Mechanisms
Memory Systems · Non-Declarative
Non-Associative Learning (Habituation and Sensitization)
Example
The first time the upstairs neighbor drops something you startle. By the fourth week you do not hear it. Nothing was learned about the noise, and you could not report how many times it happened. The response to it changed.
Non-Associative Learning in Humans
Definition
Repeated exposure to a stimulus changes the response to it, with no pairing between stimuli and no record of the individual presentations required; what changes is the response. Habituation is a response decrement to repeated stimulation that cannot be attributed to sensory adaptation or motor fatigue, and it comes in a short-term form lasting minutes to hours and a long-term form lasting days or weeks (Rankin et al., 2009). Sensitization is the increment, typically following intense or noxious stimulation; the two are treated as independent processes that sum to produce the observed response (Groves & Thompson, 1970), and a response can show facilitation before it decrements when both are active (Rankin et al., 2009).
- Habituation is a response decrement to repeated stimulation that cannot be attributed to sensory adaptation or motor fatigue (Rankin et al., 2009).
- Sensitization is the increment, typically following intense or noxious stimulation, and the two can operate concurrently on the same response.
Scope: The consensus characterisation (Rankin et al., 2009) covers habituation and dishabituation only; sensitization was excluded from the 2007 workshop and rests here on the dual-process account (Groves & Thompson, 1970). The cellular account of habituation in Aplysia is contested within the field, including by a Rankin co-author, and the row treats it as the best-characterized case rather than a settled one.
Neural Substrate
The cellular evidence is non-human: Aplysia for the synaptic mechanism, rat and mouse for the mammalian circuit. Human evidence exists only at the level of lesion effects on long-term habituation. In Aplysia, short-term habituation of the gill-withdrawal reflex results from a presynaptic decrease in the number of transmitter quanta released per impulse at sensory-to-motor synapses, with postsynaptic receptor sensitivity unchanged (Castellucci & Kandel, 1974). The mammalian acoustic startle response is mediated by a short brainstem circuit in which giant neurons of the caudal pontine reticular nucleus (PnC) receive converging sensory input and project directly to motor neurons (Koch, 1999; Simons-Weidenmaier et al., 2006). Short-term habituation is intrinsic to that pathway: it is synaptic depression at the terminals of sensory afferents onto PnC neurons, upstream of multimodal integration, and it is specific to the stimulated modality both at the synapse and in behavior (Simons-Weidenmaier et al., 2006). That is the same locus identified in the invertebrate work, presynaptic and on the sensory side, though the molecular mechanism in the mammalian case remains unresolved. Sensitization is a separate modulation that raises response magnitude (Koch, 1999), with a neuronal substrate distinct from habituation's (Groves & Thompson, 1970).
In humans, patients with midline cerebellar lesions show intact short-term habituation of the startle blink within a session but impaired long-term habituation across days, indicating that the two forms depend on different substrates (Maschke et al., 2000). All of this concerns habituation of a reflex. Habituation to complex or cognitive stimuli, which is where the everyday example sits, is less well localized.
AI at a Glance
Habituation reduces responding to a stimulus that has been repeated, without sensory or motor fatigue (Rankin et al., 2009). Language models have no component that does this: the repetition penalty is a fixed configured discount rather than a learned change (Keskar et al., 2019), and many-shot jailbreaking increases responding to a new item after repeated demonstrations of the same kind (Anil et al., 2024).
- Repetition penalty: a sampling setting that discounts the score of any token that has already appeared in the generated text by a fixed factor, so it is less likely to be chosen again (Keskar et al., 2019).
- Many-shot jailbreaking: prompting a model with hundreds of fabricated examples of a behavior so that it produces that behavior on a new request it would otherwise refuse (Anil et al., 2024).
Non-Associative Learning in AI
How Far It Has Been Measured
Measurement in AI: Indirectly Measured
AI Memory Performance: Comparison Not Possible
Where It Lives in the System
Weights, Context, Memory layer, Product level. Absent at all four. Weights: fixed while the system runs, so nothing accumulates across exposures. Context: earlier turns persist within a session as content, and repeated material raises rather than lowers the probability of the demonstrated response (Anil et al., 2024). Memory layer: stores retrievable content, not response strength. Product level: the repetition penalty (Keskar et al., 2019) lowers the probability of a token once it has appeared, but it is a fixed configured discount with no accumulation and no decay, and no published work tests it against the habituation criteria.
