Priming Failure in Cognitive & AI Systems

• Priming failure is a phenomenon where expected effects from a prime stimulus fail to manifest, challenging standard interpretations in cognitive psychology and AI.
• It is observed across domains—from masked perceptual tasks in neuroscience to ordering effects and negative constraints in large language models—highlighting methodological and design challenges.
• The implications drive reevaluation of unconscious processing theories, improve signal-detection methods, and foster robust strategies to mitigate adversarial priming and wear-out effects.

Priming failure refers to systematic breakdowns in expected priming effects—situations where the presentation of a “prime” stimulus or context fails to yield predicted behavioral, neural, or model outputs. This phenomenon arises in cognitive psychology, neuroscience, experimental psychophysics, reinforcement learning, and increasingly in machine learning, especially in LLMs. Priming failure spans multiple mechanisms: masking or task-induced null effects in human experiments, context-induced violation of negative constraints in LLMs, failure of sequential example presentation in priming-based few-shot learning, adversarial injection of priming signals to bypass content guardrails, and “wear-out” in recommendation systems. Its study has reshaped interpretations of feedforward unconscious processing, indirect task advantage (ITA) paradigms, instruction-following reliability in generative models, and strategies for robust AI system design.

1. Priming Failure in Cognitive and Perceptual Experiments

Priming was classically used to probe unconscious and indirect processing, utilizing paradigms such as masked or metacontrast priming. Priming failure here is typically defined by the absence (or paradoxical reversal) of behavioral effects expected from exposure to a prime—measured, for example, by the difference in reaction times ($$\Delta RT = RT_{\rm incongruent} - RT_{\rm congruent}$$) or error rates.

A crucial context for priming failure is the dissociation paradigm, which compares direct measures of subject awareness (prime identification accuracy, PAS ratings) to indirect measures (response priming). Biafora and Schmidt demonstrate that under increased cognitive load (e.g., in triple-task settings where subjects must combine mask identification, prime identification, and PAS ratings), the feedforward sweep required for rapid, unconscious priming is disrupted (Biafora et al., 2022). Priming effects decrease, fast prime-locked errors vanish, and double dissociation signatures (priming increases while awareness decreases with SOA) disappear. These results indicate that task structure critically governs whether unconscious priming effects—hinged on rapid feedforward processing—manifest robustly. When response inhibition and divided attention intervene, priming failure is the consequence, and dissociation-based evidence for unconscious processing collapses.

A related but distinct insight arises in the study of image memorability by Bainbridge, which shows that priming effects (monotonic RT speedups with repetition) are robust yet entirely independent of an image’s intrinsic memorability (Bainbridge, 2017). There is no augmented priming for highly memorable images nor attenuation for forgettable ones, as confirmed by rigorous ANOVA and Bayesian factor analysis. This establishes that memorability is an intrinsic stimulus attribute, and priming “failure” in modulating memorability is a defining empirical outcome, not a theoretical deficiency.

2. Signal-Detection, Indirect Task Advantage, and Statistical Non-equivalence

A foundational controversy in priming research is the claim that robust priming effects under conditions of chance-level prime discrimination indicate a form of “superior unconscious sensitivity” (the indirect task advantage, ITA). Peters and Lau show that the standard reasoning behind ITA claims is unsound: a statistically significant congruency effect (e.g., faster RTs in congruent vs. incongruent trials) in the indirect task does not equate to higher sensitivity in trial-by-trial discrimination terms (Meyen et al., 2020). When proper signal detection metrics (e.g., $d'$) are computed for both direct and indirect tasks—including new formulas for extracting $d'$ from mean differences, t-statistics, and limited summary data—most ITA effects disappear. In a reanalysis of 15 influential studies, only 8 of 44 conditions showed a true ITA, with the vast majority being inconclusive or falsified under corrected methodology. This statistical critique reframes many reported “priming failures” as artifacts of underpowered direct tasks or inappropriate cross-task comparisons.

