Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 87 of the 90 references it cites were found in a registry or an index; 13 of the 25 sources its claims cite were checked for support.

Testing predictive coding theories of autism spectrum disorder using models of active inference

Open access

Abstract

Several competing neuro-computational theories of autism have emerged from predictive coding models of the brain. To disentangle their subtly different predictions about the nature of atypicalities in autistic perception, we performed computational modelling of two sensorimotor tasks: the predictive use of manual gripping forces during object lifting and anticipatory eye movements during a naturalistic interception task. In contrast to some accounts, we found no evidence of chronic atypicalities in the use of priors or weighting of sensory information during object lifting. Differences in prior beliefs, rates of belief updating, and the precision weighting of prediction errors were, however, observed for anticipatory eye movements. Most notably, we observed autism-related difficulties in flexibly adapting learning rates in response to environmental change (i.e., volatility). These findings suggest that atypical encoding of precision and context-sensitive adjustments provide a better explanation of autistic perception than generic attenuation of priors or persistently high precision prediction errors. Our results did not, however, support previous suggestions that autistic people perceive their environment to be persistently volatile.

Overview

CoChat Summary

## TL;DR

This paper uses active-inference-style computational models to test competing predictive-coding accounts of autism across two sensorimotor tasks: object lifting and anticipatory eye movements in a VR interception task. It found no autism-related differences in object-lifting priors or sensory weighting, but did find task- and context-dependent differences in eye-movement beliefs, learning rates, and volatility estimates.

## Key findings

  • Experiment 1 (object lifting, n = 55 analyzed): there were no significant autism-related differences in beliefs about p(weight|size), decision temperature (ζ), or learning-related parameters ω2, ω3, α2, α3, α4; for example, beliefs about p(weight|size) showed t(53) = 1.37, p = .18, d = 0.37, and ζ showed t(53) = 0.07, p = .95, d = 0.02.
  • In Experiment 1, the HGF4 Bayesian model was the best fit, with protected exceedance probability = 1.00; the veridical-input variants had higher log model evidence than felt-input variants in all cases.
  • Experiment 2 (interceptive eye movements, n = 59 analyzed): the HGF3 model was best, with protected exceedance probability = 1.00, and it fit better than associative learning models.
  • In Experiment 2, autism-related differences appeared in posterior beliefs and learning under volatility: group differences were found for μ2 (F(1,57) = 6.60, p = .01, ηp2 = 0.10), μ3 (F(1,57) = 2.28, p = .14, ηp2 = 0.04 overall, but t(57.0) = 3.00, p = .008, d = 0.72 in volatile trials), ω (F(1,57) = 4.75, p = .03, ηp2 = 0.08), and learning rates α2 (F(1,57) = 6.40, p = .01, ηp2 = 0.10) and α3 (H(1) = 3.90, p = .048), with neurotypical participants showing higher values in volatile trials.

## Methods

  • The study reanalyzed two previously collected datasets using computational modeling: an object-lifting size-weight illusion task and a VR ball-interception task with eye tracking.
  • Experiment 1: 58 participants initially, with 55 included in analyses after excluding 3 autistic participants for invalid grip-force data; groups were matched on age, gender, and dominant hand. The main measures were peak grip force rate (pGFR) and heaviness ratings, modeled with Hierarchical Gaussian Filter (HGF) and Rescorla-Wagner / Sutton K1 alternatives.
  • Experiment 2: 90 participants initially, with 59 included in analyses after excluding datasets with missing gaze data; the main measure was gaze pitch angle at bounce, modeled with HGF3/HGF4 and Rescorla-Wagner / Sutton K1. Model comparison used Bayesian model selection.

## Limitations

  • The paper states that Experiment 1 had minimal changes in environmental contingencies, so firm conclusions about volatility beliefs could not be drawn there.
  • The authors note that the study had limited power for smaller effects: in Experiment 1, power was 85% for d ≈ 0.8 but only 50% for d ≈ 0.5; in Experiment 2, power was 85% for f ≈ 0.4 but only 50% for f ≈ 0.25.
  • Evident from the text: this is a reanalysis of existing datasets, so conclusions depend on the original task designs and available measures rather than newly designed experiments.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1371/journal.pcbi.1011473

Cites

87 of 87 references verified0 citation errors · 0 not checkable

Statistics

31 tests recomputed · 0 with the reported and recomputed p on opposite sides of .05

