Contribution
It gives active inference an explicit model of mental action and meta-awareness. Earlier “parametrically deep” models (Hesp et al. 2021) let higher-level states condition lower-level precisions. This paper adds policy selection at the attentional level, so that the agent can switch its own attentional state, and a third, meta-awareness level whose states set the precision with which the agent observes its attentional state. Its claimed novelty is control: “What is novel about our model is the coupling of this deep inferential architecture with ‘action and its control’” (Discussion, “Expanding the model to other parameters”).
Key insight
A mental action has the same form as a bodily one, one level up. The agent’s attentional state is treated as a hidden state of a generative process that the agent cannot see directly, that drifts on its own, and that the agent’s policy can push. The agent infers which attentional state it is in from observations of it and from how its perceptual beliefs are updating, and acts to bring about the outcome it prefers, “focused”. What the attentional state does at the level below is set a precision: “There is nothing special about these states, other than they generate outcomes at lower levels that speak—not to which state the world is in but—to the precision of beliefs about which state the world is in” (Methods, “Formalizing meta-awareness …”).
Assumptions
- Discrete-state POMDPs with a likelihood A, transitions B, preferences C, initial states D, policy prior E and expected free energy G (Methods).
- Attention is likelihood precision. “In active inference, ‘attentional processes’ have been formulated in terms of the precision (here denoted γA) of—or confidence in—the likelihood mapping (the A matrix; Parr and Friston 2017b)” (Methods, “Generative models”). The precision-weighted likelihood is A with each element raised to γA and renormalized (Fig. 2 caption).
- Purely parametric levels. “higher-level state factors are only those that condition precisions of the level below” (Introduction).
- Two levels per factor. Perceptual: standard or deviant. Attentional: focused or distracted. Meta-awareness: high or low (Fig. 6 caption). Focused and distracted set γ to 2.0 and 0.5 (β = 0.5 and 2.0; Methods, after Eq. 1).
- Action only at level 2. “Action here is only defined on the second level, with two possible actions u(2), ‘stay’ and ‘switch’, which condition the attentional transition matrix B(2) as shown” (Fig. 6 caption).
- A fixed world. “The generative process of this state is predetermined, with a deviant stimulus shown every 20th time step” (Results, Level 1). “Sensory habituation is not modelled here.”
- Meta-awareness set by hand. In the Level 3 simulation it is high for the first half of the trial and low for the second, by design; “we could have let the generative model roam its meta-awareness states freely” (Results, Level 3).
- No learning. No likelihood, transition or precision hyperparameter is learned anywhere. The model changes its states, not its parameters.
Key results
All results are single illustrative simulations. No statistics are reported and no data are fitted.
- Level 1 (Fig. 7). A passive oddball with no action. Focused (high γA on A(1)) for the first half, distracted for the second. When focused, the agent updates its belief on the deviant immediately. When distracted, its beliefs are “not fully adjusted when the deviant is presented”. Precision acts as “a ‘rate constant’ in evidence accumulation and consequent belief updating” (Results, Level 1).
- Level 2 (Figs. 8–10). A preference in C(2) to observe “focused”. The generative process moves the true attentional state with a policy-dependent probability. The agent becomes distracted, keeps believing it is focused for a few steps (mind-wandering), infers “distracted” from higher-order observations and from ascending evidence about first-level precision, and selects “switch”. The deviant speeds this up, because the prediction error it causes strengthens ascending evidence (Results, Level 2).
- Level 3 (Fig. 11). With low meta-awareness (low precision on A(2)), mind-wandering lasts longer. “In this example, the agent does not realize their mind has wandered until the contents of their perception changes (i.e. the oddball is presented)” (Results, Level 3).
- Equation 1. The posterior over attentional states combines a prior from B(2), ascending evidence from the first level’s precision update (computed by Bayesian model reduction), and “perceptual evidence” γA(2)·ln A(2)·o(2) from a direct observation of the attentional state. This last term is available only when γA(2) > 0. That is the paper’s formal line between a transparent state (inferred only from below) and an opaque one (also observed).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Mental action can be formalized as policy selection over higher-level states that condition lower-level likelihood precision | moderate | Methods, Figs. 4–6; a construction, shown to run |
| C2 | The focus, mind-wandering, noticing and refocusing cycle emerges from this architecture | weak | one simulation (Fig. 10), qualitative match to Lutz et al. and Hasenkamp et al. |
| C3 | Lower meta-awareness lengthens mind-wandering | weak | one simulation (Fig. 11), with meta-awareness set by hand |
| C4 | Opacity versus transparency maps onto zero versus non-zero higher-level likelihood precision | moderate | Eq. 1 and the paragraph after it; definitional |
| C5 | The scheme generalizes to precision on B, C, D, E and G | weak | asserted (“Mathematically, it is perfectly valid …”); not implemented, by the authors’ statement |
Method
- Level 1 POMDP over the stimulus, with likelihood A(1) raised to the precision γA(1).
- Level 2 states (focused, distracted) fix γA(1) through β(f,d). Their own likelihood A(2) maps them to an observation o(2) of the attentional state. B(2) depends on the policy u(2) ∈ {stay, switch}. C(2) prefers “focused”.
- Level 3 states (high, low) fix γA(2) the same way.
- Each level’s state posterior combines prior, ascending precision evidence and its own observation (Eq. 1). Policies are scored by expected free energy.
