iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Yes—but only in specific experimental designs. Some AI agents can exchange numerical hidden-state representations or coordinate by observing one another’s actions instead of sending natural-language messages. This is usually called latent communication or latent-space inter-agent communication. It does not mean that every model can connect to every other model, that machines possess human-like telepathy, or that the methods are ready for unrestricted deployment.
What “without words” means in AI
Natural-language text is only one possible communication channel. A model can also pass internal numerical representations produced during inference, write those representations into a shared latent memory, or reveal information through behavior that another agent learns to interpret.
| Approach | What is transmitted | How the partner receives it | What the cited work demonstrates |
|---|---|---|---|
| Direct latent-state exchange | Continuous hidden states, often compressed before transmission | One model sends representations directly to another model | Interlat presents a feasibility study of entirely latent-space communication |
| Shared latent workspace | Latent thoughts derived from last-layer embeddings | Agents read and write a common latent working memory | LatentMAS reports training-free latent collaboration across nine benchmarks |
| Behavioral signaling | Actions that reveal information about a private state | A partner infers the signal from observed behavior | Policy Belief Learning studies cooperative games in which explicit messaging is disabled |
| Human-facing nonverbal interaction | Speech-linked motion, gesture, or other audiovisual cues | A person or system interprets visible or audible behavior | DyaPlex and Meta’s audiovisual work address social interaction, not necessarily hidden-vector exchange between LLMs |
The last category is easy to confuse with latent inter-agent communication. A robot that turns toward a person or produces a gesture is communicating without written words, but that does not prove that two language models are passing private hidden states to each other.
How latent-state communication works
In ordinary multi-agent prompting, one model writes a message, another model reads the text, and the conversation continues. Latent systems replace, supplement, or bypass that text interface.
#1 Best Overall
- Generate an internal representation. While processing an input, the sending model produces vectors in its hidden layers. These vectors encode information used by the model’s next computations, but they are not normally readable sentences.
- Choose a transfer point. A system may take a last-layer embedding, a sequence of continuous hidden states, or a learned latent “thought” produced for collaboration.
- Compress or align the representation. Hidden states can be large. A communication method may learn a compression function or another alignment mechanism so the receiving model can use fewer numbers and interpret them consistently.
- Deliver the representation. The vector sequence can be sent directly to another agent or stored in a shared latent working memory.
- Condition the receiving agent. The second model incorporates the received representation into its own inference and produces a decision, another latent contribution, or a user-visible answer.
The key difference from text messaging is that the channel carries numerical activations rather than a vocabulary-encoded string. The recipient does not necessarily decode a sentence first; it uses the representation as part of its computation.
Interlat: direct communication in latent space
Interlat is described by its authors as “a feasibility study of entirely latent space inter-agent communication.” The method transmits continuous last hidden states between agents and applies learned compression. Its goal is to reduce the cost of converting internal reasoning into text and then parsing that text again.
The ACL 2026 paper reports inference acceleration of up to 24× in its evaluated setup. That figure belongs to the paper’s models, compression method, and workloads; it is not a guaranteed speedup for arbitrary model pairs. Calling the work a feasibility study is important: the result shows that the channel can work under defined conditions, not that latent exchange is a solved interoperability problem.
LatentMAS: a shared latent working memory
LatentMAS takes a different organizational approach. Rather than requiring every exchange to become a natural-language message, it generates latent thoughts from last-layer embeddings and transfers them through shared latent working memory. The authors write: “We introduce LatentMAS, an end-to-end training-free framework that enables pure latent collaboration among LLM agents.”
According to the Proceedings of Machine Learning Research paper, LatentMAS was evaluated on nine benchmarks and reported:
Rank #2
- brand: Pearson
- ARTIFICIAL INTELLIGENCE: A MODERN APPROACH, 4TH EDITION
- Up to 14.6% higher accuracy than the compared communication setup.
- 70.8%–83.7% fewer output tokens.
- 4×–4.3× faster end-to-end inference.