Figures

Fig 1 Basic HGF framework. Panel A shows the conceptualisation of an agent connected to the world by the sensory information it receives (u) and the actions it takes (y), as described in active inference. Beliefs about the world depend on inferences about true states (x) based on sensory input. The sensory (u) and active (y) states correspond to the ‘Markov blanket’ which connects the agent with, but also distinguishes it from, the surrounding environment [52]. Panel B shows the basic structure of the HGF. The HGF estimates the inference processes that best explain the behavioural responses of an agent given a time series of inputs [50,53]. Model parameters can be estimated by inverting the model to infer participant-specific parameters and belief trajectories [50]. The perceptual model (χ in Panel A) is described via beliefs (x) represented at multiple layers that evolve across time (k), scaled by variance parameters (ω, ϑ). In B, the variance parameter ϑ controls the rate of change of change in x3, while ω controls the rate of change in x2.
Fig 1 Basic HGF framework. Panel A shows the conceptualisation of an agent connected to the world by the sensory information it receives (u) and the actions it takes (y), as described in active inference. Beliefs about the world depend on inferences about true states (x) based on sensory input. The sensory (u) and active (y) states correspond to the ‘Markov blanket’ which connects the agent with, but also distinguishes it from, the surrounding environment [52]. Panel B shows the basic structure of the HGF. The HGF estimates the inference processes that best explain the behavioural responses of an agent given a time series of inputs [50,53]. Model parameters can be estimated by inverting the model to infer participant-specific parameters and belief trajectories [50]. The perceptual model (χ in Panel A) is described via beliefs (x) represented at multiple layers that evolve across time (k), scaled by variance parameters (ω, ϑ). In B, the variance parameter ϑ controls the rate of change of change in x3, while ω controls the rate of change in x2.CC BY · PubMed Central
Fig 2 Results of the model fitting. Panels A-C show the results of the model fitting and comparison procedures, illustrating the log-model evidence (LME) for all models (A) and for the veridical compared to felt models (B). Panel C illustrates the probabilities of the different models in the participant population based on Bayesian model selection, where HG4 was the most likely generative structure. Panel D shows a schematic of the winning HGF4 model. Panel E shows the parameter identifiability matrix for the HGF4, which shows that no model parameters were highly correlated (i.e., one could not simply be substituted for another). Panel F shows a comparison of the log model evidence per participant between the veridical and felt versions of the HGF4, illustrating the veridical version was better for most participants.
Fig 2 Results of the model fitting. Panels A-C show the results of the model fitting and comparison procedures, illustrating the log-model evidence (LME) for all models (A) and for the veridical compared to felt models (B). Panel C illustrates the probabilities of the different models in the participant population based on Bayesian model selection, where HG4 was the most likely generative structure. Panel D shows a schematic of the winning HGF4 model. Panel E shows the parameter identifiability matrix for the HGF4, which shows that no model parameters were highly correlated (i.e., one could not simply be substituted for another). Panel F shows a comparison of the log model evidence per participant between the veridical and felt versions of the HGF4, illustrating the veridical version was better for most participants.CC BY · PubMed Central
Fig 3 Results of the parameter comparison. Panel A: Illustration of grip and load force on the force transducers. Panel B: The stimulus objects. Panel C: Illustration of the relationship between mass and volume in the stimuli. Panel D: Plots show individual data points with mean and 95% confidence intervals for comparisons of model parameters between autistic and neurotypical groups.
Fig 3 Results of the parameter comparison. Panel A: Illustration of grip and load force on the force transducers. Panel B: The stimulus objects. Panel C: Illustration of the relationship between mass and volume in the stimuli. Panel D: Plots show individual data points with mean and 95% confidence intervals for comparisons of model parameters between autistic and neurotypical groups.CC BY · PubMed Central
Fig 4 Results of the model fitting. Panel A shows the task environment and Panel B shows the winning HGF3 model structure. Panels C and D illustrate the log-model evidence (LME) for all models and the probabilities of the different models in the participant population as indicated by Bayesian model selection. Panel E shows the belief trajectory of the fitted model for a single participant. In the bottom section, observed outcomes are shown in green (1 = normal ball, 0 = bouncy). The inferred posterior belief about the likelihood of a normal or bouncy ball (i.e., s(μ2)) is in red and the binary distributed response variable (pitch angle: above or below median) in pink. The thin black line shows the learning rate. Mean and 95% CIs for μ2 and μ3 are shown in blue in the middle and top panels.
Fig 4 Results of the model fitting. Panel A shows the task environment and Panel B shows the winning HGF3 model structure. Panels C and D illustrate the log-model evidence (LME) for all models and the probabilities of the different models in the participant population as indicated by Bayesian model selection. Panel E shows the belief trajectory of the fitted model for a single participant. In the bottom section, observed outcomes are shown in green (1 = normal ball, 0 = bouncy). The inferred posterior belief about the likelihood of a normal or bouncy ball (i.e., s(μ2)) is in red and the binary distributed response variable (pitch angle: above or below median) in pink. The thin black line shows the learning rate. Mean and 95% CIs for μ2 and μ3 are shown in blue in the middle and top panels.CC BY · PubMed Central
Fig 5 Mean and 95% CIs of model-derived estimates. Figures show the decision parameter, ζ, (panel A), the random walk variance parameter, ω, (panel B), posterior beliefs about x2 (ball bounciness; panel C) and x3 (environmental volatility; panel D), and learning rates at levels 2 (panel E) and 3 (panel F) of the HGF. Thicker significance bars indicate group-level effects and thinner bars indicate pair-wise differences. *p < .05, **p < .01, ***p < .001.
Fig 5 Mean and 95% CIs of model-derived estimates. Figures show the decision parameter, ζ, (panel A), the random walk variance parameter, ω, (panel B), posterior beliefs about x2 (ball bounciness; panel C) and x3 (environmental volatility; panel D), and learning rates at levels 2 (panel E) and 3 (panel F) of the HGF. Thicker significance bars indicate group-level effects and thinner bars indicate pair-wise differences. *p < .05, **p < .01, ***p < .001.CC BY · PubMed Central