Concepts
- Mental action. “Goal-directed internal behaviour, without necessarily having a direct consequence on the external environment (e.g. as through muscular action). A paradigmatic example of mental action is covert attention” (Introduction). Formally: “the deployment of precision” (Methods).
- Meta-awareness. Third-order states that make second-order attentional states opaque, as attentional states make perceptual states opaque (Introduction, the bracketed note).
- Opacity and transparency. After Metzinger. A process is transparent when only its content is available and opaque when the process itself can be accessed “as data for further inference”.
- Parametric depth. Higher levels take lower-level posterior and precision estimates as data (Introduction, after Hesp et al. 2021).
Connections
It extends Hesp et al. (2021) from model precision (γG) to likelihood precision (γA), and takes from Limanowski and Friston (2018, 2020) the proposal that mental action is the deployment of precision. Its attention-as-precision rests on Parr and Friston (2017; see LIT-1195 for their 2019 review, read in NOTE-956). It differs from Whyte and Smith (2020) by having purely parametric levels and a third level. It explicitly sets aside “the role of selective attention in epistemic foraging” (Introduction). It does not use the term “Markov blanket”.
CLAIM-337 (the counter). The claim quotes the abstract exactly, and its paraphrase is right in outline. Three things the reading changes.
- Where mental action sits. Only the attentional level has policies. Meta-awareness is not acted on: the third-level state is set by the generative process in the simulation. So “mental action selects among higher-level cognitive states” is true of one level, with two actions, “stay” and “switch”.
- It is literally action one level inward. The paper models the agent’s own attentional state as a generative process, “we have programmed a ‘generative process’ that regulates the ‘true’ state transitions that the agent undergoes” (Results, Level 2). The agent acts on that process to bring about a preferred observation o(2) of itself. This is the action–perception cycle with the attentional state in the place of the world, and the claim’s “one blanket inward” is a fair description of it. The paper’s own phrase is that “there is no fundamental difference between a simple reflex, a series of complex movements, and a mental action” (Discussion).
- No structure learning. Nothing in this model is learned. It cannot support the claim’s second strand, the observer reshaped without action. That strand rests on Smith et al. (THEORY-564), not on this paper.
CLAIM-341 (the thesis). This is exactly the line the thesis draws, and the paper falls on the side the thesis needs at the level of perception. At the first level the observation’s content is fixed before and independently of any mental action: the stimulus sequence is “predetermined”, with a deviant every 20th step. Mental action changes only the precision with which that content updates belief, a “rate constant” in evidence accumulation. The authors note that the same effect could come from changing the entries of A(1) “in response to incoming evidence”, but they model it as a single precision applied to all entries, not as a change to the mapping. So, at level 1, mental action does not produce what the agent observes, or change which mapping turns causes into observations. It changes how much the fixed observation counts. At level 2 the picture differs. Mental action changes the true attentional state, and so it does change the content of the observation o(2) the agent then receives of itself. That is the synchronic form the thesis declines to claim, and it takes the action–perception shape. For the diachronic form, in which a judgement reshapes the observer that produces later observations, the paper offers a weak analogue. The attentional state persists through B(2) and conditions later perception, so a mental act now changes how later stimuli are weighed. But it is a two-valued gain on an unchanging likelihood, reversible and unlearned. It changes what the observer believes, not which objects it individuates. On the claim’s own terms, the paper’s mental action is closer to action than to judgement. It targets a state of the agent and is chosen to make a preferred observation come true, whereas the claim says judgement targets the observer and is not chosen for its outcome.
CLAIM-342 and QUESTION-121. In this model what corrects a false belief about one’s own attention is evidence not routed through that belief. Two kinds do it: ascending precision evidence from the level below, and the direct observation o(2). Under low meta-awareness the correction waited for an exogenous change in the stimulus: the agent “does not realize their mind has wandered until the contents of their perception changes”. This is a small illustration of the question’s first candidate, unselected input, doing the anchoring. Note that it is input the agent could not select, since the stimulus schedule is fixed. But the loop is also built to confirm a preference: the agent acts until it observes itself “focused”. Nothing in the model checks whether “focused” is a correct reading of the world.
Bearing on the record
- CLAIM-337: supports its first strand, and more literally than the claim states it. Mental action here is action on a modelled internal generative process. It qualifies the strand to one acting level, with precision-setting only and no learning.
- CLAIM-341: at the perceptual level, the observation is fixed before the mental action and only its precision changes, as the claim needs. At the attentional level, the mental action produces the content of the self-observation, and the loop has the action–perception shape. The paper offers no mechanism by which judgement reshapes the observer’s likelihood.
- LIT-1191’s “Why this is here”: its question is answered. The observation weighted at level 1 is fixed independently of the mental action.
- No instruction for machine-learning practice.
Limitations
These are proofs of principle: single runs, hand-set parameters, and a meta-awareness schedule imposed from outside. There is no fit to behavioural, phenomenological or neural data, which the authors leave to future work. The generalization to other precisions is stated, not built. The equations are given for the attentional level only. The perceptual task is a passive oddball, so overt action is never combined with mental action in a simulation, although the Discussion says the framework allows it.
Open questions
- With the meta-awareness level left free and given its own policies, does the model still settle, or does it lock in a preferred self-observation? The authors chose not to run this for clarity.
- If A(1) were learned (Dirichlet counts) while attention fluctuated, would a habit of attention leave a lasting change in the likelihood? That would be the diachronic reshaping CLAIM-341 needs, and this paper does not test it.