The cueing task is a robust experimental paradigm for investigating attention. A centrally presented valid cue, correctly indicating the location of an upcoming target stimulus, leads to quicker responses than an invalid cue. A feature of this paradigm is that increasing the delay between a peripheral cue and a target reverses this effect, where responses become slower for a valid cue, a phenomenon termed inhibition of return (IOR). Using GEMS, a system that utilises genetic programming techniques, we generated potential strategies underlying the facilitation and IOR effects in the cueing paradigm. Models were generated for three experiments differing in their experimental designs, all with good fit to behavioural data. Our approach helps address current issues in the field of attention regarding how it is defined and what mechanisms underlie it. Additional benefits and limitations of this method are discussed.

Research on expertise has identified pattern recognition and selective search as key components in decision making. This article compares the processes of human experts in chess with those of DeepMind’s AlphaZero (AZ), an AI program that has achieved superhuman performance in chess, shogi, and Go by learning through self-play, knowing only the rules of the games. Like humans, AZ predominantly relies on pattern recognition and selective search, raising the possibility that it exhibits intuition. However, at a micro level, there are notable differences. In addition to being much stronger than even top grandmasters in these games, AZ conducts more extensive look-ahead search. Moreover, AZ does not utilize several types of knowledge (e.g., episodic and semantic knowledge) that human experts draw upon. Despite these differences, AZ research has significant implications for the psychology of human expertise, including enhancing human decision-making, developing better training methods, and potentially transforming the way experts understand their domain. Keywords:

From early in development, children acquiring English as their first language show a preference for words with trochaic stress (strong-weak: doctor, candy, broken) compared to words with iambic stress (weak-strong: giraffe, police, forgot). Since children develop this trochaic bias as they become more familiar with English, it could be an emergent property of the statistical learning mechanisms that allow infants to discover words embedded in their continuous input. We tested this hypothesis using simulations with the CIPAL architecture, which represents the patterns it encounters by learning a diverse collection of chunks. The model was trained with samples of English child-directed speech, then presented with 24 trochaic and 24 frequency-matched iambic words. CIPAL showed a consistent preference for the trochaic items; they were more likely to be represented with single chunks in long-term memory, they were processed using fewer chunks, and they had faster processing times compared to the iambic items. CIPAL also demonstrated an iambic bias when the phonemes in the training samples were presented in reverse order, consistent with evidence that children’s stress preferences reflect the prosodic structure of their input. These results suggest that the trochaic bias in English can emerge through an incremental chunking process, which favours trochaic sequences because they are the most efficient way of processing English input.

Artificial intelligence is increasingly able to perform cognitive tasks that were once thought to require human expertise. This raises the question: what happens to human expertise when AI systems equal or exceed expert human performance? We examine this question from the perspective of chess, a domain in which superhuman AI has existed for more than 25 years. We argue that chess provides not a single “canary in the coalmine,” but a set of canaries for how AI may affect human expertise. Chess shows that superhuman AI emerged earlier than expected, developed through multiple approaches, and can operate through mechanisms that share some functional similarities with human expertise. It also recalibrated assumptions about expert judgment, and challenged the expectation that human + AI will always outperform AI alone. Beyond performance, AI has reshaped chess training and knowledge generation, as engines now guide preparation, opaque AI outputs can be translated into human-understandable concepts, and AI-generated ideas can be taught to grandmasters. Finally, chess illustrates institutional consequences of superhuman AI, especially around cheating, governance, and the changing status of human expertise. We consider how these chess-based canaries may generalize to other domains while also emphasizing limits to generalization.

