Remembering the "when": Hebbian memory models for the time of past events.

In a dynamic world with continuous changes, the human brain constantly tries to predict what is going to happen next to optimise neural resource deployment. The Bayesian Brain framework conceptualises predictive processing based on the integration of a probability distribution, capturing the expectations as the prior, which is combined with current sensory input to update the expectation, i.e. the posterior []. Applications of the framework to Autism Spectrum Disorder (ASD) have been very promising [–]. ASD is a neurodevelopmental disorder characterised by two symptom clusters: difficulties in social interactions and communication and restrictive or repetitive interests and behaviour []. Central to the Bayesian Brain framework is how the system responds when predictions and incoming sensory input diverge. This mismatch results in a prediction error weighted by the estimated precision of the input, determining the extent to which predictions are updated []. However, the update depends not only on the precision of the input but also on the perceived stability of the environment. This environmental volatility can be separated into tonic and phasic components to distinguish between transient noise and sudden environmental changes []. Tonic volatility represents the stable, background level of change that is expected to persist over time, acting as a baseline expectation of environmental stability, where increased levels of tonic environmental volatility reflect a belief that the environment is inherently prone to frequent change []. In contrast, phasic volatility refers to the dynamic and situational fluctuations in uncertainty that occur in response to specific, surprising events []. Optimal learning requires a balance between tonic and phasic volatility to appropriately adjust behaviour to the underlying stability of the world. In the early 2010s, Pellicano and Burr [] proposed that both hyper- and hyposensitivity in the same autistic individual can be explained by hypo-priors, i.e., less precise priors, resulting in autistic people relying more heavily on sensory input due to decreased influence of the prior. In response, Brock [] noted that this pattern could be explained by both top-down influences, i.e., hypo-priors, or bottom-up influences like reduced sensory noise resulting in an increased weight attributed to sensory input. Subsequently, theories were proposed considering hierarchical levels to reflect the cortical hierarchy central to predictive coding, the neurobiological formulation of the Bayesian Brain framework. Van de Cruys and colleagues [] proposed that autistic individuals inflexibly attribute higher weight to prediction errors. Van de Cruys and colleagues argued that this pattern explains a wide array of symptoms, including deficits in executive functioning and social cognition. Lawson and colleagues [] emphasised that precision develops dynamically on different hierarchical levels and proposed that symptoms of autism can be explained by an imbalance of precision attributed to the sensory input relative to the priors [], specifically on higher levels of the hierarchy []. In a recent meta-analysis, these accounts were summarised under the label “imbalance hypothesis” with the authors arguing all of them proposing imbalances leading to a greater reliance on sensory inputs []. This meta-analysis found mixed results regarding the integration of priors possibly to low power and inconsistent approaches, but various studies showed differences in learning between autistic and non-autistic individuals []. Recently, Shi et al. [] proposed the atypical iterative prior updating account following which autistic individuals integrate prior knowledge similarly to non-autistic individuals for immediate perceptual judgements yet revise their priors more dynamically on a trial-by-trial basis, placing greater weight on new sensory information when updating beliefs about the environment. In their central tendency task, this elevated update rate initially weakens reliance on accumulated priors, yielding reduced central tendency early in learning, but converges with the performance of non-autistic individuals as stable expectations accrue. This pattern suggests that priors form typically in ASD, despite overweighting of sensory information in believe updating. Most research focusing on predictive processing in ASD used tasks not directly capturing core symptoms of autism. For instance, Lawson et al. [] adapted a probabilistic associative learning task to experimentally manipulate changes in learned expectations and sensory noise to assess effects thereof on behaviour in a non-social context. In their binary classification task, participants had to differentiate between a house or a human face (see also []). The visual stimuli were preceded by either a low or a high tone, which was predictive of the next decision. In the task designed by Lawson and colleagues [], learning occurred on three levels simultaneously: