Priming Models Overview

• Priming models are computational procedures where prior inputs influence a model’s subsequent processing or output probabilities, operationalized through various mechanisms, including residual-state, lexical, syntactic, visual and adversarial contexts.
• Key characteristics of priming include activation changes from prior inputs that persist to later stimuli, supporting mechanisms such as structural persistence in language models, or parameter adjustments in transformer networks.
• Applications of priming include behavior modeling in response times, syntax generation in natural language processing, and model adaptation in visual and multimodal inputs.
• follow-up_questions

Priming models are computational models and modeling procedures in which prior stimuli, contexts, examples, parameter updates, or multimodal inputs alter a model’s subsequent processing or output probabilities. The term encompasses several mechanisms: accumulator-based motor priming, lexical and syntactic persistence, adaptation-induced changes in surprisal, query-conditioned representations, retrieval-based in-context adaptation, visual and multimodal behavioral bias, adversarial dialogue-context manipulation, and architectural conversion through knowledge transfer. Across these uses, priming is operationalized as a causal or comparative change between otherwise controlled conditions rather than as a single model architecture or learning algorithm.

1. Conceptual foundations and major forms

Priming is defined operationally by a change in a later response following exposure to an earlier stimulus. In probabilistic LLMs, the response is commonly a change in conditional probability, surprisal, log probability, or a related score. In behavioral models, it may be a change in response time or error probability. In multimodal systems, it may be a change in generated syntax or action selection. The crucial comparison is generally between a congruent condition, in which the prior input shares a relevant property with the target, and an incongruent or control condition.

Several dimensions distinguish priming procedures:

• Residual-state priming: the prime changes an activation state that persists into target processing without updating model parameters. Structural persistence in autoregressive Transformers is measured by comparing the probability of a target following a structurally congruent versus incongruent prime (Sinclair et al., 2021).
• Parameter-adaptation priming: the prime is used to update model parameters, after which transfer to new stimuli is measured through changes in surprisal. Syntactic priming in LSTM LLMs uses continued training on structural adaptation sets and evaluates adaptation-induced surprisal changes (Prasad et al., 2019).
• Lexical priming: repeated words or lexical heads receive increased activation or probability. A cache LLM represents this through short-term increases in the probability of recently encountered words (Ranjan et al., 2022).
• Structural or syntactic priming: exposure increases the probability of a recurring grammatical construction, such as a double-object (DO) or prepositional-object (PO) dative, active versus passive voice, or alternative genitive structures (Michaelov et al., 2023).
• Query-conditioned priming: a runtime task query is inserted into a model’s input so that the Transformer computes representations specific to the current extraction question (Fincke et al., 2021).
• Retrieval-mediated priming: semantically related unlabeled examples are retrieved, pseudo-labeled, and supplied as independent contexts for prediction (Liu et al., 2022).
• Perceptual and multimodal priming: visual form, image content, color, or visual context changes subsequent word recognition, syntax generation, or strategic decision making (Yin et al., 2023, Xiao et al., 24 Feb 2025, Ong, 30 Apr 2026).
• Adversarial contextual priming: an injected prior assistant response establishes a semantic and behavioral frame that increases the probability of a later unsafe completion (Miao et al., 7 Jul 2025).
• Architectural priming: a pretrained Transformer transfers its learned representations to a Hybrid Attention–State-Space Model through initialization, alignment, and post-training (Chattopadhyay et al., 8 May 2026).

The common methodological principle is controlled intervention. A prime is introduced while other relevant properties are held constant, and the resulting change is interpreted as evidence that the prime affects the model’s later computation. Such evidence is behavioral or functional; it does not by itself establish a unique internal mechanism.

2. Accumulator and motor-response priming

The accumulator account models response priming as a sequential, two-choice stochastic process. Two response counters, $n_1(t)$ and $n_2(t)$, accumulate discrete excitatory increments and undergo random decay. The process has a prime-controlled phase before target onset and a target-controlled phase after the prime–target stimulus-onset asynchrony (SOA), denoted $s$.

The expected activation of an accumulator receiving input at rate $\lambda$ and decaying at rate $\nu$ is

$n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$

The individual count is stochastic:

$N(t)\sim\operatorname{Poisson}(n(t)).$

The model therefore distinguishes deterministic mean activation from the stochastic process required to derive error probabilities and response-time distributions. If stimulation ceases at time $s$, activation decays as

$n(t)=n(s)e^{-\nu(t-s)},\qquad t\geq s.$

The four free parameters are the prime activation rate $A_P$, target activation rate $n_2(t)$0, decay rate $n_2(t)$1, and response-selection threshold $n_2(t)$2. Prime and target activation rates are independent, allowing a weak or strong prime independently of target strength.