These are reported benchmark results, not universal guarantees. Token reduction can be valuable when agents otherwise spend large amounts of time serializing and rereading intermediate reasoning. However, the quality of a latent workspace depends on whether the participating models and the learned interfaces use compatible representations.
Why model compatibility matters
Text is an unusually portable interface: two different systems can exchange a string if they agree on an encoding such as UTF-8 and on the meaning of the words. Hidden states do not have that same built-in portability. A vector at one model’s layer 24 may have no directly usable meaning at another model’s layer 24.
A practical latent communication system therefore has to specify several compatibility conditions:
- Backbone and layer layout: the sending and receiving models may need matching or deliberately mapped hidden-state dimensions.
- Representation alignment: a projection, adapter, shared training procedure, or learned compressor may be required.
- Tokenization and sequence handling: models can organize information over different token boundaries and sequence lengths.
- Training and inference assumptions: a method advertised as training-free for a particular framework may still depend on pre-existing model compatibility or a prepared interface.
- Failure detection: the recipient needs a way to recognize malformed, out-of-distribution, or semantically mismatched representations.
The cited Interlat and LatentMAS records report experiments with particular model configurations and alignment choices. They do not establish that arbitrary commercial or open models can exchange hidden states reliably without adaptation.
Communication through actions instead of hidden vectors
A separate line of work lets an agent communicate by choosing actions that another agent can observe. The action may have an ordinary task purpose while also signaling information about a private state, intention, or belief. The partner learns to infer the hidden information from the behavior.
The AAAI paper Learning to Communicate Implicitly by Actions describes “collaborative games where explicit communication is disabled.” In this setting, an agent cannot send a message such as “the target is on the left.” It can, however, move, select, wait, or manipulate an object in a way that a trained partner learns to interpret.
Free tools Windows power users keep installed
One-click scans. No signup required.
This is implicit signaling, not direct transfer of an LLM’s hidden state. The channel is the observable action sequence. It can be useful when a system has no communication API, when messages are disallowed by the environment, or when behavior itself must coordinate a team. It also introduces ambiguity: an action chosen for task reasons may be mistaken for a signal, and a signal may be hard to distinguish from noise.
Related work broadens the idea to embodied cooperation. The Latent Theory of Mind paper studies latent communication in cooperative robot manipulation, where one robot’s internal estimate of another agent’s state can support joint physical work. The mechanism is still distinct from simply sending readable prose between two chatbots.
Nonverbal interaction with people is a different problem
AI systems can communicate with humans through motion, gesture, gaze, prosody, and other audiovisual cues. NVIDIA’s DyaPlex project describes continuous, full-duplex speech-and-motion interaction. Meta’s Seamless Interaction work likewise focuses on dyadic audiovisual motion modeling and datasets.
These projects matter because communication does not have to be typed. They do not, by themselves, show that two LLM agents exchange private numerical activations. Human-facing nonverbal interaction asks whether a system can perceive or generate social cues; latent inter-agent communication asks whether model computations can be connected through internal representations.
What the reported numbers actually tell you
| System or paper | Reported result | How to interpret it | Source |
|---|---|---|---|
| Interlat | Up to 24× inference acceleration | Result of the paper’s additional compression and evaluated setup; authors characterize the work as a feasibility study | ACL Anthology |
| LatentMAS | Up to 14.6% higher accuracy | Reported across the authors’ benchmark comparisons, not a universal improvement | PMLR |
| LatentMAS | 70.8%–83.7% fewer output tokens | Measures output-token use in the evaluated latent-collaboration setup | PMLR |
| LatentMAS | 4×–4.3× faster end-to-end inference | End-to-end speed reported for the paper’s tasks and implementation | PMLR |
Accuracy, token count, and latency are related but not interchangeable. A system can emit fewer tokens yet spend substantial compute producing latent vectors. Conversely, a speedup measured on one sequence length or hardware configuration may shrink when network transfer, compression, or validation is added. Any deployment decision should reproduce the relevant workload rather than copy a headline multiplier.
