The Hard Case: Autism and Stimming
This is where I worked the model out first, and it's the application I hold with the least confidence. Both of those are worth saying up front.
A framework earns its keep on hard cases. Autistic stimming is the hardest one I know, because the behavior is common, the accounts of it conflict, the intervention history is genuinely ugly, and the people it describes are right there and able to say whether the description fits.
Everything below is a hypothesis about organization. Nothing in it should change what anyone does tomorrow, and the last section says so in plain terms.
What is stimming, and what is it for?
Stimming is repetitive, usually rhythmic, self-generated activity across any sensory channel: hand flapping, rocking, finger flicking, spinning, pressing, humming, muttering, watching moving light. It isn't exclusively autistic. What distinguishes autistic stimming is frequency, intensity, form, and, decisively, the social response it draws. Autistic adults describe it as regulation rather than distress relief.
Kapp and colleagues interviewed 32 autistic adults about it, and the finding that matters is directional (Kapp et al., 2019). Stimming responded to intense emotion in both directions. Some participants stimmed when anxious or distressed. Others when excited or happy.
As Rebecca put it:
"[s]timming is just a release of any high emotion, so really anxious, really agitated, really happy, really excited, just any high emotion, that's when I stim."
The valence varied across participants. The potency of the state was what stayed constant. Two participants described happy hand flapping as hands open and arms out, unlike the tighter, closer movement of distress. Same behavior, different topography by valence.
Sinead's account of childhood spinning isn't a description of soothing at all:
"I remember as a child spinning all the time and loving spinning and loving swinging and feeling that movement all the time, but then I also realised that there was a point where it wasn't acceptable to be spinning anymore … so it actually still feels glorious if there's nobody around and I can skip or I can spin and it's like I'm breaking the rules."
Two words get used interchangeably for what stimming does, and they aren't equivalent. Self-soothing is an observer's inference, what the behavior looks like it's for, from outside, in a person who appears distressed. Self-regulation is closer to what autistic people report, and broader in a way that matters.
I'll own my share of that first one. I spent years describing this behavior with a word nobody had ever tested against what the person doing the behavior reported. The label traveled on how plausible it sounded from across the room.
A behavior that occurs as reliably in pleasure as in distress is doing something other than soothing. It's bringing an aroused system back toward a workable range, from whichever side it departed.
The term carries a history worth knowing. "Self-stimulatory behaviour" entered the literature through behavior-analytic research on autistic children, where it named repetitive behavior that appeared to compete with instruction (Lovaas et al., 197190035-0)). If a behavior is reinforced by its own sensory feedback and competes with learning, the intervention logic points straight at extinction. Generations of autistic children were taught to sit on their hands.
"Stimming" is the community's reclamation of that term, and the shortening isn't cosmetic. It drops the pathologizing frame and keeps the descriptive content.
What does the control model say stimming is?
A limit cycle. A loop with high gain and low damping doesn't sit still at its target. It overshoots, corrects, overshoots again, and settles into a self-sustaining oscillation with a characteristic frequency. Rocking, hand movements, pacing and vocal repetition all have that form: rhythmic, roughly periodic, self-terminating when the disturbance passes. The claim isn't that stereotypy is caused by oscillation. It's that stereotypy is oscillation, in the engineering sense.
Which makes it measurable. If stereotypy is oscillation it should have frequency structure, and that structure should respond to loop parameters rather than to how unpleasant the room is.
Two other features fall out without extra assumptions.
Insistence on sameness. An under-damped loop is fine until something perturbs it. Its vulnerability is the ringing that follows a disturbance. A system that oscillates badly after every perturbation has a rational interest in preventing perturbations. Read that way, insistence on sameness stops looking like rigidity as a trait and starts looking like disturbance avoidance in a system whose recovery is expensive.
Meltdown escalation. Escalation is what an under-damped loop does when the disturbance persists, or when corrections generate new error. A meltdown that ends in exhaustion rather than resolution maps onto instability better than onto emotional intensity.
That distinction is testable, and cheaply. An intensity account predicts response size scales with trigger size. An instability account predicts that past a threshold, response size becomes largely independent of trigger size.
Why would more feedback make performance worse?
Because raising the gain on a loop that's already near its stability limit pushes it over. This is the load-bearing inference in the whole autism application, and it's what separates a stability reading from a sensory-overload reading. Overload predicts degradation that tracks how aversive the stimulus is. Instability predicts degradation of a particular kind: more oscillation and more regular output, steeper at the high-gain end.
Four observations separate the two accounts, and they lean toward stability.
The eyes-open result. Under overload, closing your eyes removes an input and should reduce distress, so sway should worsen with the eyes closed, as it does in nearly every other population, because vision is stabilizing. Bloomer and colleagues found the group difference more apparent with the eyes open (Bloomer et al., 2025). Adding a feedback channel made it worse. That's what raising loop gain does, and not what removing sensory burden does.
The dynamic-only result. Lidstone and colleagues found the deficit for moving targets and not static ones, with the same display in both (Lidstone et al., 2020). Intensity was constant. The demand on closed-loop tracking was not.
The form of the degradation. Excess power below 0.25 Hz, more periodic reactive adjustments, and reduced entropy with increased regularity all describe output becoming more rhythmically structured. Overload and distraction produce noisier output. Instability produces more structured output, because a limit cycle is highly regular.
Degradation at both extremes. Mosconi and colleagues found worse performance at highly amplified and highly degraded feedback gain (Mosconi et al., 2015). Too much loop gain destabilizes. Too little leaves error uncorrected. An intensity account has no reason to predict that reducing feedback resolution also hurts.