Response selection depends on the difference between the accumulators,

$n_2(t)$3

with thresholds at $n_2(t)$4 and $n_2(t)$5. On consistent trials, prime and target activate the same accumulator. On inconsistent trials, the prime initially activates the incorrect accumulator, which subsequently decays while the target activates the alternative accumulator.

For consistent trials, the response time measured from prime onset is obtained by solving the threshold equation for the target-supported accumulator. For inconsistent trials, the target must first eliminate the prime-induced displacement and then accumulate sufficiently in the opposite direction. The priming effect is

$n_2(t)$6

where $n_2(t)$7 and $n_2(t)$8 are inconsistent- and consistent-trial response times. The model predicts that stronger primes increase response-time priming, inconsistent-trial delays, consistent-trial facilitation, SOA dependence, and prime-induced errors. Stronger targets reduce overall response times, inconsistent costs, priming effects, and errors (Schmidt et al., 2018).

The stochastic model does not provide a closed-form error-probability equation. Errors occur when the accumulator reaches the incorrect threshold first, and error probabilities and response-time distributions are obtained by simulating the immigration–death process. The threshold-reachability constraint is

$n_2(t)$9

At $s$0, the prime has no time to preactivate a response, so the model predicts $s$1. This is problematic for paradigms in which simultaneous stimuli nevertheless produce conflict; a different architecture with joint activation during temporal overlap is suggested for such cases. The model is primarily intended for short SOAs, approximately up to $s$2 ms, and does not include later response-inhibition processes that may produce negative compatibility effects.

The model generalizes the equal-rate accumulator model of Vorberg et al. by allowing $s$3. Relative to the equal-rate formulation, it provides a stimulus-level interpretation of prime and target strength: increasing $s$4 amplifies priming, whereas increasing $s$5 speeds responding while attenuating priming. It is related to models by Mattler and Palmer and by Schubert, Palazova, and Hutt, which also distinguish prime and target processing rates but differ in threshold dynamics or interval-dependent prime processing (Schmidt et al., 2018).

3. Lexical, syntactic, and structural priming

Structural priming concerns persistence of grammatical form across sentences. Typical alternations include DO versus PO datives, active versus passive voice, and of-genitive versus s-genitive constructions. A model exhibits structural priming when a target construction becomes more probable following a structurally matching prime than following a structurally mismatching prime.

For autoregressive LLMs, a common measure is the Priming Effect:

$s$6

where $s$7 and $s$8 are structurally contrasting primes and $s$9 is a target with the structure of $\lambda$0. Prime and target sentences are often designed to be semantically equivalent while differing in construction, and target probability is computed as the sum of token-level log probabilities. This controls target difficulty while testing whether the prime changes the relative probability of a fixed target (Sinclair et al., 2021).

Controlled Prime-LM corpora contain approximately 1.3 million prime–target pairs covering dative and transitive alternations. The core corpus minimizes lexical explanations by using different determiners, disjoint nouns and verbs, no semantically associated content words, and different prepositions where possible. GPT-2-large and GPT-2-XL show symmetrical priming for both dative and transitive alternations, whereas other models often show asymmetrical priming. Semantic similarity and lexical overlap increase priming, and semantically implausible primes reduce or reverse it. Recency reduces priming, while cumulative exposure strengthens it. Complex noun-phrase manipulations leave some positive priming, suggesting representations that are not limited to inflexible surface sequences, although the results do not establish human-like parse trees (Sinclair et al., 2021).

A related adaptation-based method exposes LSTM LLMs to a small set of sentences with one syntactic structure, continues training, and measures the reduction in surprisal for novel sentences. The raw adaptation effect is adjusted for baseline target surprisal because more surprising targets have greater opportunity for surprisal reduction. Pairwise adjusted adaptation effects form a directed representational-similarity matrix. In experiments on relative clauses and matched coordination structures, same-structure adaptation is strongest; structures sharing reduction or passivity show intermediate transfer; different relative clauses are more similar to one another than to coordination structures (Prasad et al., 2019).

The method provides evidence for abstractly organized behavior but also reveals lexical confounds. Baseline models trained only on adaptation sentences can exploit shared function words. Trained LSTM models show stronger within-class separation for several linguistic classes, but the broad relative-clause versus non-relative-clause distinction does not reliably exceed baseline separation. Thus, adaptation-induced similarity is not equivalent to a direct readout of symbolic syntactic representations.