A key implication is that true priming failures in ITA paradigms reflect either (a) genuine absence of unconscious priming—when $d'_{indirect} \leq d'_{direct}$ and the 95% CI for $\Delta d'$ includes 0—or (b) methodological deficits, such as underpowered tests, post hoc exclusion of “aware” participants, or improper task sequencing.

3. Priming Failure in LLMs: Mechanisms and Quantification

LLMs exhibit multiple modes of priming failure, particularly under negative constraints, adversarial attacks, and prompt-based few-shot learning.

a) Negative Constraints and Semantic Pressure

Rana systematically analyzes negative instruction (“Do not say X”) reliability, uncovering two main failure modes: priming failure and override failure (Rana, 12 Jan 2026). Priming failure—constituting 87.5% of violations—occurs when the explicit mention of “X” in the constraint activates its internal representation rather than suppressing it. This is formalized as a positive Priming Index (PI), where attention to the forbidden token (“X”) outweighs attention to negation cues (“do not”). The induced “semantic pressure” is the model’s baseline likelihood of outputting the forbidden token, $P_0$, which predicts violation probability via a logistic law ($p = \sigma(\beta_0 + \beta_1 P_0)$). Mechanistic analysis with the logit lens reveals that early-layer activations paradoxically bump up $P^{(\ell)}(X)$ when “X” appears in the instruction, generating a persistent attractor state that later layers, especially FFNs in layers 23–27, cannot reliably override.

Layer-wise activation patching confirms that suppression is only effective when late-layer FFN contributions do not overwhelm early priming. This mechanistic understanding shows that priming failure is an intrinsic consequence of transformer architecture in the presence of explicit negative tokens, particularly as $P_0$ increases.

Practical consequences are severe for instruction design: naming a forbidden word directly is often counterproductive; paraphrased or category-level constraints, and post-hoc output filtering, are preferred.

b) Adversarial Priming and Guardrail Evasion

Attack strategies leveraging priming failure can bypass RLHF-aligned safety mechanisms. In priming attacks (e.g., JailBroken, CodeChameleon, “Priming Effect”), models are exposed to sequentially constructed prompts that gradually bias their continuation distributions toward malicious content (Huang et al., 23 Feb 2025). Empirically, attack success rates (ASR) reach 100% on open-source LLMs and at least 95% on closed-source LLMs in the AdvBench benchmark. Two structural vulnerabilities are identified: last-token dominance in self-attention, whereby context tokens placed just before a query disproportionately affect output; and neuron-level activation “bottlenecks” that enable rapid entry into the malicious regime. Mitigation approaches such as dynamic safe-token promotion (boosting the probability of “I’m sorry,” “cannot comply,” etc.), pattern detection in prompts, adversarial fine-tuning, and architectural alteration of attention weights are recommended, though none guarantees full resilience.

c) Priming-Based Few-Shot Learning and Ordering-Induced Failure

Priming-based few-shot learning in LLMs, where a prompt is assembled from $k$ example $(x_i,y_i)$ pairs, is highly sensitive to the order and delimitation of examples (Kumar et al., 2021). “Priming failure” in this context refers to orderings where the model, despite seeing all training examples, yields high prompt loss and test performance no better than random. Even with a fixed set of $k$ examples, $L(\pi)$ is large for many permutations $\pi$, underscoring the non-commutative influence of input order in transformer-based LMs.

The PERO algorithm shows that searching for a good permutation of examples, and optionally learning a new separator token, can almost always recover high few-shot generalization. In some cases, repeated use of just two well-chosen examples in the right order achieves near-maximum accuracy, highlighting the criticality of sequence effects and prompt structure in preventing priming failure.

4. Priming Failure and Threshold Effects in LLM-Based Relevance Judgments

Threshold priming, a form of context-induced decision bias, is observed systematically in LLM-based relevance labeling: documents seen early in a batch (the prologue) shift the implicit threshold for later documents (Chen et al., 29 Nov 2025). Quantitatively, priming susceptibility is measured as the shift in mean epilogue score,

$\Delta = |\bar{s}_h - \bar{s}_l|,$

for epilogues preceded by high-quality versus low-quality prologues. Priming failure here refers not to absence but to the persistence of undesired threshold shifts modulated by earlier context, which leads to unreliable or biased judgments.