Verification

Claims the paper makes

10 Valid · 0 Not valid · 6 Not checked · 68 Own results and interpretations

53 more claims

Create a free account to see all 84 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 15
  • Probably not a substantial claim. 16

Create a free account to see these claims.

Statistics that don’t add up

Each reported p-value recomputed from its test statistic and degrees of freedom.

1 of 1

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Doesn’t add upF(1, 57) = 0.45reported p = .45 · recomputed p = .505

    “F(1,57) = 0.45, p = .45”

  • Not confirmedz = 376reported p = .76 · recomputed p < .001

    “[z = 376, p = .76”

  • Not confirmedz = 238reported p = .047 · recomputed p < .001

    “[z = 238, p = .047”

  • Consistentt(53) = 1.37reported p = .18 · recomputed p = .176

    “t(53) = 1.37, p = .18”

  • Consistentt(53) = 0.07reported p = .95 · recomputed p = .944

    “t(53) = 0.07, p = .95”

5 of 31 · the rest aren’t loaded

26 more recomputed tests

Create a free account to see every test the Stats check recomputed.

See all tests — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Citations

Cites

Show all 90 references

Cited by

See all 23

Metadata

Authors
Arthur, T., Vine, S., Buckingham, G., Brosnan, M., Wilson, M., & Harris, D.
Affiliations
  • University of Exeter (GB)
  • University of Bath (GB)
Corresponding author
  • Arthur, Tom
  • Harris, David
Venue
PLoS Computational Biology
Published
2023-09-11
Volume
19
Issue
9
Pages
e1011473
ISSN
  • 1553-734X (ISSN-L)
  • 1553-7358
OpenAlex
W4386616703
Cited by
21
References
90
Citations by year
2026: 6 · 2025: 9 · 2024: 2 · 2023: 4
FWCI
3.14
Type
journal article
Language
en
Open access
Locations
  • PLoS Computational Biology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Figshare (submittedVersion, cc-by)
  • Open Research Exeter (University of Exeter) (submittedVersion, cc-by)
  • PsyArXiv (OSF Preprints) (submittedVersion, cc-by)
  • PsyArXiv (OSF Preprints) (submittedVersion)
Funders
  • Economic and Social Research Council (ES/P000630/1)
  • Leverhulme Trust (Fellowship)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Embodied and Extended Cognition
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Autism Spectrum Disorder Research
  • Social Sciences › Psychology › Social Psychology › Action Observation and Synchronization
Keywords
autism spectrum disorder, predictive coding, active inference, sensorimotor perception, object lifting, anticipatory eye movements, interception, prior beliefs, belief updating, precision weighting, prediction error, adaptive learning rate, environmental volatility, computational modelling
MeSH terms
Autism Spectrum Disorder, Autistic Disorder, Brain, Eye Movements, Humans, Learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Testing predictive coding theories of autism spectrum disorder using models of active inference. https://app.cochat.ai/papers/10.1371/journal.pcbi.1011473
The paper
Arthur, T., Vine, S., Buckingham, G., Brosnan, M., Wilson, M., & Harris, D. (2023). Testing predictive coding theories of autism spectrum disorder using models of active inference. PLoS Computational Biology. https://doi.org/10.1371/journal.pcbi.1011473

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.