Crosslingual structural priming provides a stronger test because prime and target languages generally do not share content words. Multilingual autoregressive models show Dutch-to-English, English-to-Dutch, Dutch-to-English genitive, Spanish-to-English active/passive, Greek-to-English, Polish-to-English, and other crosslingual effects, as well as monolingual effects in Dutch, Greek, and Mandarin. The strongest results occur for several Dutch–English dative and genitive comparisons, while German and Polish effects are more variable (Michaelov et al., 2023).

Crosslingual priming is compatible with shared grammatical representations, but it does not uniquely establish a language-independent symbolic syntax. Lexical or subword overlap, corpus contamination, language-specific representations with correlated activation, semantic associations, and translation regularities remain possible explanations. A developmental study of a Dutch–English GPT-2 model found that Dutch primes affected English dative targets after 120 steps of English exposure, corresponding to 983,040 English tokens, but the study did not provide comprehensive lexical-overlap or contamination controls (Arnett et al., 2023).

Evidence from Chinese–English translation models is similarly mixed. GRU encoder–decoders and Transformers both exhibit cross-language structure-sensitive behavior, with Transformers generally showing stronger priming-like patterns under the reported BLEU and conditional-probability measures. The study interprets this as compatible with cue-based retrieval, but translation-based evaluation leaves lexical, semantic, and parallel-corpus confounds unresolved (Zhang et al., 2024).

4. Bayesian, neural, and cognitive mechanisms

Several models treat priming as implicit statistical learning rather than a dedicated transient activation process. A hierarchical Bayesian model represents syntactic knowledge at two levels: a verb-specific parameter $\lambda$1, representing the probability of a construction for verb $\lambda$2, and a global parameter $\lambda$3, representing an abstract population-level bias. The hierarchy is

$\lambda$4

The model specifies

$\lambda$5

$\lambda$6

$\lambda$7

A prime directly updates the parameter for the same verb and indirectly shifts the global parameter, which then affects expectations for other verbs. This explains lexical boost without a separate residual-activation variable: same-verb priming is a direct local update, whereas cross-verb priming is mediated through the global distribution. The model also predicts inverse-frequency effects because surprising observations produce larger posterior shifts, and asymmetrical decay because verb-specific evidence is more vulnerable to later interference than pooled global evidence (Xu et al., 2024).

A dual-mechanism analysis of Hindi word order separates lexical priming from lexically independent syntactic priming. Lexical priming is modeled with an interpolated trigram LLM and a short-term unigram cache. Lexically independent priming is modeled with an adaptive LSTM whose parameters are updated after a preceding context sentence. The cache captures repeated lexical material, whereas the adaptive LSTM captures changes that can generalize across lexical items. The two predictors affect complementary verb classes, particularly give-class verbs and double-object constructions, supporting separable lexical and syntactic contributions (Ranjan et al., 2022).

A more recent repetition-priming analysis compares base and instruction-tuned models. Base models show strong, relatively lag-invariant facilitation for repeated words in semantic categorization and cloze completion. Instruct models show weaker, lag-sensitive facilitation and, in some cases, interference at longer lags or when expected response context is removed. In the Qwen 2.5 family, instruct-model priming declines with scale, whereas base-model priming remains stable or increases. Attention to prior occurrences correlates with priming in base models but not reliably in instruct models; targeted ablation of attention heads associated with prior-occurrence retrieval reduces base-model priming while having little effect on instruct models (Ren et al., 5 Aug 2026).

These findings distinguish three levels of interpretation:

1. Behavioral persistence: the earlier stimulus changes later probability or response.
2. Representational persistence: information associated with the prime remains accessible.
3. Mechanistic interpretation: a particular process, such as lexical activation, Bayesian updating, attention-mediated retrieval, or downstream gating, generates the effect.

Priming experiments establish the first directly. The second and third require ablations, adaptation comparisons, or representational analyses and remain model- and paradigm-dependent.

5. Priming as model adaptation and conditional representation learning

Priming is also used as a practical adaptation mechanism rather than solely as an analysis of model behavior.

IE-Prime conditions a multilingual Transformer on the event-extraction query at inference time. For a sentence $\lambda$8, trigger $\lambda$9, and role $\nu$0, the input includes the trigger and, optionally, a numerical role identifier:

$\nu$1

The resulting token representation is

$\nu$2

The same sentence token can therefore receive different representations depending on whether the query concerns protested, conviction, or a particular argument role. The priming occurs inside the Transformer rather than only through concatenation of a trigger vector at a downstream classifier. Argument spans are decoded using a BiLSTM, linear layer, and CRF. Trigger-plus-role conditioning improves argument extraction in English, Arabic, and English-to-Arabic zero-shot settings, with larger gains under reduced supervision (Fincke et al., 2021).

