language messages. LatentMAS demonstrates reported benchmark improvements in a training-free framework, and Interlat demonstrates a compressed latent channel with reported acceleration. Policy Belief Learning shows that agents can signal through actions when explicit communication is disabled.
What remains unproven is broad interoperability, dependable interpretation across unrelated architectures, universal quality gains and any human-like consciousness or telepathy. Treat latent communication as an experimental systems technique whose usefulness must be demonstrated for a specific pair of models, task and operating environment.
Quick Recap
Further reading
- Enabling Agents to Communicate Entirely in Latent Space (ACL Anthology, 2026)
- Latent Collaboration in Multi-Agent Systems (PMLR, 2026)
- Learning to Communicate Implicitly by Actions (AAAI)
- Latent Theory of Mind (PMLR)
- Thought Communication in Multiagent Collaboration (OpenReview PDF)
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
