An entropic explanation of insistence on sameness in autism

Purpose: An information theory-based framework is proposed in attempt to explain insistence on sameness in autism as an instance of a general behavior pattern in which an individual tries to reduce surprise and uncertainty. It offers a new definition of autism as an impairment in which cognitive functions are restricted to discrimination, memorization and prediction of tangible properties of the environment.

Methods: An analogy between insistence on sameness and constrained minimization of the entropy metric is observed and examined for a set of assumptions that describe cognitive limitations of a person with autism. The metric is given by the formula D_H(R, M) = H(R|M)+H(M|R), where R represents sequences of random stimuli, M is a memory that stores and retrieves them, and where H(·|·) denotes their conditional entropies interpreted as surprise and uncertainty, respectively.

Results: It is first inferred that to minimize the metric an individual can learn about R (and store that knowledge in M) or can restrict R to the already known M. Then, it is concluded that insistence on sameness is a manifestation of the latter. Moreover, it is shown that the proposed framework: (1) Helps to quantify the concepts of surprise, uncertainty, sensory overload and deprivation, anxiety, comfort zone, disappointment, disorientation, pedantry, rigidness, observance or aberrant precision. (2) Leads to a list of guidelines for learning therapies and daily care routines, and allows them to be defined as optimization algorithms and implemented as programs for robotic live-in caregivers. (3) Can be validated with the help of a Turing test-like approach that requires no experiments involving individuals with autism.

Conclusion: The framework—if positively validated—will provide advantages of both theoretical and practical importance: it explains the insistent on sameness as a consequence of cognitive restrictions and offers formal foundations and design guidelines for therapies aimed at improving self-reliance of individuals with autism in basic activities of daily living.

1 Introduction

Etiology and pathogenesis of autism remain unclear as does its definition (Wiggins et al., 2009; American Psychiatric Association, DSM-5 Task Force, 2013; Davis and Plaisted-Grant, 2015; Sauer et al., 2021; Volkmar, 2021; Jourdon et al., 2023; Shiraishi et al., 2024; Assary et al., 2024; Mahé et al., 2025). There is no a consensus about the behavior traits that are associated with individuals with autism either (Mottron and Bzdok, 2020; Chapman and Veit, 2021; Perochon et al., 2023). While behavior, in general, appears to be driven by a variety of templates and patterns (see e.g., Lai, 2023 and cf. Bailey et al., 2024), we focus on insistence on sameness, a phenomenon referred also to as stereotyped, repetitive, and the autistic-like behavior or the autism-like trait (cf. e.g., Uljarević et al., 2017; Volkmar, 2021; Lyu et al., 2024; Kamp-Becker, 2024) observed in non-verbal low-functioning individuals (the term low-functioning is used to denote severe (profound) instances of the Level 3 of Autism Spectrum Disorder, cf. American Psychiatric Association, DSM-5 Task Force, 2013; Betty et al., 2015; Wang et al., 2022; Singer et al., 2023).

We propose a new insight into that phenomenon and examine its resemblance to a problem of constrained minimization of uncertainty, where the following entropy metric (see e.g., MacKay, 2003):

$\begin{array}{ll}{D}_H(R,M)=H(R|M)+H(M|R),& (1)\end{array}$

is used as a measure of uncertainty. Random variables, R and M, represent the real-world environment and the memory of an individual, respectively. The terms H(R|M) and H(M|R) are conditional entropies that quantify the levels of surprise and uncertainty.

The motivation for such an approach is to devise a mathematically rigorous framework that offers an explanation of the insistence on sameness which is more precise and less arbitrary than the now dominant language-based descriptions (cf. e.g., American Psychiatric Association, DSM-5 Task Force, 2013; Mottron and Bzdok, 2020; Chapman and Veit, 2021; Volkmar, 2021; Wang et al., 2022), enables formulation of learning therapies as solutions to a constrained optimization problem and facilitates their implementations as programs for e.g., robot-based daily life assistants.

The framework is inspired by a variety of results and observations from information theory, neuroscience and psychiatry (see Table 1) and its contribution is twofold, theoretical (a relatively simple and verifiable explanation of insistence on sameness as a consequence of cognitive restrictions of a person with autism) and practical (a set of design guidelines for therapies focused on low-functioning non-verbal individuals), and may lead to a more precise definition of the disorder and better personalized therapies.

