NOTE-1062: Whyte et al. — On the minimal theory of consciousness implicit in active inference

Contribution

The paper does not start from a property of consciousness. It starts from active inference, a general framework for modelling behaviour as approximate Bayesian inference, and asks what theory of consciousness is already assumed by the active inference models of consciousness paradigms that exist. Its answer is a “hard core” of three commitments with a “protective belt” of model-specific assumptions, set out in Lakatos’s terms. It is written as a tutorial for experimentalists and a guide for modellers: “the primary aim of this paper is didactic rather than polemic” (p. 4). It adds three small simulations of its own.

Key insight

Because every active inference model minimises the same two objectives, whose terms (complexity, accuracy, risk, ambiguity, novelty) have fixed interpretations, a finding that breaks a shared assumption would break most models at once. That is what lets a modelling framework carry a theory of consciousness. On that core, a belief is conscious when it can steer a choice among discrete alternatives, slowly and precisely enough. Seeing is then a kind of looking: “To See is to Look” (p. 30).

Assumptions

• The organism as a generative model. Its phenotype is “a generative model of its niche”, and behaviour is “self-evidencing” (p. 5).
• The two objectives. Variational free energy for perception and reflex, expected free energy for any counterfactual choice, which “(normally) requires a discrete generative model” (p. 8).
• Hierarchical mixed models. Continuous predictive coding at low levels, discrete POMDPs above, joined by link functions (p. 8). Gaussian, Laplace and stationary-variance assumptions give predictive coding, which are “three different ways of phrasing the same assumption” (p. 8).
• A neuronal process theory. Expectation nodes in deep cortical layers, error nodes in superficial ones, softmax as firing rate, and ERPs as the rate of change of firing (pp. 9–11). The mapping to microcircuits “is a matter of ongoing research” (p. 9).
• Neuromodulators set precision. Noradrenaline for transition (B) precision, acetylcholine for likelihood (A) precision (pp. 23–24).
• Scope. Only numerical simulation studies are reviewed (p. 13), and only “well-studied paradigms” (p. 4). The approach is “about the properties of a system that exhibits phenomena associated with consciousness” (p. 4), not about where consciousness comes from.

Key results

By section. Results attributed to cited studies are the paper’s report.

