Contribution
The paper recasts post-traumatic stress disorder as a failure to tell which context one is in, not a failure of fear extinction. It gives a generative model (a partially observed Markov decision process under active inference) in which a traumatic experience can leave an overly precise prior for “harmful context”. That prior stops the agent sampling the evidence that would correct it. It attaches a putative neuromodulatory circuit and a cognitive account of flashbacks, and it shows the loop and two interventions in simulation.
Key insight
The maintaining link is a choice of what to look at. Under active inference, the agent picks policies by expected information gain. If the prior is already confident that the context is harmful, looking at the periphery is “falsely believed to be uninformative due to overly informative prior beliefs”. So the agent attenuates the periphery and focuses on the foreground, whose observations are the same in both contexts. “As long as the sensory attenuation policy is selected, no amount of observations will render prior beliefs obsolete.” The prior is kept up not by what it says but by the policy it licenses.
Assumptions
- Two contexts, one shared cue. The model has a harmful and a harmless context that share an ambiguous first observation (the “triangle”) and differ only in peripheral observations.
- Two actions per time step. Exploratory sensing of the periphery, or sensory attenuation of the periphery. Three-step trials, four policies.
- Context priors are learned Dirichlet counts (the D vector), with learning and forgetting. Vulnerable and resilient agents differ in “the specific weights of learning and forgetting”. The paper gives no values.
- Drugs act on precision. SSRIs, benzodiazepines and alcohol are modelled as lowered precision in the likelihood (A) and transition (B) tensors.
- Harm, not fear. The model is framed around harm expectation and defensive mobilisation, and is meant to stay compatible with fear conditioning and extinction “without relying on these constructs”.
- The circuit is putative. The PPtg–LC acetylcholine–noradrenaline loop, the dorsal raphe serotonergic “brake” and the forebrain control are hypotheses assembled from the literature. They do not drive the simulations.
Key results
- Pathogenesis (Fig. 6). Both agents start with balanced context priors, shown as two attractor basins. Then they get a short run of harmful trials, then a long run of random trials (50/50). The resilient agent gathers peripheral evidence and rebalances. The vulnerable agent enters “a ‘vicious circle’”: “the sensory attenuation policy prevents learning about the harmless context, which biases the system towards the sensory attenuation policy, and so on”. The basins collapse into one. Holds for: one illustrative run per phenotype, with unstated parameters.
- Trial schedule (Fig. 7a). Trials 1–5 harmless, 6–16 harmful, 17–60 random. By trials 50–60 the vulnerable agent selects sensory attenuation “under any condition”.
- Successful intervention (Figs. 7b–c, 7f blue). Sixty trials of “controlled exposure to harmless condition” restore the exploratory policy near the end of the block. The priors begin to rebalance and stay balanced when random trials resume.
- Illusory intervention (Figs. 7d–e, 7f orange). Lowered A and B precision under random trials gives “more grayscale (low precision policy selection) and a wider range of selected policies”, read as fewer symptoms. When precision is restored, the attenuation policy returns even on harmless trials, because “the contextual priors were not rebalanced”. The authors say this contrast “closely matches randomised controlled trial (RCT) results”, citing van der Kolk et al. (2007), EMDR against fluoxetine. It is asserted, not fitted.
- The bistable reading (Discussion). Plasticity can create unstable basins that favour sampling, or “a collapse or canalisation into a single basin”. The collapse is “self-maintaining because it precludes epistemic foraging”, and “the saddle point between basins corresponds to a reinstatement of context-sensitivity”.
- Key hypothesis. “The balance of plasticity-enhancement and plasticity-decay for aversive events is greater in the vulnerable.”
- Testable predictions stated.
- Cued traumatic recall attenuates peripheral sensing more than focal sensing (a target/distractor paradigm).
- Differently valenced contexts change startle to the same probe less in PTSD than in controls.
- Active inference predicts more weight on trauma-encoded context, and dual representation theory less.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | An overly precise harmful-context prior can make exploration look uninformative and so block the evidence that would correct it | moderate | follows from expected-information-gain policy selection; shown in simulation (Figs. 4, 6) |
| C2 | Whether the loop forms depends on the learning and forgetting rates for aversive evidence | weak | two illustrative phenotypes, parameters not reported |
| C3 | Controlled exposure that restores exploration rebalances priors durably; precision-lowering drugs suppress symptoms without rebalancing | weak | one simulated sequence each (Fig. 7); the two arms also differ in environment |
| C4 | The simulated contrast matches the EMDR–fluoxetine RCT | weak | asserted by citation, no fit |
| C5 | A PPtg–LC–DRN neuromodulatory circuit implements the elevated tonic arousal of PTSD | weak | literature synthesis, not modelled |
| C6 | Flashbacks are a blend of focal sensation and top-down recalled context in visual working memory under peripheral attenuation | weak | narrative account; the model “does not drive the simulation” on this |
Method
- A generative model over hidden context (harmful, harmless) and state, with a shared cue at the start and context-specific peripheral observations reachable only by the exploratory action (Fig. 4).