Table 1

Table 1. Works that led to the development of the framework.

Clearly, there are some limitations of the framework: it is a functional model, i.e., it does not define the underlying biological implementation. Is also requires new and carefully designed personalized validation experiments to be developed. In turn, limiting the scope of the framework to low-functioning non-verbal individuals with autism is a deliberate attempt to make it both formal and accurate.

In the following sections we present the framework, examine its basic properties and discuss its relations with insistence on sameness. Then, we devise the guidelines for learning therapies aimed at increasing autonomy of persons with autism, formulate a therapy as an optimization problem and discuss two framework validation approaches. Finally, we conjecture that insistence on sameness is a special case of a general minimizing uncertainty pattern (cf. Friston, 2010) and conclude the work with the proposal of a new definition of autism.

2 Methods

The framework consists of three components: the entropy metric (1), a stimuli processing loop in Figure 1, and a set of Assumptions 1 –4 that represent the cognitive constraints of a person with autism.

Figure 1

Figure 1. Stimuli processing loop.

2.1 Entropy metric

The metric (1) measures an entropy distance between two discrete random variables R and M. To illustrate its basic properties, we will consider two opposite scenarios:

1. R fully determines M (and vice versa). In this case knowledge of one variable removes the entire uncertainty about the other and thus both conditional entropy terms are zero (and so is the metric),

$D_H(R,M)={\displaystyle \underset{=0}{{\displaystyle \underbrace{H(R|M)}}}}+{\displaystyle \underset{=0}{{\displaystyle \underbrace{H(M|R)}}}}=0.$

2. Both R and M are independent of each other. Knowing either of them carries no information about the other (so they maintain their original uncertainties):

$D_H(R,M)={\displaystyle \underset{=H(R)}{{\displaystyle \underbrace{H(R|M)}}}}+{\displaystyle \underset{=H(M)}{{\displaystyle \underbrace{H(M|R)}}}}=H(R)+H(M).$

Hence, in order to reduce the entropic distance between R and M one can learn about R (and store that knowledge in M), or constrain R to the already known M.

2.2 Stimuli processing loop

The loop (see Figure 1) serves as a model of stimuli processing. Its two phases, perception, and prediction are associated with the components of the metric, H(R|M = m_n−1) and H(M|R = r_n), that quantify how surprise and uncertainty are affected by the incoming stimulus r_n. If levels of surprise or uncertainty exceed the respective thresholds, T_so or T_a, a sensory overload or anxiety state can occur.

2.3 Assumptions

To specify the cognitive constraints of a non-verbal low-functioning individual with autism, we assume that:

1. The stimuli, r_n, n = 0, 1, …, represent a perceived environment and form sequences. The mixture of all sequences is denoted by R.

2. A memory, M, is capable of storing sequences and retrieving their elements, m_n, n = 0, 1, ….

3. Each memory item, m_n = ρ(r_n), is a result of a classification of a single stimulus, r_n, by a nearest neighbor algorithm.

4. Both sensory overload and anxiety thresholds, T_so, T_a>0, are random variables.

2.4 Comments on assumptions

Each stimulus, r_n, can be interpreted as a “snapshot” of the environment and m_n = ρ(r_n) as its counterpart perceived by the senses and located in memory M (cf. e.g., Nobre, 2022). The sequences in the memory are represented by chains/directed graphs (see Figure 2).

Figure 2

Figure 2. An instance of a graph composed of memorized stimuli sequences, m_n = ρ(r_n), n = 0, 1, … with a set $M_{n+1}=\left\{m_{n+1}^{\prime },m_{n+1}^{\prime \prime }\right\}$ consisting of a pair of alternative events.

Remark 2.1. All stimuli are treated in a unified way, as raw (random) signals, and their sources (e.g., external/exogenous or internal/endogenous) are not distinguished. Moreover, their semantics (i.e., the non-tangible, abstract, cultural meanings and associations) are not available. Thus, only their perceptible properties are taken into account and treated as equally important; i.e., all their abstract properties and relations (see e.g., Bäck, 2014), incl. importance, hierarchy, being a part/whole are ignored.

Example 2.1. An abstract notion of nothing cannot be represented by any perceptible stimulus. This observation can also be derived from the namesake property of a nearest neighbor algorithm (cf. Assumption 3), where the classification routine always selects the closest remembered (known) item, regardless of how "far" it is in terms of a distance function.