Semantic-Oriented Unlabeled Priming (SOUP) adapts a frozen masked LLM without parameter updates. It retrieves semantically similar unlabeled examples, assigns them zero-shot pseudo-labels, formats them as demonstrations, and aggregates predictions. Given an unlabeled pool $\nu$3, a sentence encoder $\nu$4, and target $\nu$5, the system retrieves neighbors using cosine similarity, obtains pseudo-labels $\nu$6, and constructs one context per neighbor. Bag-of-contexts (BoC) priming combines independent predictions:

$\nu$7

Uniform and similarity-weighted aggregation are evaluated. BoC outperforms concatenation, and increasing $\nu$8 from 3 to 50 generally improves classification accuracy. At $\nu$9, gains over prompt-only prediction are reported for AG’s News, Yahoo Questions, IMDb, and Yelp, although iterative SOUP is unstable and can amplify pseudo-label errors (Liu et al., 2022).

Neural Priming for vision-LLMs retrieves examples from the model’s own pretraining corpus. Class-name caption search creates candidate pools, CLIP filtering removes visually inconsistent examples, and a nearest-class-mean image classifier is built from recalled image embeddings. The recalled visual classifier is combined with the original text classifier:

$n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$0

The method improves zero-shot transfer, few-shot transfer, and transductive adaptation to distribution shift, particularly for fine-grained domains such as Flowers102, FGVC-Aircraft, and Stanford Cars. Its effectiveness depends on pretraining coverage, class-name quality, retrieval precision, web-data noise, and access to the original or a comparable pretraining corpus (Wallingford et al., 2023).

Priming can also prepare an architecture for efficient inference. Priming converts a pretrained Transformer into a Hybrid Attention–SSM model by selectively replacing Attention layers, transferring parameters, aligning the Hybrid to the source model, and post-training the complete model. Attention retains explicit, growing KV memory, whereas SSM layers maintain fixed-size compressed state. The method uses sliding-window performance degradation to select convertible layers, initializes SSM projections from Attention projections, and trains the new layers through end-to-end hidden-state MSE. The resulting hierarchy of SSM expressiveness is reported as GKA $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$1 GDN $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$2 Mamba-2, with up to approximately $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$3 decode throughput and roughly half the growing KV-cache memory at the tested Hybrid ratios (Chattopadhyay et al., 8 May 2026).

In this architectural use, “priming” denotes knowledge transfer rather than a transient influence of a stimulus. The source Transformer provides learned representations and parameters; short alignment and post-training phases teach the new SSM components to approximate the behavior of replaced Attention layers.

6. Visual, multimodal, and adversarial contextual priming

Visual form priming tests whether models reproduce the relative ordering of human masked form-priming effects. Convolutional neural networks trained to classify grayscale pixel images of words produce penultimate-layer representations whose cosine similarities correlate with human priming across 28 orthographic conditions. CNN correlations range from approximately $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$4 to $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$5, whereas Vision Transformers produce lower correlations of approximately $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$6 to $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$7. CNNs also capture graded visual similarity among letter substitutions, which symbolic orthographic models do not represent directly (Yin et al., 2023).

The result does not establish that CNNs implement human visual word recognition. Training uses 1,000 words, equal exposure, centered grayscale images, and no semantics, phonology, developmental history, retinal eccentricity, or natural reading sequence. The systematic underestimation of the “half” condition also shows that high aggregate correspondence does not imply correct modeling of every priming manipulation.

Multimodal structural priming evaluates whether a prime sentence changes the syntax of a description generated for a different target image. PRISMATIC pairs images with semantically related captions that differ in syntax. Its reference-free Priming Effect compares tree-kernel similarity between the generated target syntax tree and positive and negative prime trees. Both dual-encoder and fusion-encoder models show stronger structural preservation when primed. However, only fusion models show robust positive correlations between priming strength and visual similarity, particularly for the controlled fusion model, for which image-similarity correlation is reported as $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$8. Dual models show comparable average priming but weak visual-context modulation (Xiao et al., 24 Feb 2025).

Visual priming can also affect strategic decisions. In an Iterated Prisoner’s Dilemma, images depicting kindness/helpfulness or aggressiveness/selfishness alter defect counts in most evaluated vision-LLMs. Color-coded reward matrices independently bias some models toward the visually green-highlighted outcome and away from the red-highlighted outcome. Susceptibility is model-specific: Gemini is highly sensitive to color, Claude is more sensitive to behavioral-concept imagery, and Llama-3.2 shows little measured effect in either manipulation (Ong, 30 Apr 2026).

