Priming: Cross-Disciplinary Effects

• Priming is a broad concept where prior exposure biases later responses, impacting perception, language processing, machine learning, and seed treatment.
• It spans diverse methodologies from behavioral experiments and neural network conditioning to agronomic seed priming, each with domain-specific measurement protocols.
• Key applications include modulating reaction times in cognitive tasks, enhancing model adaptation and accuracy in ML, and improving germination and stress tolerance in agriculture.

Priming denotes a broad family of effects and interventions in which prior exposure, contextual cues, or preparatory computation changes subsequent processing. In psychology, it is explicitly defined as “the psychological process by which one exposure to a certain stimulus has an influence on the reaction to the exposure of another stimulus” (Mantione-Holmes et al., 2022). In response priming, a prime begins activating a response before the target arrives (Schmidt et al., 2018). In structural priming, a sentence makes a later sentence with the same structure more likely or easier to produce or process (Michaelov et al., 2023). In machine learning, priming can mean an intermediate adaptation stage, runtime conditioning of a model by task-specific clues, or retrieval of relevant pretraining data for later attunement (Huang et al., 2022). In agriculture, seed priming is a pre-germination treatment that partially hydrates the seed while preventing radicle emergence (Singh et al., 2018). Taken together, these usages suggest that priming is best understood as a general concept of preparatory influence rather than a single mechanism.

1. Conceptual scope and domain-specific meanings

Across the literature, the term retains the common idea of a prior state altering a later response, but the object being altered differs sharply by field. In psychophysics and psycholinguistics, the altered object is typically perception, response selection, or syntactic choice. In economics and political science, it is the interpretation or salience of a decision context. In machine learning and numerical algorithms, it is the parameter state, representational geometry, search space, or input-conditioned computation. In biology and agriculture, it is the physiological state of a cell or seed.

Domain	Operational meaning of priming	Representative source
Response tasks	Prime activates a response before the target	(Schmidt et al., 2018)
Structural linguistics	Prior sentence biases later syntactic structure	(Sinclair et al., 2021)
Multimodal generation	Prime biases generated syntax for a target image	(Xiao et al., 24 Feb 2025)
ML adaptation	Intermediate conditioning between pretraining and downstream adaptation	(Huang et al., 2022)
Retrieval-based adaptation	Recall relevant pretraining examples and adapt on them	(Wallingford et al., 2023)
Seed science	Partial hydration before sowing, without radicle emergence	(Singh et al., 2018)

This breadth has an important methodological consequence. The same word can denote a behavioral effect, a causal intervention, a training stage, a retrieval procedure, or a physiological treatment. A plausible implication is that definitions of priming are inseparable from the measurement protocol used in a given domain.

2. Priming in perception, action, and cognitive performance

In speeded response tasks, priming has been formalized as a time-dependent competition between response accumulators. A free-rate accumulator model extends Vorberg et al. (2003) by allowing independent prime-driven and target-driven response activation rates, denoted by $\lambda_P$ and $\lambda_T$, together with decay $v$ and threshold $c$ (Schmidt et al., 2018). The model uses two counters, $n_1(t)$ and $n_2(t)$, and responses are emitted when $n_1(t)-n_2(t)=c$ or $n_1(t)-n_2(t)=-c$. Priming is the response-time difference

$P(s)=I(s)-C(s),$

where $s$ is the prime-target SOA. The model predicts that stronger primes mainly increase priming effects in response times and error rates, whereas stronger targets mainly diminish response times and priming effects. It also predicts $\lambda_T$0 for simultaneous prime and target presentation (Schmidt et al., 2018).

Related work on crossmodal interaction treats priming as a top-down contextual manipulation rather than a masked prestimulus. In a sequence reproduction task using pitch–brightness and pitch–elevation correspondences, conceptual priming and perceptual priming were presented before the task. Conceptual priming was generally more effective than perceptual priming in enhancing crossmodal perception, and subjective evaluations of helpfulness were reported to be in contradiction with objective behavioural data (Feng et al., 2020). When two correspondences were placed in conflict, the results suggested selective integration under priming and more additive, distraction-prone integration without priming (Feng et al., 2020).

A major controversy concerns unconscious priming. A methodological reanalysis argues that the standard inference from near-chance direct-task performance and significant indirect-task congruency effects to an “indirect task advantage” is flawed because the two tasks are often compared on different statistical scales (Meyen et al., 2020). The corrected proposal is to estimate sensitivity for both tasks on the same metric, typically $\lambda_T$1, using expressions such as

$\lambda_T$2

and then test the difference

$\lambda_T$3

In a replication of the behavioral part of Dehaene et al. (1998), the congruency effect remained significant, but the corrected comparison did not support an indirect task advantage (Meyen et al., 2020). This directly limits strong claims that priming alone establishes superior unconscious sensitivity.