Remark 2.2. Assumptions 2–3 define a cognitive model with the inference abilities limited to replication of the memorized stimuli and their sequences. They also imply a non-vanishing (persistent) character of the memory.

Remark 2.3. Defining the thresholds, T_so, T_a, as random variables allows them to be unknown, varying in time, specific for an individual and driven by the factors not included in the model (cf. e.g. Kim and Kim, 2023; Lai, 2023).

2.5 Analysis

The goal of this section is to inspect how the metric changes during perception and prediction phases; cf. the processing loop in Figure 1. The amount of surprise in the former, given the memorized stimulus, m_n−1, is

$\begin{array}{ll}{D}_H(R,M|M=m_{n-1})=H(R|M=m_{n-1})+{\displaystyle \underset{=0}{{\displaystyle \underbrace{H(M=m_{n-1}|R)}}}},& (2)\end{array}$

while the level of uncertainty in the latter, given the incoming stimulus, r_n, is

$\begin{array}{ll}{D}_H(R,M|M=r_n)={\displaystyle \underset{=0}{{\displaystyle \underbrace{H(R=r_n|M)}}}}+H(M|R=r_n),& (3)\end{array}$

so that during perception we need to consider only the first term, H(R|M = m_n−1), and only the other, H(M|R = r_n), during prediction. We examine them in the following three cases; cf. Section 2.1:

1. The first, deterministic, takes place when in perception phase knowledge of m_n−1 determines r_n so that there is no surprise, and when there is no uncertainty during prediction, because knowledge of r_n determines m_n+1. Then, both terms are zero:

$\begin{array}{ll}H(R=r_n|M=m_{n-1})=0\text{ and }H(M=m_{n+1}|R=r_n)=0.& (4)\end{array}$

2. The second case occurs when, in the perception phase, m_n−1 determines a set R_n of possible stimuli or, when in prediction, the set of memorized possibilities M_n+1 occurs given the stimulus r_n (cf. Figure 2). The respective metric terms are:

$\begin{array}{l}H(R=R_n|M=m_{n-1})=H(R_n)\text{ and}\\ H(M=M_{n+1}|R=r_n)=H(M_{n+1}).& (5)\end{array}$

3. The last case encompasses, amongst others, the learning-like situations when r_n is a new (unknown) stimulus, so that the current knowledge of m_n−1 carries no information about it during perception and the stimulus itself does not reduce the level of prediction uncertainty:

$\begin{array}{ll}H(R|M=m_{n-1})=H(R)\text{ and }H(M|R=r_n)=H(M).& (6)\end{array}$

It should be noted here that because the sets of known options, R_n and M_n+1, are subsets of R and M, the amounts of surprise or uncertainty in 2 seem to always be smaller than in 3 (when unknown stimuli occur). Nevertheless, for some realizations of m_n−1 and r_n the following inequalities may hold (see e.g., MacKay, 2003, Ch. 8):

$\begin{array}{ll}H(R_n|M=m_{n-1})\ge H(R)\text{ and }H(M_{n+1}|R=r_n)\ge H(M),& (7)\end{array}$

and hence the surprise and uncertainty levels in (5) can be larger than in (6).

Example 2.2. Such a situation can take place when, in prediction phase, the set M_n+1 consists of multiple options, $\left\{m_{n+1}^{\prime },m_{n+1}^{\prime \prime },\dots ,m_{n+1}^{(\nu )}\right\}$, some ν, that are equally probable; (cf. Figure 2). The amount of uncertainty is then increased (predictions are the least efficient) and can easier lead to disorientation and anxiety (especially when ν is large).

So far, we have conditioned the metric terms on the last realizations of the stimulus r_n (or on the latest memory item m_n−1), as if the amounts of surprise and uncertainty are independent of the past [like in Markov chains (see e.g., Ekroot and Cover, 1993)]. However, to better comply with Assumptions 2 –3 and Remarks 2.1–2.2, one should take into account the previous events as well.

Example 2.3. If, for instance, in the perception phase, we condition the surprise on the set M_n−1 = {m_n−1, M_n−2, …} of all previous items in a sequence, then the term H(R|M = M_n−1) can quantify a disappointment that the sequence does not continue in a way it has been memorized.