What Has Been Measured
- No study located gives an AI system the habituation paradigm psychologists use, which checks a response decrement against the defining hallmarks (stimulus specificity, spontaneous recovery, dishabituation); nothing was found for training, the current conversation, or saved memory (stored notes an assistant can look up later), so existing work measures related effects under other names or in only one of these places (Rankin et al., 2009).
- Within the context window, meaning the text a model can see in the current conversation, GPT-2 and BART showed a sensitization-like "self-reinforcement effect": the more times a sentence had already been repeated, the more likely the model was to produce it again, a pattern studied as a text-generation flaw, not as learning (Xu et al., 2022).
- A peer-reviewed security paper tested chatbots by the name habituation: repeatedly exposing a model to a request, or escalating it in small steps, raised average jailbreak success (getting the model past its safety rules) from 0.1 to 0.5 and from 0.2 to 0.4; the abstract does not report which models were tested or how they differed, and there was no human baseline or hallmark testing (Zhang et al., 2024).
- In training, a 2026 preprint on a single small open model (1.1 billion parameters) reported "progressive adversarial sensitization": ten rounds of preference tuning raised jailbreak rates by 20%, an analogy to kindling rather than a test of the behavioral definition, with no human comparison (Cheung, 2026).
- The closest same-task comparison had 100 people and GPT-2 predict the next word of passages repeated up to four times: agreement was high on first reading (r = .87) but fell to r = .05 by the third, because GPT-2 became near-perfect after one repetition while people improved gradually; only one AI model was tested, so no spread across systems can be given (Vaidya et al., 2023).
- That comparison cannot be used for a rating: it measures memory for repeated text (closer to repetition priming) rather than a waning or growing response, and GPT-2 could still see the earlier passage while people saw words one at a time and had to remember them, so the AI's better score only favours the side the test was easier for (Vaidya et al., 2023).
- Response suppression to repeated input has been shown in AI mainly when researchers build it in: an image-recognition network given built-in suppression reproduced adaptation aftereffects seen in people, but this is sensory adaptation, which the psychology definition separates from habituation, and the component is not part of today's AI assistants (Vinken et al., 2020).
Where a Mechanism Can Live
04 Product
The whole configuration, and the only thing a benchmark observes directly.
01
Weights
Across deployments
02
Context Window
Within a session
03
Memory Layer
Across sessions
Which of the three a result belongs to is inferred from it, not seen in it.
Read more
Brainstem
Division
The stalk of midbrain, pons and medulla joining the cerebrum to the spinal cord and carrying the cranial nerve nuclei.
The brainstem runs from the diencephalon down to the foramen magnum and is conventionally divided into midbrain, pons and medulla oblongata. Every ascending and descending pathway between brain and body passes through it; ten of the twelve cranial nerves attach to it; and its tegmentum holds the reticular formation and the monoaminergic and cholinergic cell groups that project diffusely upward. The density of critical structures is why small brainstem strokes yield such distinctive syndromes, and why brainstem function is the criterion in clinical determinations of death.
Cerebellum
Cerebellum
The densely folded structure behind the brainstem holding most of the brain's neurons in a uniform, repeating circuit.
The cerebellum sits under the occipital lobes behind the pons and medulla, attached by three pairs of peduncles. It is about a tenth of brain volume yet contains roughly four-fifths of all neurons, largely because of its granule cells. Its cortex uses the same circuit everywhere, which makes it unusually tractable: input arrives as mossy and climbing fibers, output leaves through Purkinje cells to the deep nuclei. Classically a structure of motor coordination and timing, it is now widely implicated in cognition and language on imaging and lesion evidence, though what it contributes beyond motor control is actively argued.
Pontine reticular nucleus
Nucleus in the pons
The reticular formation of the pontine tegmentum, source of the reticulospinal tract and of REM sleep atonia.