None of this is decisive. Reduced entropy has other explanations, and stability is the more economical reading rather than the proven one. Overload has not been excluded.
How good is this evidence, honestly?
Weak, and weak in a specific way. Every finding above was collected to answer a different question and interpreted in different terms. Reinterpreting existing data is legitimate and it's how fields often move, but it's much weaker than predicting a result before you see it. Four findings that fit aren't impressive if there are forty that don't and nobody counted. Nobody counted.
The model has not been tested. No study has manipulated observer gain and controller gain independently in autistic participants, measured damping directly, or predicted a treatment response from loop parameters and then confirmed it.
None of the researchers cited above set out to test a gain-and-damping model, and none would necessarily accept this reading of their data.
The specific risk is flexibility. Four free parameters, plus delay and arbitration, can accommodate almost any outcome after the fact.
Sensory intervention works: observer gain. It fails: wrong parameter. Exercise helps: damping. It doesn't: wrong dose.
Each move is individually reasonable and collectively fatal.
The model also doesn't handle hypo-reactivity cleanly. Low registration can be modeled as low gain on a channel, but it can equally reflect attentional allocation, a detection threshold, or competition from a channel currently consuming the available precision. The framework accommodates hypo-reactivity. It doesn't predict it, and that asymmetry is a real weakness.
What would show this is wrong?
The strongest single test is cheap and nobody has run it. Raise display gain experimentally and see whether observable repetitive behavior rises dose-dependently within the same session. Established apparatus, an unambiguous manipulation, and a result that came back flat would remove the model's claim on the autism phenotype outright rather than merely complicating it. That is the test I would run first if I had a lab.
Six more criteria sit alongside it in the full document. Three are worth naming here.
If reducing sensory intensity reliably outperformed increasing predictability, the account is wrong, because the model says the problem is loop stability rather than input magnitude.
If repetitive behavior turned out to have no characteristic frequency structure, the oscillation claim collapses. Stereotypy would be repetitive without being periodic, and a limit cycle is nothing if not periodic.
If observer-side and controller-side atypicality never dissociated in real people, the two dials aren't two dials.
These only work as protection if they're treated as commitments rather than as a list. That's the whole reason Part 7 exists.
What should a clinician actually do with this?
Almost nothing differently, and that's deliberate. DO use it to explain to a family why predictability helps, and why suppressing a behavior may cost more than it saves. DON'T use it as a basis for preferring any treatment over any other. Every intervention worth recommending stands or falls on its own evidence, and in most cases that evidence is thin.
Two cautions that have nothing to do with control theory and outrank everything above.
The framing in autism research has moved, unevenly, from a deficit model toward a neurodevelopmental-difference model, in which much of the disability sits in the mismatch between person and environment (Pellicano & den Houting, 2022). That consequence is concrete. Under a deficit model the target of intervention is the person's behavior. Under a difference model the target is the fit. Same assessment, two different plans, and a mechanistic model of stimming can be pressed into service for either one. I intend it for the second.
And Milton's double empathy problem is the corrective worth carrying into any room where this gets discussed: the breakdown in mutual understanding between autistic and non-autistic people runs in both directions, and the clinician is one of the two parties (Milton, 2012).
Any sentence beginning "autistic people are" is a hypothesis to check with the person in front of you. That includes every sentence in this article.
Next, the same three failures, in a business.
References
Bloomer, B. F., Bolbecker, A. R., Gildea, E. L., Kennedy, D. P., Wisner, K. M., O'Donnell, B. F., & Hetrick, W. P. (2025). Postural sway dynamics in adults across the autism spectrum: A multifactor approach. Molecular Autism, 16(1), 44. https://doi.org/10.1186/s13229-025-00676-y
Kapp, S. K., Steward, R., Crane, L., Elliott, D., Elphick, C., Pellicano, E., & Russell, G. (2019). 'People should be allowed to do what they like': Autistic adults' views and experiences of stimming. Autism, 23(7), 1782–1792. https://doi.org/10.1177/1362361319829628
Lidstone, D. E., Miah, F. Z., Poston, B., Beasley, J. F., Mostofsky, S. H., & Dufek, J. S. (2020). Children with autism spectrum disorder show impairments during dynamic versus static grip-force tracking. Autism Research, 13(12), 2177–2189. https://doi.org/10.1002/aur.2370
Lovaas, O. I., Litrownik, A., & Mann, R. (1971). Response latencies to auditory stimuli in autistic children engaged in self-stimulatory behavior. Behaviour Research and Therapy, 9(1), 39–49. https://doi.org/10.1016/0005-7967(71)90035-0
Milton, D. E. M. (2012). On the ontological status of autism: The 'double empathy problem'. Disability & Society, 27(6), 883–887. https://doi.org/10.1080/09687599.2012.710008
Mosconi, M. W., Mohanty, S., Greene, R. K., Cook, E. H., Vaillancourt, D. E., & Sweeney, J. A. (2015). Feedforward and feedback motor control abnormalities implicate cerebellar dysfunctions in autism spectrum disorder. The Journal of Neuroscience, 35(5), 2015–2025. https://doi.org/10.1523/JNEUROSCI.2731-14.2015
Pellicano, E., & den Houting, J. (2022). Annual Research Review: Shifting from 'normal science' to neurodiversity in autism science. Journal of Child Psychology and Psychiatry, 63(4), 381–396. https://doi.org/10.1111/jcpp.13534
The full documents, free and ungated