2.5.1 Classification

The influence of the classification phase (i.e., the assignment m_n = ρ(r_n) in Figure 1) has so far been ignored. This is because it resembles the impact of prediction. First, observe that the metric reduces to the same term; cf. (3):

$D_H(R=r_n,M)={\displaystyle \underset{=0}{{\displaystyle \underbrace{H(R=r_n|M)}}}}+H(M|R=r_n).$

Next, note that if the stimulus, r_n, is already known, then it fully determines the corresponding m_n (cf. Assumption 2 and Remarks 2.1–2.2). In turn, if r_n is new, it carries no information about the memory. Hence, cf. (4) and (6):

$H\left(M|R=r_n\right)=\{\begin{array}{l}0\hfill \\ H\left(M\right)\hfill \end{array}\text{ if }{r}_n\text{ is }\begin{array}{l}\text{known,}\hfill \\ \text{new.}\hfill \end{array}$

Remark 2.4. Although the presence of randomness, even in a fixed and arranged environment, seems to be inevitable (cf. e.g., Davis and Plaisted-Grant, 2015; Walker et al., 2023), the modified environments and/or the sequences can still be perceived as known as long as those random changes do not affect the results of classification; cf. Remark 2.1.

3 Results

3.1 Analogies with insistence on sameness

In case of non-verbal low-functioning individuals with autism, both sensory overload and anxiety can lead to violent, (self-)aggressive behaviors [sometimes labeled as meltdowns or tantrums (Volkmar, 2021)]. Insistence on sameness can be seen as a set of actions aimed at decreasing the risk of these events, either by observing (learning) the environment (i.e., making M better approximate R), or by constraining it to the already known one (that is, making R better resemble M). For example (cf. Uljarević et al., 2017):

1. Pedantry, together with a routinized and repetitive behavior [inc. strict following only known sequences of activities and keeping the known environments unchanged; cf. an aberrant precision in (Lawson et al., 2014)], are the means to reduce the levels of surprise or uncertainty; see Ex. 2.2, 2.3 and Remark 2.4.

2. Ritualistic and stereotyped behaviors stem from the lack of the ability to ignore items; see Remarks 2.1-2.2 and cf. Guideline 3.

3. Restrictive and rigid behavior includes actions like forcing deterministic scenarios in order to further control surprise and uncertainty [and to effectively zero their levels as in (4)], to avoid disappointments and to select a preferred (the most frequent) option in multiple option cases; cf. (5), (7) and Ex. 2.3.

4. Vigilance and observance reduce the number of options, that is, the uncertainty level caused by them in (5); see Ex. 2.2.

5. Aversion to learning is a way to avoid potentially high surprise/uncertainty levels caused by new options or new stimuli; cf. (5)-(7).

3.2 Therapeutic guidelines

In practical terms, the goal of the learning therapy is twofold:

1. Increasing autonomy of an individual (by learning new vital activities in a supervised way),

2. Reduction of aversion to exploration (by controlling uncertainty of the new environments designed for semi-supervised and unsupervised activities).

Assumptions 1-4, Remarks 2.1–2.3, and the subsequent analysis in Section 2.5, imply that the therapy is a long-term incessant process with several limitations on how it can be implemented (in particular, they exclude rule-based learning; cf. Remarks 2.1-2.2 and Betty et al., 2015; Galitsky, 2016):

1. Only tangible objects and their realistically simulated counterparts should be used in learning environments and in sequences of activities.

2. Artifacts such as pictures (especially, in simplified forms of drawings or labels), but also spoken words and sentences, should a priori be assumed as unrelated with the tangible items they represent or refer to. Hence, if necessary, these relations need to be learned separately. Abstract notions, like being a child or a parent, a teacher or a caretaker, and relations like love, trust or friendship, are not available either.

3. The meaning of tangible items depends on the context (created by other items and/or their past sequences; cf. Remark 2.2). Such a correlative-like reasoning leads to ”cum/post hoc ergo propter hoc” fallacies; see Example 3.2.

4. Remembering the exact sequences (of events and activities; cf. Remarks 2.1 and 2.2) implies they are treated as a whole and that ”shortcuts” in sequences or other ”by-the-way” activities will likely be perceived as new (and will increase levels of surprise and/or uncertainty); cf. Ex. 2.3 and (Lawson et al., 2014; Noel et al., 2025).