3. Structural priming and abstract grammatical representation

Structural priming occupies a central place in psycholinguistics because it links repeated structure use to abstract grammar. In multilingual LLMs, crosslingual structural priming has been used as a behavioral-style test for shared abstract grammatical representations. One study replicated 8 crosslingual structural priming experiments covering 6 languages and 4 monolingual structural priming experiments in 3 non-English languages, using XGLM 4.5B as the main multilingual model together with other XGLM sizes and PolyLM models (Michaelov et al., 2023). Priming was operationalized as a change in the model’s probability of generating one syntactic construction over another after a matching versus non-matching prime. The reported conclusion was evidence for abstract monolingual and crosslingual grammatical representations that function similarly to those found in humans (Michaelov et al., 2023).

A complementary line of work studies structural persistence in autoregressive Transformers without weight updates between prime and target. For prime-target pairs in dative and transitive alternations, the Priming Effect is defined as

$\lambda_T$4

Using the Prime-LM corpus of about 1.3 million prime-target pairs, positive PE was found for many Transformer models, with stronger and more symmetrical effects in GPT2-large and GPT2-xl (Sinclair et al., 2021). The same work shows that priming strength is strongly modulated by semantic similarity and lexical overlap, weakens with recency, and increases with cumulative congruent primes. It further reports that PE generally remains positive even when prime and target differ in phrase complexity, which was interpreted as evidence for some hierarchical syntactic information beyond flat sequential sensitivity (Sinclair et al., 2021).

Priming has also been turned into a representational probe. In an adaptation-based framework for LSTM LLMs, adapting on sentences of structure $\lambda_T$5 and testing on held-out sentences of structure $\lambda_T$6 yields a directed similarity quantity $\lambda_T$7, corrected for baseline surprisal as

$\lambda_T$8

Pairwise $\lambda_T$9 values reconstruct a syntactic similarity matrix, and the reported organization of relative-clause constructions was hierarchically interpretable, with passivity contributing more to similarity than reduction (Prasad et al., 2019). At the same time, the comparison to untrained baselines led to a more limited conclusion: for the class “all RCs,” similarity was no greater than in baseline models, suggesting that some apparent abstraction may still be driven mainly by lexical items rather than a fully abstract gap representation (Prasad et al., 2019).

Multimodal structural priming extends the same logic to image-conditioned generation. The PRISMATIC dataset contains 4,208 sentences paired with 1,710 aligned images across 16 distinct syntactic structures, with a test set of 1,006 annotated sentences (Xiao et al., 24 Feb 2025). Instead of fixed references, priming is measured by a reference-free tree-kernel score comparing the syntax tree of a generated sentence to the positive and negative prime structures. The reported result is that both dual-encoder and fusion-encoder multimodal models show comparable syntactic priming effects, but only fusion-encoded models exhibit robust positive correlations between priming effects and visual similarity; for Model 2, the correlation between visual similarity and PE is $v$0, $v$1 (Xiao et al., 24 Feb 2025). The paper interprets this as more aligned with human psycholinguistic findings, where similar contexts strengthen structural priming.

4. Priming as computational conditioning in machine learning

In machine learning, priming often denotes a deliberate mechanism for conditioning model computation before or during downstream inference. For cross-lingual event extraction, priming is implemented as runtime augmentation of the transformer input with the current trigger, and optionally an event-role identifier, so that the same sentence is encoded differently depending on the question being asked (Fincke et al., 2021). The method improves trigger and argument extraction in low-resource and zero-shot settings; in English$v$2Arabic, trigger classification improves from 42.4 to 51.0 F1, and argument classification from 30.2 to 32.4 (Fincke et al., 2021). The central claim is that priming lets candidate arguments be represented conditionally on the current event query rather than only on sentence context.

A related idea is cue-dependent top-down modulation of internal features. In object detection and segmentation, a cue vector is mapped to channel-wise modulation coefficients so that feature maps are transformed by

$v$3

with $v$4 (Rosenfeld et al., 2017). The paper contrasts this with pruning, where the cue affects only the decision stage. Priming early layers was found to be especially important, and under severe Gaussian noise with $v$5, priming reached 34.8% mAP while pruning reached 24.1% mAP (Rosenfeld et al., 2017). In this formulation, priming is not a post hoc filter but a cue-dependent change to the internal computation.