The pontine reticular nucleus occupies the central tegmentum of the pons and is divided into oral and caudal parts. It gives rise to the medial (pontine) reticulospinal tract, which facilitates extensor tone, and its caudal part contains the giant neurons that mediate the acoustic startle reflex; this is one of the sites at which prepulse inhibition is standardly assumed to act. Its subcoeruleus region drives the descending inhibition that produces muscle atonia during REM sleep; loss of that inhibition is the proposed mechanism of REM sleep behavior disorder. The paramedian pontine reticular formation beside it generates horizontal saccades.
Anil et al. (2024)
Many-shot Jailbreaking
In
Advances in Neural Information Processing Systems 37, 129696-129742
Abstract
We investigate a family of simple long-context attacks on large language models: prompting with hundreds of demonstrations of undesirable behavior. This attack is newly feasible with the larger context windows recently deployed by language model providers like Google DeepMind, OpenAI and Anthropic. We find that in diverse, realistic circumstances, the effectiveness of this attack follows a power law, up to hundreds of shots. We demonstrate the success of this attack on the most widely used state-of-the-art closed-weight models, and across various tasks. Our results suggest very long contexts present a rich new attack surface for LLMs.
Find it
Castellucci & Kandel (1974)
A Quantal Analysis of the Synaptic Depression Underlying Habituation of the Gill-Withdrawal Reflex in Aplysia
In
Proceedings of the National Academy of Sciences, 71(12), 5004-5008
Abstract
Habituation, one of the simplest behavioral paradigms for studying memory, has recently been examined on the cellular level in the gill-withdrawal reflex in the mollusc Aplysia and in the escape response in cray-fish. In both cases short-term habituation involved a decrease in excitatory synaptic transmission at the synapses between the sensory neurons and their central target cells. To analyze the mechanisms of the synaptic depression in Aplysia, we applied a quantal analysis to synaptic transmission between the sensory and motor neurons of the gill-withdrawal reflex. Our results indicate that short-term habituation results from a presynaptic mechanism: a decrease in the number of...
Find it
Cheung (2026)
Kindling in Neural Systems: Progressive Adversarial Sensitization During LLM Alignment Mirrors Psychiatric Progression
Abstract
Abstract Objective Reinforcement learning from human feedback (RLHF) is widely used to make large language models safer, yet repeated preference tuning could also make them easier to breach. Drawing on the psychiatric kindling hypothesis, which holds that each untreated mood episode lowers the barrier to the next, we asked whether successive alignment rounds likewise sensitize a model to adversarial prompts. Methods A 1.1-billion-parameter chat model (TinyLlama−1.1B-Chat) equipped with LoRA adapters completed ten preference-tuning cycles. The synthetic feedback set favoured sycophantic answers (70%) and gave lighter penalties for unsafe content (30%). Three experimental arms were...
Find it
Groves & Thompson (1970)
Habituation: A dual-process theory.
In
Psychological Review, 77(5), 419-450
Abstract
Presented a dual-process theory of response plasticity to repeated stimulation. 2 hypothetical processes, 1 decremental (habituation) and 1 incremental (sensitization), are assumed to develop independently in the CNS and interact to yield the final behavioral outcome. Behavioral experiments are presented, using both the hindlimb flexion reflex of acute spinal cat and the acoustic startle response of intact rat, which are consistent with this theory. Neurophysiological experiments indicate that the 2 processes have separate and distinct neuronal substrates. The dual-process theory and other current theories of response habituation are evaluated in terms of these and other recent findings....
Find it
Keskar et al. (2019)
CTRL: A Conditional Transformer Language Model for Controllable Generation
In
arXiv
Abstract
Large-scale language models show promising text generation capabilities, but users cannot easily control particular aspects of the generated text. We release CTRL, a 1.63 billion-parameter conditional transformer language model, trained to condition on control codes that govern style, content, and task-specific behavior. Control codes were derived from structure that naturally co-occurs with raw text, preserving the advantages of unsupervised learning while providing more explicit control over text generation. These codes also allow CTRL to predict which parts of the training data are most likely given a sequence. This provides a potential method for analyzing large amounts of data via...