5. Multiple choice situations (branches in sequences; see Figure 2) introduce uncertainty even if the environment remains unchanged; cf. Cases 2–3, and Exs. 2.2–2.3 and 3.1.

Example 3.1. In order to reduce the risk of anxiety induced by predictions in multiple option/choices situations, several countermeasures could be applied: making all sequences distinct (with the help of amulet- or talisman-like artifacts that will allow to distinguish the otherwise same activities), or locating the branches at the sequence beginnings (to reduce these uncertainties as early as possible), or limiting the numbers of possible options in them.

Example 3.2. An instance of an immediate reward will likely turn into a part of a routine while its delayed version will become a part of (possibly unrelated) sequence it will occur in. This is because the delayed gratification concept relies on an abstract concept of an award and its causal relation with the past events.

3.3 Self-stimulation activities and comfort zone

The guidelines 1–5 are aimed at reducing risk of tantrums during learning/therapy. However, in case of non-verbal individuals, it is difficult to assess whether the levels of surprise and/or uncertainty are sufficiently low (w.r.t. their thresholds T_soandT_a) to safely start a new sequence of activities. The problem is of special importance at the beginning of the therapy when the environments and activities the individuals are familiar with (inc. their routines stored in M) are not known. Nevertheless, if the stimuli are treated in a unified way, so that their exogenous or endogenous nature is not distinguished (cf. Assumption 1 and Remark 2.1), we can conjecture that (cf. Davis and Plaisted-Grant, 2015):

Conjecture 3.1. Self-stimulation activities [inc. ”flapping, waving, finger wiggling, mouth opening, orofacial movements, head nodding” (see Volkmar, 2021)], are a sign of low levels of perceived surprise and uncertainty.

The rationale behind this claim is following: without external stimuli (in a state of sensory deprivation), the internal ones become dominant. If sufficiently random, they can carry an amount of entropy large enough to exceed either sensory overload or anxiety threshold. Self-stimulations can therefore (in case of low-functioning individuals, in particular) be a way to generate known (zero- or low-entropy) stimuli that keep the surprise/uncertainty levels below these thresholds (i.e., inside a ”comfort zone”).

Remark 3.1. One of the Reviewers raised a question about the role of white noise. That is an interesting and open problem as one can consider two potentially opposite, negative and positive, effects:

1. Presence of a white noise in stimuli increases their conditional entropy and the risk of sensory overload. It can affect the results of the classification and increase the risk of anxiety as well; cf. Remark 2.4.

2. In turn, as pointed out by (Davis and Plaisted-Grant 2015), noise can help avoid getting stuck in local minima during optimization and, for instance, enable generalization (abstraction) abilities. In our framework, such effect can be associated with jumping between the known sequences or finding the shortcuts within them.

White noise, if present in exogenous visual or aural stimuli, could be detected with the help of basic audio-video codecs [the more noisy environment, the less effective entropy coders; (see MacKay, 2003, Ch. 1)].

3.4 Learning therapy as optimization problem

In the framework's terms, learning a low-functioning person with autism is equivalent to presenting new stimuli and their sequences. Therefore, a learning therapy with the goals 1–2, can be formulated as the following optimization problem: Problem 2.1. Given an initial state of a memory M, find R such that

$\begin{array}{ll}{\displaystyle \underset{R}{argmax}}I(R;M),& (8)\end{array}$

subject to

$H(R|M=m_{n-1})<T_{so}\text{ and }H(M|R=r_n)<T_a,$

given Assumptions 1–4 and the processing loop in Figure 1.

The objective function (8) is the mutual information of R and M, and is related to the entropy metric in (1) via identity D_H(R, M) = H(R, M)−I(R; M), where H(R, M) is the joint entropy of R and M (see MacKay, 2003, Ch. 8; and Friston, 2010; Ramstead et al., 2023). The practical value of such a formulation is that it allows learning therapies and daily care routines to be defined as optimization algorithms and implemented as programs for robotic live-in caregivers.