Another family of methods uses auxiliary task-relevant features to redirect optimization. PrimeNet introduces a priming variable

$v$6

computed from key inputs $v$7, and trains the main model as $v$8 (Wen et al., 2022). The method is designed to “create a better shortcut” so that SGD is pulled away from poor shortcuts. On NICO image classification, PrimeNet reports 71.11% in-domain accuracy and 49.00% OOD accuracy, compared with 66.11 and 42.61 for vanilla ResNet18 (Wen et al., 2022). On CARLA Nocrash-Dense, PrimeNet reports 49.3% success versus 34.1% for vanilla behavioral cloning (Wen et al., 2022).

Priming also appears as an explicit training stage. A general parameter-efficient framework inserts priming between pretraining and downstream PEFT, yielding the pipeline

$v$9

Using BART-base on the 160-task CrossFit Challenge, the reported conclusions are that priming generally improves parameter-efficient methods, multi-task learning outperforms meta-learning in this setup, and priming only the PLM is the strongest general strategy (Huang et al., 2022). This usage differs from psycholinguistic priming: the effect is not residual activation from a prior stimulus but an intermediate adaptation stage designed to improve few-shot generalization.

A retrieval-based formulation goes further by treating priming as recall of relevant pretraining examples. Neural Priming retrieves examples from the model’s own pretraining corpus using class names or unlabeled test images, filters them with the model itself, and constructs a classifier from nearest-class means mixed with the zero-shot text head (Wallingford et al., 2023). Reported gains include a 2.45% improvement on ImageNet, a 3.81% average accuracy improvement across standard transfer learning benchmarks, and a 1.41% improvement on ImageNetV2 (Wallingford et al., 2023). The method is presented as test-time re-attunement to a downstream task or shifted distribution.

In numerical linear algebra, priming can mean the initial approximate stage of a two-step algorithm. Primed-PCA first runs any approximate-PCA method to obtain a candidate low-dimensional subspace and then performs exact PCA within that subspace. The reported experimental result is an average speedup of 7.2 over Oja’s rule and 10.5 over EigenGame (Máté et al., 2021). Here priming narrows the search space rather than changing a behavioral response.

At the architecture level, Priming has also been used to convert pre-trained Transformers into Hybrid Attention + State-Space Models. The method selects Attention layers to replace with SSMs, transfers weights, aligns the hybrid with the source model, and then post-trains it, using less than 0.5% of the source model’s pre-training token budget (Chattopadhyay et al., 8 May 2026). In controlled comparisons of Gated KalmaNet, Gated DeltaNet, and Mamba-2, the expressiveness hierarchy $c$0 is reported to predict downstream long-context reasoning performance. A Hybrid GKA 32B improves over its source Qwen3-32B by +3.8 average reasoning points while enabling up to 2.3x higher decode throughput (Chattopadhyay et al., 8 May 2026). In this setting, priming is knowledge transfer from one architecture family to another.

5. Priming as contextual intervention in decisions, strategy, and interface behavior

In experimental economics, priming is used to manipulate context interpretation. In two social dilemma experiments in the US and Israel based on the Rubinstein (2006) layoff dilemma, economic cues were induced by word-search or recall tasks with terms such as “inflation,” “monopoly,” and “market,” whereas communal cues used terms such as “equality,” “solidarity,” and “fairness” (Snir et al., 2024). Economic priming made participants retain about 10.69 fewer workers in Israel and 10.27 fewer workers in the US, while communal priming led to about 11.84 more workers retained in Israel and 8.37 more in the US (Snir et al., 2024). In both experiments, the interaction between treatment and economics major was not statistically significant. The interpretation offered is that small cues shift which norm becomes active in an ambiguous setting, rather than merely revealing fixed selfish preferences (Snir et al., 2024).

In continual NLP learning, priming has been repurposed as a class-ordering principle for catastrophic forgetting. The approach maps semantic priming, associative priming, and repetition priming onto ways of ordering classes before incremental training (Mantione-Holmes et al., 2022). The best-performing method is associative interleaved priming: on RMHD it outperforms random ordering by about 15%, and on CCAT-50 by about 20%; within associative priming, interleaving outperforms block priming by about 20% on RMHD and about 10% on CCAT-50 (Mantione-Holmes et al., 2022). This usage treats priming as exposure sequencing that prepares the network for later classes with smaller parameter disruption.

Political models use the term in yet another sense: campaign spending that changes issue salience rather than perceived candidate quality. In a multi-issue, multi-party model, baseline issue salience $c$1 is shifted by campaign investment through

$c$2

and relative salience determines expected vote share (Shaki et al., 2024). For parliamentary elections, pure-strategy Nash equilibrium always exists; for two parties, there exists an equilibrium in which each party invests only in a single issue, computable in $c$3 time; in most presidential settings, no equilibrium exists (Shaki et al., 2024). Priming here is agenda-setting through salience reweighting.