- Policies scored by expected free energy, so exploration is chosen when it is expected to resolve uncertainty about context.
- Context priors D learned across trials as Dirichlet counts with forgetting. Vulnerable and resilient agents differ in learning and forgetting weights.
- Interventions: a block of harmless-only trials (exposure), or lowered precision of the A and B tensors under random trials (drug), followed by restoration.
Concepts
- Context. “Elements of the sensory or cognitive periphery that provide evidence for the states of affairs prior to belief-updating.”
- Contextual prior (D). The learned prior probability of being in each context at the start of a trial.
- Sensory (peripheral) attenuation. A policy that lowers precision on the periphery while keeping focus on foreground elements. The authors read it as hypervigilance: “rapid shifting of foreground focus without resolving contextual uncertainty”.
- Counterfactual blocking. From their earlier work (Moutoussis et al. 2018): an impaired ability to accumulate evidence that a context is benign, so that Bayesian learning “breaks”.
- Pathogenesis, not aetiology. The process by which a response that was adaptive becomes persistent, as opposed to its causes.
- Return slope. The decay of arousal after harm becomes unlikely. Vulnerability is “an inappropriately sluggish decay”.
Connections
The paper extends Linson, Parr and Friston (2020) and Linson and Friston (2019) on PTSD under active inference, and Moutoussis et al. (2018) on computational psychotherapy. It cites Kaye et al. (2023) and Seriès (2019) as kindred computational accounts. It names Carhart-Harris’s “canalisation” for the collapsed basin. It sets itself against dual representation theory on how much weight trauma-encoded context gets.
Bearing on the record
- QUESTION-045. The paper is an individual-scale model of exactly the
structure the question names, and it answers part of it in simulation.
- Behaviour that blocks its own disconfirmation. The sensory attenuation policy is selected because the prior is precise, and it withholds the only observations that could lower that precision.
- Capacity to leave degrades. The basins collapse to one, so the exploratory policy, the way out, stops being selected.
- The maintaining link, not the content. The two interventions both leave the prior’s content (harmful) in place at the start. The drug lowers the symptoms; exposure restores the exploratory policy. Only restoring the policy rebalances the prior. That is the question’s contrast in model form.
- Limit. The two arms are not matched. Exposure runs in a controlled harmless-only environment and the drug in a random one, so the simulation does not hold exposure fixed while cutting the link, as the question asks. And it is a model, with no individual data and no measured hysteresis.
- THEORY-058 holds that disorders are kept up by self-reinforcing loops and outlast their triggers. This model is a worked mechanism of one such loop, with a stated tipping point. It is consistent with THEORY-058 and adds no data.
- THEORY-053 says only specific change in the theorised process, ahead of the outcome, bears on a theory of change. The paper’s drug arm is an instance: the outcome (symptoms) moves without the process (context priors), and the symptoms come back. The paper reads it the same way.
- LIT-983’s filing description holds. The status note said it “models a maintained disorder as impaired belief-updating with an explicit adaptive/maladaptive boundary”. It does, but the “boundary” is a qualitative tipping point in a simulation, not an estimated threshold.
- Filed from this reading: THEORY-285, the self-sustaining prior and the drug–exposure contrast, at simulation strength.
- Anthology. The paper carries treatment suggestions (psychotherapy that restores context sensitivity, drugs as adjuncts, caution about immediate debriefing). They are clinical, not machine-learning practice, so nothing goes to the anthology.
Limitations
- Simulation only. No data are fitted. The phenotypes, schedules and precision changes are chosen to illustrate, and the parameter values are not given in the main text.
- Version 2 only. The current version 3 was not reached and may differ.
- The circuit is not modelled. The neuromodulatory account and the flashback account are narrative, assembled from cited findings.
- The RCT match is by citation, against one trial, with no quantitative comparison.
- Unmatched intervention arms, as above.
- Two contexts. Generalisation to many contexts, or to a graded harm probability, is not shown.
- The Experts by Lived Experience feedback is endorsement of the account’s fit to experience, not evidence about mechanism, and the paper does not present it as such.
Open questions
- The paper’s own tests: peripheral against focal attenuation under cued recall, context-modulated startle, and the opposite predictions from dual representation theory.
- Can learning and forgetting rates for aversive context be estimated in people after trauma, and do they predict who develops PTSD? That would turn C2 from illustration into a finding.
- In the model, does exposure under a random environment, matched to the drug arm, still rebalance the priors? That is the comparison QUESTION-045 needs, and the model could run it.