4 Discussion

4.1 Validation

One can consider the following two (complementary) approaches to validate/reject the framework's assumptions and the resulting analogies:

1. A traditional (direct) variant, where the realistic environments and interactively generated sequences of new activities are presented through a “digital window” (e.g., a wall-like display with a haptic interface). The intensity (and the pace the sequences are presented with) depends, in accordance to recommendations 1 –5, on the levels of surprise and anxiety assessed from e.g., the observed behavior or from a frequency of interactions between the individual and the generated world. The long-term therapeutic effects, like increased autonomy in these environments (with possible help of appropriately programmed robotic live-in caregivers), is evaluated in a traditional way (e.g., in a form of progress questionnaires) by parents and therapists.

This is a standard approach that requires an access to individuals with autism living in their homes (or other familiar environments; cf. the analogies 1–5). Getting permissions for such arrangements is usually difficult (especially at early validation stages). Moreover, in order to satisfy reproducibility conditions, each experiment needs a tedious and time-consuming individual setup (cf. Assumption 2 and Ex. 2.3). Therefore, as an alternative (as a preliminary test phase) we propose:

1. A Turing test-like approach, in which the framework is used to create avatars (“digital twins”) residing in realistic virtual environments, where their simulated behavior is observed and validated by therapists or by early-detection computer tools; cf. (Dillion et al., 2023) and (Perochon et al., 2023), respectively.

Remark 4.1. Validation methods which are useful to assess higher-functioning individuals and which (implicitly) assume an individual is able to perceive abstracts and operate on them, e.g., by examining their propensity to ignore opportunity costs (see e.g., Da Silva et al., 2025), cannot be used here.

4.2 Final remarks

We conclude the work with a claim that reducing uncertainty is not specific to autism.

Conjecture 4.1. The phenomenon of insistence on sameness in autism is a special case of the generic behavior pattern described in Friston (2010) where “Adaptive agents must occupy a limited repertoire of states and therefore minimize the long-term average of surprise associated with sensory exchanges with the world [and that] action under the free-energy principle reduces to suppressing sensory prediction errors that depend on predicted (expected or desired) movement trajectories.”

Therefore, a distinctive character of insistence on sameness should rather be attributed to cognitive limitations of a person with autism. Thus, if correct, the claim corroborates the following definition of the disorder; cf. (Mottron and Bzdok, 2020), Assumptions 2–3 and Remark 2.1:

Definition 4.1. Autism is an impairment in which cognitive functions are restricted to discrimination, memorization and prediction of tangible properties of the environment.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

Author contributions

PŚ: Writing – original draft, Writing – review & editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This study was supported by the Department of Control and Quantum Computing, Faculty of Information and Communication Technology, Wrocław University of Science and Technology, under No. K101W04ND03, 8251051500.

Acknowledgments

The author wants to thank Prof. Jerzy Rozenblit and Prof. Linda Restifo from the University of Arizona in Tucson for numerous and influential discussions, Prof. Lynn Kern Koegel from Stanford University for her encouraging opinion about the manuscript, and the Reviewers for their inspiring and valuable comments.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

American Psychiatric Association, DSM-5 Task Force. (2013). Diagnostic and Statistical Manual of Mental Disorders, 5th Edn. American Psychiatric Publishing, Inc. Available online at: https://doi.org/10.1176/appi.books.9780890425596 (Accessed December 15, 2025).

Google Scholar

Assary, E., Oginni, O. A., Morneau-Vaillancourt, G., Krebs, G., Peel, A. J., Palaiologou, E., et al. (2024). Genetics of environmental sensitivity and its association with variations in emotional problems, autistic traits, and wellbeing. Mol. Psychiatryp 29, 2438–2446. doi: 10.1038/s41380-024-02508-6

PubMed Abstract | Crossref Full Text | Google Scholar

Bäck, A. (2014). Aristotle's Theory of Abstraction, vol. 73. Cham: Springer. doi: 10.1007/978-3-319-04759-1

Crossref Full Text | Google Scholar

Bailey, D. H., Jung, A. J., Beltz, A. M., Eronen, M. I., Hamaker, E. L., et al. (2024). Causal inference on human behaviour. Nat. Hum. Behav. 8, 1448–1459. doi: 10.1038/s41562-024-01939-z

PubMed Abstract | Crossref Full Text | Google Scholar

Betty, J. S., Ho, P. V., and Carter, M. (2015). Cognitive–behavioural approach for children with autism spectrum disorder: a literature review. J. Intellect. Dev. Disabil. 40, 213–229. doi: 10.3109/13668250.2015.1023181