Find it
Koch (1999)
The neurobiology of startle
In
Progress in Neurobiology, 59(2), 107-128
Abstract
Startle is a fast response to sudden, intense stimuli and probably protects the organism from injury by a predator or by a blow. The acoustic startle response (ASR) of mammals is mediated by a relatively simple neuronal circuit located in the lower brainstem. Neurons of the caudal pontine reticular nucleus (PnC) are key elements of this primary ASR pathway. The ASR in humans and animals has a non-zero baseline, that is, the response magnitude can be increased or decreased by a variety of pathological conditions and experimental manipulations. Therefore, the ASR has been used as a behavioral tool to assess the neuronal basis of behavioral plasticity and to model neuropathological...
Find it
Maschke et al. (2000)
Involvement of the human medial cerebellum in long-term habituation of the acoustic startle response
In
Experimental Brain Research, 133(3), 359-367
Abstract
Animal studies have shown an involvement of the cerebellar vermis in long-term habituation of the acoustic startle response, but not in short-term habituation. The aim of the present study was to investigate whether short-term and long-term habituation of the acoustic startle response are impaired in patients with medial cerebellar lesions. Five patients with midline cerebellar lesions due to surgery for astrocytoma and ten healthy, age- and sex-matched subjects were studied. Subjects received 40 acoustic startle stimuli each day on five successive days. Peak amplitudes of the startle response recorded at the orbicularis oculi and the sternomastoid muscles were obtained. Data were...
Find it
Rankin et al. (2009)
Habituation revisited: An updated and revised description of the behavioral characteristics of habituation
In
Neurobiology of Learning and Memory, 92(2), 135-138
Abstract
The most commonly cited descriptions of the behavioral characteristics of habituation come from two papers published almost 40 years ago (Thompson and Spencer, 1966; Groves and Thompson, 1970). In August 2007, the authors of this review, who study habituation in a wide range of species and paradigms, met to discuss their work on habituation and to revisit and refine the characteristics of habituation. This review offers a re-evaluation of the characteristics of habituation in light of these discussions. We made substantial changes to only a few of the characteristics, usually to add new information and expand upon the description rather than to substantially alter the original point.
Find it
Simons-Weidenmaier et al. (2006)
Synaptic depression and short-term habituation are located in the sensory part of the mammalian startle pathway
In
BMC Neuroscience, 7(1)
Abstract
**Background** Short-term habituation of the startle response represents an elementary form of learning in mammals. The underlying mechanism is located within the primary startle pathway, presumably at sensory synapses on giant neurons in the caudal pontine reticular nucleus (PnC). Short trains of action potentials in sensory afferent fibers induce depression of synaptic responses in PnC giant neurons, a phenomenon that has been proposed to be the cellular correlate for short-term habituation. We address here the question whether both this synaptic depression and the short-term habituation of the startle response are localized at the presynaptic terminals of sensory afferents. If this is...
Find it
Vaidya et al. (2023)
Humans and language models diverge when predicting repeating text
In
arXiv
Abstract
Language models that are trained on the next-word prediction task have been shown to accurately model human behavior in word prediction and reading speed. In contrast with these findings, we present a scenario in which the performance of humans and LMs diverges. We collected a dataset of human next-word predictions for five stimuli that are formed by repeating spans of text. Human and GPT-2 LM predictions are strongly aligned in the first presentation of a text span, but their performance quickly diverges when memory (or in-context learning) begins to play a role. We traced the cause of this divergence to specific attention heads in a middle layer. Adding a power-law recency bias to these...
Find it
Vinken et al. (2020)
Incorporating intrinsic suppression in deep neural networks captures dynamics of adaptation in neurophysiology and perception
In
Science Advances, 6(42)
Abstract
Adding a cellular adaptation mechanism to artificial neural networks explains temporal context modulation of biological vision.
Find it
Xu et al. (2022)
Learning to Break the Loop: Analyzing and Mitigating Repetitions for Neural Text Generation
In
arXiv
Abstract
While large-scale neural language models, such as GPT2 and BART, have achieved impressive results on various text generation tasks, they tend to get stuck in undesirable sentence-level loops with maximization-based decoding algorithms (\textit{e.g.}, greedy search). This phenomenon is counter-intuitive since there are few consecutive sentence-level repetitions in human corpora (e.g., 0.02\% in Wikitext-103). To investigate the underlying reasons for generating consecutive sentence-level repetitions, we study the relationship between the probabilities of the repetitive tokens and their previous repetitions in the context. Through our quantitative experiments, we find that 1) Language...
Find it