• Section 2: the objectives (pp. 5–8). Free energy decomposes into divergence plus surprise, and into complexity minus accuracy. Expected free energy decomposes into risk, ambiguity and novelty, so goal-seeking and information-seeking share one objective.
• Section 3: belief updating (pp. 8–13). A single-layer predictive coding network is derived as gradient descent on free energy, and the POMDP update is given with its neural reading. Policy posterior is a softmax of summed expected free energy (eq. 14).
• Section 4.1: content (pp. 13–19).
  • Troxler fading and binocular rivalry (Parr et al. 2019) follow from imprecise transitions plus attention- or fixation-dependent likelihood precision. The unattended stimulus gains epistemic value until attention switches. The step to awareness is explicit: “Assuming a correspondence between the posterior over states and the content of awareness” (p. 14).
  • Appendix 1 re-implements the rivalry model and fits Levelt’s second and fourth propositions (A precision 0.7/0.3 when attended, uniform when not; B precision ω = 0.8). Appendix 2 adds reward (c = 0.25) and predicts that reward biases dominance and that lowering the rewarded stimulus’s precision compensates: “a predicted violation of Levelt’s second proposition under reward” (p. 33).
  • Hemineglect as a lesion of novelty-driven saccades (Parr & Friston 2018b), with a prediction about TPJ–FEF connectivity reported as confirmed by Parr, Mirza et al. 2019 with DCM and MEG.
  • Conscious access (Whyte & Smith 2021): a two-level POMDP reproduces Dehaene’s taxonomy and a P3b-like “ignition”, and predicted that valid expectations lower P3b, reported as “independently confirmed by Schlossmacher et al, (2020)” (p. 17).
  • No-report (Whyte et al. 2022): when report is not required, second-level firing and the P3b-like response drop, yet “the posterior at the second level of the model was still well above the threshold for close to 100% reportability of the stimulus” (p. 18). Of its first prediction the text says it “is exactly what was observed in a simultaneous EEG-fMRI experiment” (p. 18), with the citation missing from the manuscript. The second is reported as confirmed by Rowe et al. 2024.
• Section 4.2: self (pp. 19–22). Homeostatic and allostatic control (Tschantz et al. 2022), building on Seth 2013; fitted interoceptive precision in patients (Smith et al. 2020), replicated in a pre-registered sample; a gastric-vibration study; a cardio-visual model (Allen et al. 2022); emotion concepts learned in an “in silico childhood” (Smith et al. 2019); valence inferred from the change between prior and posterior over policies (Hesp et al. 2021); and meta-awareness as a third level that sets attentional precision (Sandved-Smith et al. 2021). The last is called “a kind of computational conceptual analysis” (p. 22), not a fit to data.
• Section 4.3: state (pp. 22–24). “Active inference models have so far not been applied to experiments that manipulate conscious state” (p. 22). The paper extends a local–global auditory model: lowering second-level B precision (ω = 0.9), standing in for low noradrenaline in sleep, removes the P3b-like response and spares the MMN-like one. The authors say it “does not in-itself explain why participants lose consciousness in parts of sleep” (p. 24). A cholinergic story for NREM and REM is offered in words, not simulated.
• Section 5: the minimal theory (pp. 24–31).
  • Hard core 1: the free-energy imperatives, so “all behaviour, both conscious and unconscious, must, in some minimal sense, be aimed at minimising these quantities” (p. 25).
  • Hard core 2: “The most minimal such link implicit in all the models reviewed above is between the contents of consciousness and the inferred hidden state” (p. 25). It “rules out that a change in consciousness could occur without a change in the posterior over states” and “is consistent with non-conscious creatures exhibiting behaviour governed by active inference” (p. 26).
  • Hard core 3: “the conscious/unconscious distinction is tied to the mechanism(s) of conscious access, and that it occurs at the discrete-continuous interface” (p. 26). A conscious posterior must be slow enough to abstract from sensory flux and precise enough to tell counterfactual states apart. The threshold varies with the task. The criterion is dispositional: “we are not claiming that being the target of a counterfactual policy selection process is what makes a posterior belief conscious. Rather it is the capacity (in a counterfactual sense) for such a posterior belief to inform policy selection processes at the right timescale and with sufficient precision that makes it conscious” (pp. 26–27). Its scope: “a necessary and sufficient condition for content to become conscious in humans (and non-human animals with similar neuronal architectures)”, while “we remain agnostic as to whether this condition is both necessary and sufficient in systems vastly different to ourselves” (p. 27).
  • Status of the parts: the first two are “in some sense logically entailed” by the framework; the third is “an inference to the best explanation” (p. 27). Most of the explanatory work “will be done by the protective belt of the theory” (p. 28).
  • The flexibility worry is stated: “active inference may be too flexible to be productive as a theory of consciousness” (p. 29). Two answers: shared assumptions (attention as A precision, used by “almost all” models of attention) would fall together; and there are close-to-core predictions.
  • Prediction 1: the risk–ambiguity trade-off in attention policies, as in Appendix 2. If reward did not trade off against epistemic value, “active inference would be seriously challenged” (p. 29).
  • Prediction 2: no sensory precision, no change in content. A stimulus given high precision by an attention or saccade policy “will enter the contents of consciousness sooner than a retinally-matched stimulus that is not the target of a saccade” (p. 30), which the authors call “unique to active inference as a theory of consciousness” (p. 30). It is part of an adversarial collaboration (INTREPID).
  • Failure modes named: overfitted models, or “if models of conscious and non-conscious perception look indistinguishable” (p. 31).
• Code (p. 32). The new simulations use SPM’s spm_MDP_VB_X.m. “Complete code necessary to reproduce the simulations reported in this paper will be made available upon full publication.”

Claims

id	claim	strength	support
C1	Active inference models of rivalry, Troxler fading, neglect, access, no-report, interoception, emotion, valence and meta-awareness share two objectives and a small set of interpretable terms	strong as a description of the reviewed models	the review, pp. 13–22
C2	Conscious contents correspond to posterior beliefs over hidden states; no change in content without a change in posterior	an assumption made explicit, not tested	”implicit in all the models”, p. 25
C3	A posterior is conscious when it has the capacity to inform counterfactual, discrete policy selection at a slow enough timescale with enough precision; necessary and sufficient in humans	weak: inference to the best explanation from what the models needed	pp. 26–27
C4	The no-report pattern (prefrontal and P3b loss with access intact) follows from report as a precision-raising mental action	moderate as a simulation result; empirical support at second hand, one citation missing	Whyte et al. 2022; pp. 18–19
C5	The rivalry model fits Levelt’s second and fourth propositions	demonstrated in simulation, figures not viewed	Appendix 1
C6	Reward should bias rivalry and trade off against stimulus precision	a prediction, untested	Appendix 2; p. 29
C7	Lowering second-level transition precision removes the P3b-like response and spares the MMN-like one	demonstrated in simulation; the link to sleep is by analogy	section 4.3, Figure 4
C8	An attended or saccade-targeted stimulus reaches awareness sooner than a retinally matched one	a prediction, untested here; claimed unique	p. 30

Concepts

• Variational and expected free energy: the two objectives, the second decomposed into risk, ambiguity and novelty.
• Precision: inverse variance; A precision is likelihood precision, B precision is transition precision. Attention is set by A precision.
• Hard core and protective belt: Lakatos’s division between the commitments a programme keeps and the assumptions it revises.
• Discrete–continuous interface: where continuous sensory beliefs are mapped to discrete states that can enter policy selection, the paper’s locus of conscious access.
• Dispositional criterion: what makes a belief conscious is its capacity to inform policy selection, not its use (after Carruthers and Prinz).
• “To See is to Look”: conscious perception uses precision assignments of a counterfactual policy selection process, overt or covert (p. 30).