Interface-security research has used priming as a non-intrusive nudge. In PassPoints-style graphical passwords, the “presentation effect” or “drawing the curtain” gradually reveals the background image from left to right or right to left over 20 seconds at password-creation time (Parish et al., 2021). In a study with 710 analyzed participants, the priming techniques did not harm usability: mean SUS was 74.68 for Control and 77.30 for Primed, with no significant difference (Parish et al., 2021). Security effects, however, were image- and direction-dependent. On the Highway image, both RTL and LTR substantially improved security compared with control, whereas on Barn, LTR improved security but RTL often hurt it (Parish et al., 2021). This shows that a priming intervention may have measurable behavioral consequences even when pointwise distribution shifts are not statistically significant.

6. Biological and agricultural priming

In plant science, seed priming is a pre-germination treatment that partially hydrates the seed, stimulates initial germination processes, but prevents radicle emergence (Singh et al., 2018). The reported effects include faster germination, better seedling vigor, more uniform emergence, improved viability, and better tolerance to biotic and abiotic stresses. In chickpea, hydropriming and chemical priming with NaCl, KNO3, and urea were compared, and biospeckle analysis was proposed as a fast, non-destructive, low-cost assessment method (Singh et al., 2018). The study reports hydropriming for 24 h as the most effective hydropriming treatment and KNO3 at 1% for 6 h among the most effective chemical treatments, with higher biospeckle index corresponding to higher germination percentage and lower mean germination time (Singh et al., 2018).

In innate immune cells, priming refers to sensitization by a low-dose pretreatment. A coarse-grained three-node dynamical model defines priming operationally as a case where the first low dose produces a small response, but the later high-dose response is at least 50% greater than the response to a single high dose (Fu et al., 2012). Systematic parameter search identified three major mechanisms for priming: pathway synergy, suppressor deactivation, and activator induction; and one for tolerance: inhibitor persistence (Fu et al., 2012). In the authors’ murine bone marrow-derived macrophage experiment, low-dose pretreatment followed by high-dose challenge produced about a 36% augmentation of IL-6 induction, whereas high-dose pretreatment caused about an 80% reduction (Fu et al., 2012). The central systems-level conclusion is that sensitization and tolerance are emergent properties of network topology and dynamics rather than of any single molecular identity.

These biological usages preserve the core preparatory logic of priming but at a different timescale. The prime is no longer a sentence or cue seen moments earlier, but a physiological pretreatment that changes the state of a seed or cell before later challenge.

7. Methodological disputes and unifying themes

A recurring issue in priming research is that the observed effect can be contaminated by factors adjacent to the intended mechanism. Structural priming studies in multilingual LLMs explicitly treat crosslingual priming as a stronger test of abstraction because it reduces lexical overlap and repeated function words within a single language (Michaelov et al., 2023). Other work shows that lexical overlap and semantic similarity can greatly increase priming, sometimes more strongly than abstract structure alone, and that anomalous but grammatical primes produce weaker or asymmetric effects (Sinclair et al., 2021). Adaptation-based probing likewise found that some apparent structural classes were no more similar in trained LSTMs than in baseline models, suggesting that lexical cues can drive part of the effect (Prasad et al., 2019). In multimodal generation, the use of a reference-free tree-kernel metric was motivated by the fact that a single fixed target sentence can be misleading when many valid structural realizations exist (Xiao et al., 24 Feb 2025).

A second methodological dispute concerns inference from significance to sensitivity. The reanalysis of unconscious priming demonstrates that a significant congruency effect in an indirect task does not by itself imply that the indirect task is more sensitive than a direct task; like-for-like sensitivity estimates are required (Meyen et al., 2020). This critique generalizes beyond masked priming. Whenever priming is inferred from changed probabilities, reduced surprisal, increased vote share, or improved model accuracy, the meaning of the effect depends on the chosen metric and baseline.

A third theme is that priming often changes interpretation rather than only signal strength. In the layoff dilemma, economic and communal primes were argued to change which norm is activated in an ambiguous context (Snir et al., 2024). In voter priming, campaign spending alters issue salience rather than beliefs about candidate competence (Shaki et al., 2024). In machine learning, many methods called priming do not claim residual activation in a cognitive sense; instead they prepare the model by altering the search space, parameter state, or available memory, as in pPCA, PEFT priming, retrieval-based adaptation, or Transformer-to-hybrid transfer (Máté et al., 2021).

Taken together, the literature supports a broad but technically precise characterization. Priming is a preparatory intervention or carry-over state that changes subsequent computation, behavior, or physiology by biasing what information is activated, weighted, retained, or searched. The common structure is temporal asymmetry: an earlier event does not merely coexist with a later one, but alters the conditions under which the later one is processed. The diversity of operationalizations shows that priming is not a unitary mechanism; it is a cross-disciplinary framework for studying how prior states shape later responses.