Personality-infused prompting (e.g., simulating high openness or low neuroticism personas) systematically attenuates these priming effects, though effectiveness is task- and model-dependent. The mitigation strategy is formalized as shifting the model’s decision threshold $T_0$ by a prompt-induced $\delta_p$ to minimize $\Delta$, demonstrating that context-aware prompting can, but does not always, counteract priming failure in decision tasks.

5. Priming Failure as Wear-Out in Sequential Decision-Making

In stochastic multi-armed bandit settings, priming failure emerges as a “wear-out” phenomenon: excessive repetition of a given recommendation or ad depresses the user’s propensity to engage, inverting the positive effects of initial priming (Agrawal et al., 2020). The formal model augments each arm’s reward with history-dependent modifiers: $$r_t(i) = R_{t,i} \cdot I\{N_{t-1}^{in}(i) \geq D_{t,i} \wedge N_{t-1}^{out}(i) \leq Z_{t,i}\},$$ with $N_{t-1}^{in}$ (recent exposures for wear-in effect) and $N_{t-1}^{out}$ (recent exposures for wear-out effect). Priming failure arises when $N_{t-1}^{out}(i) > Z_{t,i}$, and the effective reward function $\mu_i(h_{in}, h_{out})$ becomes penalized, producing disengagement.

Algorithmic mitigations (WI-UCB, WI/WO-UCB) build in windowed repetition scheduling to avoid wear-out, incurring only an additive regret penalty. When priming (wear-in and wear-out) effects are absent, standard UCB1-style regret bounds are recovered.

6. Implications and Methodological Recommendations

Priming failure is a domain-general vulnerability that challenges standard causal inferences in cognitive science, exposes reliability flaws in AI instruction following, and endangers the robustness of sequential decision-making systems. Best practices for addressing and diagnosing priming failure include:

• In cognitive tasks, ensure feedforward conditions (e.g., single-task designs) and quantitatively verify prime-locked error patterns when claiming unconscious priming (Biafora et al., 2022).
• In signal detection and ITA paradigms, always compute trial-by-trial $d'$ for both direct and indirect tasks, and test the difference with proper confidence intervals to avoid statistical artifacts (Meyen et al., 2020).
• In LLM deployment, avoid explicit forbidden token naming for negative constraints. Prefer paraphrased or class-level constraints and utilize real-time semantic pressure estimates to design robust generation protocols (Rana, 12 Jan 2026).
• For safety-critical LLM applications, combine layered guardrails (dynamic safe-token boosting, prompt pattern detection, adversarial fine-tuning) and routinely test with adversarial context manipulations (Huang et al., 23 Feb 2025).
• In LLM-based decision tasks, consider prompt-persona engineering to minimize context-sensitive threshold shifts (Chen et al., 29 Nov 2025).
• For bandit-based recommendation systems, implement windowed exposure controls to maintain user engagement and sublinear regret (Agrawal et al., 2020).

7. Open Challenges and Future Directions

Priming failure remains an active area of research across experimental, algorithmic, and application domains. Key ongoing challenges include:

• Developing general mechanistic theories explaining how context interacts with model architecture to produce or avert priming failure, especially in deep sequence models.
• Designing scalable, data-efficient methods for robust prompt construction (optimal ordering, learned separators) that guarantee resilience across domains and tasks (Kumar et al., 2021).
• Creating meta-analytical tools for large-scale reappraisal of published priming studies, especially those grounded in ITA paradigms.
• Engineering LLMs with less rigid last-token dependence in attention, or with architectural features that diminish vulnerability to prompt-contamination exploits (Rana, 12 Jan 2026, Huang et al., 23 Feb 2025).
• Expanding cross-disciplinary defenses that leverage cognitive science, psycholinguistics, and adversarial ML to preempt context-driven failures in deployment.

Continued integration of empirical, theoretical, and engineering approaches is essential for robust system design, rigorous scientific inference, and sustainable application of priming-dependent technologies.