Connections

It builds on Hohwy’s self-evidencing and his account of access as settling on a posterior to drive action (Hohwy 2013; Marchi & Hohwy 2022), on Parr and Friston’s attention as precision, and on Seth 2014 for the depth of counterfactual policy selection. It takes its programme framing from Negro 2024 and Lakatos 1968. The record holds several works it reviews or relies on: Seth 2013 (LIT-1263, NOTE-1000), the source of the interoceptive models; Sandved-Smith et al. 2021 (LIT-1191, NOTE-958), reviewed in section 4.2; Parr & Friston on attention (LIT-1195, NOTE-956); and Friston et al. 2010 on action (LIT-1181). Laukkonen, Friston & Chandaria (LIT-1363, NOTE-1063) cite an earlier version of this paper as a “complimentary view”, and Safron’s IWMT (LIT-1345) is named among the conceptual models it leaves out.

Bearing on the record

• QUESTION-254 and CLAIM-411. The paper offers a criterion for which posteriors are conscious, and it is the access criterion. It grants the counter-claim’s premise outright: hard-core part 1 covers “both conscious and unconscious” behaviour, and part 2 “is consistent with non-conscious creatures exhibiting behaviour governed by active inference”. Its separating condition (capacity to inform counterfactual policy selection, at the right timescale and precision) is stated without phenomenal terms, which CLAIM-411’s defeater requires. But it is not independent of report: it was read off the models’ report policies, and the authors scope it to humans and similar animals. So it does not meet that defeater, and it is a candidate for the organizational half of QUESTION-254, not for the phenomenal half. The authors do not claim more.
• CLAIM-412 (constitution). The paper speaks of a “correspondence” and a “link” between contents and posteriors (pp. 14, 25), not of identity. It is neutral between constitution and production, and its method (map model variables to phenomenology) is open to either. What it adds is a worked example of A7’s demand that the qualifying organization be specified: discreteness, timescale and precision are specified, and they are specified for access.
• TERM-073 (interpretation). Posterior inference here has all three of the term’s parts: a distinction (discrete states), a context (priors and precision) and a consequence (policy selection). It is a formal case of interpretation defined without experience, and of the cost the term admits.
• CLAIM-414 (four facts come apart). The dispositional criterion puts consciousness in the capacity of a belief to be effective, not in its use. That claim shows that ordinary behaviour cannot reveal whether an interpretation is effective, only an intervention can. A capacity “in a counterfactual sense” is that kind of fact. So the criterion, applied to a system that is not reporting, needs an intervention to check, which the paper’s no-report simulation supplies by sweeping stimulus precision.
• QUESTION-048 and THEORY-094 (report and no-report). The no-report model gives one mechanism for THEORY-094’s point: prefrontal activity and P3b track the precision raised by report as mental action, while the posterior that would support report stays above threshold. It does not separate experience from access. It reads access off a posterior and calls that conscious, and so meets THEORY-577’s demand that a theory say how its mechanism produces reports by building the report in. The paper is silent on illusionism (THEORY-633).
• THEORY-767 and THEORY-056. The interoceptive models start from LIT-1263. Tschantz et al. put changing beliefs about set points and acting on the body under one objective, a modelling counterpart to THEORY-056’s single-process view of regulation, not evidence for it. Nothing reviewed bears on body ownership.
• CLAIM-413 (grounding). The paper says nothing about grounding. Its criterion lets a posterior driven by priors alone be conscious (silent counting, p. 30), but the precision prediction ties change in content to sensory precision.
• The exchange’s description. The exchange cited the paper “for extracting a minimal, testable theory of consciousness from what active inference models have in common”. That is accurate. What it left out is the theory’s scope: it is a theory of conscious access in humans, which says nothing about systems unlike us and does not claim to say what makes an organization experiential.
• No instruction for machine-learning practice.

Limitations

• A review with three small new simulations, whose code was not yet released at v2. The empirical confirmations are cited, not shown, and one citation (the EEG-fMRI study, p. 18) is missing from the manuscript.
• The third hard-core commitment is read off the report policies of the reviewed models, so it is tied to report by construction. The authors grant it is an inference to the best explanation.
• Every model of content assumes a correspondence between posterior and awareness. None tests it.
• The sleep simulation reproduces one ERP pattern and, by the authors’ account, does not explain loss of consciousness.
• The authors concede that many models fit the same data and that other formalisms can be formally equivalent in specific domains (p. 16). The answer to the flexibility worry rests on two predictions, neither tested here.
• Read from the submitted version; the journal version was not reached.

Open questions

• Can the dispositional criterion be stated for a system that never reports and was not built with report policies, so that it reaches the systems the authors set aside?
• Does the “To See is to Look” prediction survive matching expectation between attended and unattended stimuli, the confound LIT-1276 found in the rubber hand illusion?
• Is the discrete–continuous interface one locus, or a gradient along the hierarchy? The criterion’s task-dependent threshold suggests the second.