Crossref Full Text | Google Scholar

Chapman, R., and Veit, W. (2021). The essence of autism: fact or artefact? Mol. Psychiatry 26, 1440–1441. doi: 10.1038/s41380-020-00959-1

PubMed Abstract | Crossref Full Text | Google Scholar

Da Silva, S., Fiebig, M., and Matsushita, R. (2025). Opportunity costs, cognitive biases, and autism. J. Mind Med. Sci. 12:11. doi: 10.3390/jmms12010011

Crossref Full Text | Google Scholar

Davis, G., and Plaisted-Grant, K. (2015). Low endogenous neural noise in autism. Autism 19, 351–362. doi: 10.1177/1362361314552198

PubMed Abstract | Crossref Full Text | Google Scholar

Dillion, D., Tandon, N., Gu, Y., and Gray, K. (2023). Can AI language models replace human participants? Trends Cogn. Sci. 27, 597–600. doi: 10.1016/j.tics.2023.04.008

PubMed Abstract | Crossref Full Text | Google Scholar

Ekroot, L., and Cover, T. M. (1993). The entropy of Markov trajectories. IEEE Trans. Inform. Theory 39, 1418–1421. doi: 10.1109/18.243461

Crossref Full Text | Google Scholar

Friston, K. (2010). The free-energy principle: a unified brain theory? Nat. Rev. Neurosci. 11, 127–138. doi: 10.1038/nrn2787

PubMed Abstract | Crossref Full Text | Google Scholar

Galitsky, B. (2016). Computational Autism. Springer. Available online at: https://doi.org/10.1007/978-3-319-39972-0 (Accessed December 15, 2025).

Google Scholar

Jourdon, A., Wu, F., Mariani, J., Capauto, D, Norton, S., Tomasini, L., et al. (2023). Modeling idiopathic autism in forebrain organoids reveals an imbalance of excitatory cortical neuron subtypes during early neurogenesis. Nat. Neurosci. 26, 1505–1515. doi: 10.1038/s41593-023-01399-0

PubMed Abstract | Crossref Full Text | Google Scholar

Kamp-Becker, I. (2024). Autism spectrum disorder in ICD-11–a critical reflection of its possible impact on clinical practice and research. Mol. Psychiatry 29, 633–638. doi: 10.1038/s41380-023-02354-y

PubMed Abstract | Crossref Full Text | Google Scholar

Kim, E. J., and Kim, J. J. (2023). Neurocognitive effects of stress: a metaparadigm perspective. Mol. Psychiatry 28, 2750–2763. doi: 10.1038/s41380-023-01986-4

PubMed Abstract | Crossref Full Text | Google Scholar

Lai, M.-C. (2023). Mental health challenges faced by autistic people. Nat. Hum. Behav. 7, 1620–1637. doi: 10.1038/s41562-023-01718-2

PubMed Abstract | Crossref Full Text | Google Scholar

Lawson, R. P., Rees, G., and Friston, K. J. (2014). An aberrant precision account of autism. Front. Hum. Neurosci. 8:302. doi: 10.3389/fnhum.2014.00302

PubMed Abstract | Crossref Full Text | Google Scholar

Lyu, T.-J., Ma, J., Zhang, X.-Y., Xie, G.-G., Liu, C., Du, J., et al. (2024). Deficiency of FRMD5 results in neurodevelopmental dysfunction and autistic-like behavior in mice. Mol. Psychiatry 29, 1253–1264. doi: 10.1038/s41380-024-02407-w

PubMed Abstract | Crossref Full Text | Google Scholar

MacKay, D. J. (2003). Information Theory, Inference and Learning Algorithms. Cambridge University Press. Available online at: https://www.inference.org.uk/mackay/itila/ (Accessed December 15, 2025).