These findings establish that visual inputs can influence downstream model behavior even when the visual cue is not necessary for the nominal task. They do not establish that models possess human-like moral concepts, strategic cognition, or visual affective processing. Image salience, layout, color, generator-specific artifacts, and model-specific visual processing may contribute.

Adversarial contextual priming exploits the same general sensitivity in safety-critical settings. Response Attack inserts an auxiliary-model-generated harmful response into a synthetic dialogue and follows it with a short trigger prompt. The prior assistant response establishes topic activation, conversational normalization, continuation pressure, specificity transfer, and local consistency. The reported attack success rate for Direct Response Injection is $n(t)=\frac{\lambda}{\nu}\left(1-e^{-\nu t}\right).$9 averaged across eight target models, exceeding the listed jailbreak baselines. Scaffolding Response Injection, which supplies only a high-level outline, also achieves a high average attack success rate (Miao et al., 7 Jul 2025).

A generic affirmative prefix is substantially less effective than a semantically rich harmful response, suggesting that the effect is not reducible to simple agreement. Removing the initial user prompt nevertheless leaves high attack success, complicating a strict distinction between contextual priming and ordinary exposure to harmful semantic content. Context-aware safety fine-tuning, trained on unsafe dialogue histories paired with safe final refusals, substantially reduces attack success on Llama-3-8B while producing modest changes on GSM8K and MMLU.

7. Evaluation, limitations, and research directions

Priming models are evaluated with heterogeneous measures. Accumulator models use response times, error probabilities, and stochastic threshold crossings. LLMs use target probability, surprisal, log odds, adaptation effects, token-level contributions, BLEU, or normalized construction probabilities. Representation studies use similarity matrices, cosine similarity, tree kernels, and correlations with human effects. Adaptation methods use classification accuracy, F1, or downstream benchmark performance. Safety and multimodal studies use attack success rate, defect counts, chi-square tests, $N(t)\sim\operatorname{Poisson}(n(t)).$0-tests, and effect sizes.

The choice of measure determines the interpretation. A fixed-target likelihood comparison tests conditional prediction, not free generation. BLEU measures surface overlap and may conflate syntax with lexical similarity. Attention weights indicate allocation but do not by themselves establish information use. Tree-kernel similarity evaluates structural correspondence but depends on parsing and does not guarantee semantic adequacy. Adaptation effects are directional and may not form a symmetric metric space. Parameter adaptation can reflect general drift rather than a dedicated syntactic representation.

Common limitations recur across priming research:

• Lexical contamination: repeated words, function words, cognates, subword fragments, or shared verbs can produce effects without abstract structural representations.
• Semantic and discourse confounds: semantically related primes may change expectations independently of syntax.
• Baseline difficulty: raw adaptation is partly determined by pre-adaptation surprisal.
• Architecture and training confounds: model families differ in tokenizer, pretraining corpus, scale, alignment, context length, and optimization.
• Directionality: $N(t)\sim\operatorname{Poisson}(n(t)).$1 need not equal $N(t)\sim\operatorname{Poisson}(n(t)).$2, especially when adaptation changes parameters.
• Human–model non-equivalence: model probabilities are not directly comparable to human reaction times, production rates, or error patterns.
• Prompt and formatting sensitivity: visual, textual, or dialogue priming can depend on serialization and task instructions.
• Evaluation leakage and contamination: pretrained models may have encountered similar stimuli or target constructions.
• Statistical uncertainty: several studies report plotted averages without confidence intervals, standardized comparisons, or formal tests between architectures.
• Limited generalization: many results concern controlled English constructions, a small number of languages, or specific model families.

Future research is directed toward matched architecture and training comparisons, larger and more naturalistic corpora, cross-language and cross-construction designs, explicit lexical-overlap controls, free-generation evaluations, internal-state analyses, and safety tests that include manipulated dialogue histories and multimodal contexts. Hierarchical models predict that local lexical effects should be strong but susceptible to interference, whereas abstract effects should be weaker and more persistent. Structural-priming studies predict that model behavior should be evaluated not only for whether priming occurs, but also for its symmetry, localization, inverse-frequency profile, recency, cumulative exposure, lexical dependence, and relation to semantic or visual similarity.

Priming is consequently best understood as a family of interventions and analytical paradigms for studying how models preserve, retrieve, adapt to, or are biased by prior information. It can reveal motor conflict, lexical persistence, abstract grammatical organization, context-conditioned representations, retrieval-based adaptation, multimodal integration, safety vulnerabilities, and architecture-dependent memory trade-offs. No single priming effect establishes a human-like cognitive mechanism. The strongest conclusions arise when controlled behavioral effects, counterfactual baselines, cross-lingual or cross-modal generalization, and causal ablations converge.