Potential advantages of latent channels
- Lower serialization overhead: agents do not have to turn every intermediate result into a long textual explanation.
- Higher bandwidth for internal detail: a continuous representation can carry information that would be awkward or verbose to express in words.
- Reduced visible token traffic: LatentMAS’s reported token reductions show why teams are interested in this design.
- Faster coordination in a closed system: the reported Interlat and LatentMAS speed results suggest that avoiding repeated text generation and parsing can reduce latency in suitable setups.
- Coordination when explicit messages are unavailable: action-based signaling can let agents cooperate under an environment’s communication restrictions.
These benefits are conditional. They depend on aligned models, reliable interfaces, suitable tasks, and evaluation that includes the cost of producing, transmitting, decoding, and checking the latent data.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Trade-offs and failure modes
Opaque messages are harder to inspect
A text transcript lets an operator search for a false claim, a leaked secret, or an instruction that violates policy. A hidden vector is not naturally readable. Debugging may require probes, decoders, activation analysis, or a parallel textual trace, which can remove part of the efficiency benefit.
Semantic mismatch can look like success
A receiving model may accept a vector with the correct dimensions while interpreting it incorrectly. It can then produce a confident but wrong answer. Dimension matching is not semantic compatibility.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Compression can discard crucial information
Interlat’s use of learned compression illustrates the central trade-off: fewer values can mean lower bandwidth and faster inference, but an aggressive compressor may remove details needed for a rare case or safety check.
Best Value
Security assumptions change
Latent communication is not inherently safer than text. A malicious or compromised agent could encode unwanted instructions in a representation, and a monitoring system may have less visibility into that payload. Conversely, a hidden channel could make some classes of prompt injection harder to detect. Security must therefore be tested for the specific interface rather than inferred from the absence of words.
Behavioral signals can be ambiguous
In action-based communication, the same movement may be both a useful task action and an attempted signal. Policies trained in one game or environment may fail when the layout, partner, or incentives change.
How to evaluate a claim that models “talk without words”
- Identify the channel. Ask whether the system transfers hidden states, embeddings, shared-memory entries, actions, gestures, or ordinary audio.
- Check the participants. Record the exact model family, architecture, layers, adapters, and whether the agents are homogeneous or heterogeneous.
- Separate training from inference. “Training-free” can describe a framework’s use after the models and interfaces are prepared; it does not mean no alignment or engineering was needed anywhere.
- Inspect the baseline. Compare against a clearly defined text-message or single-agent baseline using the same tasks, hardware, and stopping criteria.
- Measure complete cost. Include latent generation, compression, transfer, memory access, decoding, validation, and retries—not only output-token count.
- Test generalization. Vary partners, prompts, task distributions, sequence lengths, and failure conditions to see whether the channel survives beyond the reported benchmarks.
- Keep an audit path. For high-stakes uses, retain enough observable information to investigate errors, even if production coordination uses latent messages.
What this does—and does not—prove
The evidence supports a precise conclusion: selected AI systems can exchange latent representations or communicate indirectly through actions under defined experimental conditions. The work does not show that all AI models can communicate without words, that models understand one another perfectly, or that latent channels demonstrate consciousness or human-like private thoughts.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteThe phrase “latent thought” is useful shorthand for a learned internal representation in these papers. It should not be read as evidence that a model has a human mind. A vector can support coordination without being a sentence, an experience, or an independently meaningful thought.
For now, the most accurate description is engineering-specific: researchers are building alternative interfaces between agents. Some interfaces pass hidden states, some provide a shared latent workspace, and some infer information from behavior. Each must be validated for its own models, tasks, alignment method, and safety requirements.
Quick Recap
Further reading
- Enabling Agents to Communicate Entirely in Latent Space (Interlat)
- Latent Collaboration in Multi-Agent Systems (LatentMAS)
- Learning to Communicate Implicitly by Actions
- DyaPlex: continuous speech-motion interaction
- Latent Theory of Mind
- Seamless Interaction
- Thought Communication in Multiagent Collaboration
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.
