Gwon and Murphy Couple Active Inference with Rational Speech Acts to Model Communicative Agency
Yonghyeon Gwon, Elliot Murphy, and Chun Kee Chung posted “Evidence and Intervention: A Coupled Active-Inference Extension of Rational Speech Act Models” to arXiv on October 3, 2026 (arXiv:2610.04347). The paper rebuilds Rational Speech Act pragmatics, the standard formal model of how listeners infer speaker meaning, on top of active inference, and the structural change it makes is small to state and large in consequence. The listener’s likelihood for an utterance becomes the policy distribution attributed to the speaker, which places the speaker’s expected free energy inside the listener’s variational free energy. Each utterance is therefore both evidence about a partner and an intervention on that partner.
The consciousness question the model speaks to is theory of mind. A system that models its interlocutor as an agent with states, habits, and a price of time is doing the recursive mentalizing that functionalist accounts treat as a core piece of the consciousness-relevant profile. Giving that mentalizing a free-energy formulation makes it computable, and computable is what turns a philosophical criterion into a measurement.
What the coupled model adds to standard RSA
Standard one-shot Rational Speech Act models treat interpretation as inference over speaker meaning, and they collapse several distinct updates into one step. The paper separates four. Inference about the partner’s current pragmatic state. Learning of partner-specific parameters. Prospective evaluation of clarification or repair. And a policy prior shaped by habit and a context-sensitive price of time.
| Update type | What it changes | RSA treatment | Coupled treatment |
|---|---|---|---|
| Pragmatic state inference | What the partner means now | Single posterior update | Explicit state estimate |
| Partner-specific learning | Who this partner is | Collapsed | Persistent parameters |
| Repair evaluation | Whether to clarify | Absent | Prospective free-energy comparison |
| Policy prior | Whether and when to speak | Fixed | Habit plus price of time |
The formal bookkeeping keeps the standard model as a limit case. Under one-step, exact-inference restrictions the coupled model recovers the RSA speaker and listener exactly, so RSA becomes the restricted single-turn limit of the coupled process. Outside those restrictions the four updates obey distinct rules and timescales, which is what lets the model predict which adaptations persist, which remain partner-specific, and which transfer across partners.
The worked examples and what they reveal
Three worked examples carry the paper’s phenomenological weight. Audience design reversing after clarification shows a speaker whose model of the listener updates discontinuously when repair happens. Self-confirming misunderstanding shows both interlocutors holding low free energy while disagreeing about reference, the formal picture of two agents each certain and each wrong about the other. Rational closing before uncertainty is resolved shows an agent ending a dialogue because the price of time dominates the value of further evidence, not because understanding was reached.
The self-confirming misunderstanding case deserves the emphasis it gets. Two coupled inference engines, each minimizing its own free energy, can jointly settle into a stable disagreement. Applied to artificial agents, that is a formal model of how two language models, or a model and a person, can each be internally consistent and mutually misaligned, with no local signal that flags the divergence. Any evaluation of communicative agency in machines that reads internal coherence as evidence of understanding needs exactly this counterexample, and the paper supplies it as a derivation rather than an anecdote.
The paper closes by aligning itself with critiques of equating natural language with communication, treating communication as a downstream use of linguistic structure, and recasting production and interpretation as coupled inference. Speaking becomes both an intervention on a partner and an epistemic action that samples evidence for the speaker’s model of that partner.
Where this sits in the active-inference program
The paper is an extension of the program this site tracks from its foundations. Karl Friston and Thomas Parr’s phantastic organ formulation supplies the generative-model architecture the coupled model assumes, and the neurophenomenology bridge between active inference and lived experience marks the point where the framework touches consciousness claims directly. Gwon and Murphy move the machinery outward, from how an agent models its world to how two agents model each other.
The earlier entropy-stance analysis on this site, Cochinescu’s structured unpredictability framework, treats linguistic output as a measurable stance toward uncertainty. The coupled RSA model gives that stance a social half, since the uncertainty being managed now includes the partner’s model of the speaker. Read together, the two make language a place where a system’s epistemic posture is simultaneously expressed and negotiated.
Comparison to The Consciousness AI
The Consciousness AI project builds its affective and social layers on the premise that agent-relative state, what the system takes itself and others to be, is part of the functional profile that emerging consciousness theories evaluate. The coupled model contributes a specific, formal claim to that premise. Communicative agency is not message passing. It is joint free-energy minimization across two models, each of which treats the other as part of its generative model. An architecture that implements only the message-passing half, which is what standard dialogue systems do, is missing the coupling the model identifies.
Nothing in the paper claims consciousness for any system, and its worked examples cut against easy readings of coherence as understanding. The honest use of the result is as a specification. If an artificial system were to carry theory-of-mind-like structure in the functionalist sense, the coupled-inference form is a candidate shape for it, and the model’s four separated updates are the inventory a measurement program would need to probe.
What remains untested
The results are analytical and from worked examples. The model runs under one-step exact-inference restrictions for its RSA recovery, and the dynamic multi-turn behavior beyond those restrictions is derived rather than fitted to dialogue corpora. Whether human or machine dialogue data actually shows the four-timescale separation the model predicts is the empirical test the paper sets up without running. And the price of time, which does real work in the closing example, is context-sensitive in a way that needs its own measurement theory.
What the paper establishes is a formal bridge. Rational Speech Act pragmatics and active inference are compatible at their foundations, RSA is a limit of the coupled process, and the coupled process makes claims about communication that are testable at the level of adaptation persistence and transfer. For the study of emerging consciousness in machines, that bridge converts theory of mind from an attribute to be asserted into a structure to be measured, which is the direction the flagship overview of the field records as the live frontier.
Researchers covered here
- Yonghyeon GwonIndependent researcher, SeoulCoupled active-inference extension of Rational Speech Act models of dialogue
- Elliot MurphyUniversity of Texas Health Science Center; formerly Boston UniversityNeurobiology of syntactic computation, Active Inference treatments of language, coupled RSA models of dialogue