Google Scholar

Mahé, O., Morel-Kohlmeyer, S., Briend, F., and Houy-Durand, E. (2025). Predictors of psychotropic medication use among autistic adults. J. Autism Dev. Disord. 1–12. doi: 10.1007/s10803-025-06966-x

PubMed Abstract | Crossref Full Text | Google Scholar

Mottron, L., and Bzdok, D. (2020). Autism spectrum heterogeneity: fact or artifact? Mol. Psychiatry 25, 3178–3185. doi: 10.1038/s41380-020-0748-y

PubMed Abstract | Crossref Full Text | Google Scholar

Nobre, A. C. (2022). Opening questions in visual working memory. J. Cogn. Neurosci. 35, 49–59. doi: 10.1162/jocn_a_01920

PubMed Abstract | Crossref Full Text | Google Scholar

Noel, J.-P., Balzani, E., Acerbi, L., Benson, J., Angelaki, D., et al. (2025). A common computational and neural anomaly across mouse models of autism. Nat. Neurosci. 28, 1519–1532. doi: 10.1038/s41593-025-01965-8

PubMed Abstract | Crossref Full Text | Google Scholar

Palmer, C. J., Lawson, R. P., and Hohwy, J. (2017). Bayesian approaches to autism: towards volatility, action, and behavior. Psychol. Bull. 143, 521–542. doi: 10.1037/bul0000097

PubMed Abstract | Crossref Full Text | Google Scholar

Perochon, S., Di Martino, J. M., Carpenter, K. L. H., Compton, S., Davis, N., Eichner, B., et al. (2023). Early detection of autism using digital behavioral phenotyping. Nat. Med. 29, 2489–2497. doi: 10.1038/s41591-023-02574-3

PubMed Abstract | Crossref Full Text | Google Scholar

Ramstead, M. J., Sakthivadivel, D. A., Heins, C., Koudahl, M., Millidge, B., Da Costa, L., Klein, B., and Friston, K. J. (2023). On bayesian mechanics: a physics of and by beliefs. Interface Focus 13:20220029. doi: 10.1098/rsfs.2022.0029

PubMed Abstract | Crossref Full Text | Google Scholar

Sauer, A. K., Stanton, J., Hans, S., and Grabrucker, A. (2021). Autism Spectrum Disorders: Etiology and Pathology. Exon Publications, 1–15. Available online at: https://doi.org/10.36255/exonpublications.autismspectrumdisorders.2021.etiology (Accessed December 15, 2025).

PubMed Abstract | Google Scholar

Shiraishi, T., Katayama, Y., Nishiyama, M., Shoji, S., Miyakawa, T., Mizoo, T., et al. (2024). The complex etiology of autism spectrum disorder due to missense mutations of CHD8. Mol. Psychiatry 29, 2145–2160. doi: 10.1038/s41380-024-02491-y

PubMed Abstract | Crossref Full Text | Google Scholar

Singer, A., Lutz, A., Escher, J., and Halladay, A. (2023). A full semantic toolbox is essential for autism research and practice to thrive. Autism Res. 16, 497–501. doi: 10.1002/aur.2876

PubMed Abstract | Crossref Full Text | Google Scholar

Uljarević, M., Richdale, A. L., Evans, D. W., Cai, R. Y., and Leekam, S. R. (2017). Interrelationship between insistence on sameness, effortful control and anxiety in adolescents and young adults with autism spectrum disorder (ASD). Mol. Autism 8: 36. doi: 10.1186/s13229-017-0158-4

PubMed Abstract | Crossref Full Text | Google Scholar

Volkmar, F. R. (2021). Encyclopedia of Autism Spectrum Disorders. Cham: Springer. doi: 10.1007/978-3-319-91280-6

Crossref Full Text | Google Scholar

Walker, E. Y., Pohl, S., Denison, R. N., Barack, D. L., Lee, J., Block, N., et al. (2023). Studying the neural representations of uncertainty. Nat. Neurosci. 26, 1857–1867. doi: 10.1038/s41593-023-01444-y

PubMed Abstract | Crossref Full Text | Google Scholar

Wang, Z., Liu, J., Zhang, W., Nie, W., and Liu, H. (2022). Diagnosis and intervention for children with autism spectrum disorder: a survey. IEEE Trans. Cogn. Dev. Syst. 14, 819–832. doi: 10.1109/TCDS.2021.3093040

Crossref Full Text | Google Scholar

Wiggins, L. D., Robins, D. L., Bakeman, R., and Adamson, L. B. (2009). Brief eport: sensory abnormalities as distinguishing symptoms of autism spectrum disorders in young children. J. Autism Dev. Disord. 39, 1087–1091. doi: 10.1007/s10803-009-0711-x

Crossref Full Text | Google Scholar