Large language models fine-tuned on human behavioural data have emerged as general-purpose cognitive proxies, but the scale this requires, and whether these models process task structure or exploit statistical shortcuts, remain open questions. We train fourteen models from 135M to 14B parameters across four architecture families on Psych-101, a dataset of 10.7 million trial-level choices from 160 experiments. In-distribution, scale barely matters. The models fall within a narrow band, as though against a ceiling, and 0.6B to 1B parameters suffice to match a 70B baseline on held-out participants. Out-of-distribution, that band opens into a markedly steeper scaling gradient, with larger models clearly advantaged in generalisation to novel task structure. To determine what information these models use, we run two diagnostics. We progressively strip four prompt channels – task instructions, experimental stimuli, outcome feedback, and choice history – across 27 experiments, and permute trial order. Masking the content of stimuli and feedback destroys 75.7

Elite athletic performance is linked to advanced cognitive functioning, yet cognition's role in motor skill acquisition among novices remains underexplored. This study examined how cognition influences learning a complex field hockey skill in participants with minimal prior experience. Forty novices (mean age = 20.04 years) completed a hockey ball control task assessed across six timepoints, three before and three after viewing a coaching video of an expert performing the task. Cognitive measures included fluid intelligence (Raven's Progressive Matrices), crystallized intelligence (Spot-the-Word), working memory (OSpan), perceptual speed (Inspection Time), and psychomotor ability (Fitts's task). Declarative and procedural knowledge were also recorded. Performance was evaluated using positional and technical scoring systems. Gains through repetition were modeled via linear regression. No overall significant improvement occurred during uncoached repetition. However, crystallized intelligence predicted individual differences in pre-coaching gains while fluid intelligence was positively associated with immediate coaching effects. Post-intervention improvements were predicted by working memory capacity and number of practice attempts. Psychomotor ability predicted gains through repetition both before and after intervention. Distinct cognitive domains support different phases of motor learning. Cognitive profiling may inform talent identification in early skill acquisition.

According to chunking theories, children discover their first words by extracting sub-sequences embedded in their continuous input. However, the mechanisms proposed in these accounts are often incompatible with data from other areas of language development. We present a new theory to connect the chunking accounts of word discovery with the broader developmental literature. We argue that (a) children build a diverse collection of chunks, including words, multi-word phrases, and sub-lexical units; (b) these chunks have different processing times determined by how often each chunk is used to recode the input; and (c) these processing times interact with short-term memory limitations and incremental processing to constrain learning. We implemented this theory as a computational modelling architecture called CIPAL (Chunk-based Incremental Processing and Learning). Across nine studies, we demonstrate that CIPAL can model word discovery in different contexts. First, we trained the model with 70 child-directed speech corpora from 15 languages. CIPAL gradually discovered words in each language, with cross-linguistic variation in performance. The model’s average processing time also improved with experience, resembling the developmental changes observed in children’s speed of processing. Second, we showed that CIPAL could simulate seven influential effects reported in statistical learning experiments with artificial languages. This included a preference for words over nonwords, part words, frequency-matched part words, phantom words, and sub-lexical units. On this basis, we argue that incremental chunking is an effective implicit statistical learning mechanism that may be central to children’s vocabulary development.

A key issue in cognitive science concerns the fundamental psychological processes that underlie the formation and retrieval of multiple types of concepts in short-term and long-term memory (STM and LTM, respectively). We propose that chunking mechanisms play an essential role and show how the CogAct computational model grounds concept learning in fundamental cognitive processes and structures (such as chunking, attention, STM and LTM). This is done in two ways. First are the in-principle demonstrations, with CogAct automatically adapting to learn a range of categories – from simple logical functions, to artificial categories, to natural raw (as opposed to natural pre-processed) concepts in the dissimilar domains of literature, chess and music. This kind of adaptive learning is difficult for most other psychological models, e.g., with cognitive models stopping at modelling artificial categories and (non-GPT) models based on deep learning requiring task-specific changes to the architecture. Secondly, we offer novel ways of designing human benchmarks for concept learning experiments and simulations accounting for subjectivity, ways to control for individual human experiences, all while keeping to real-life complex categories. We ground CogAct in simulations of subjective conceptual spaces of individual human participants, capturing humans subjective judgements in music, with the models learning from raw music score data without bootstrapping to pre-built knowledge structures. The CogAct simulations are compared to those obtained by a deep-learning model. These findings integrate concept learning and adaptation to complexity into the broader theories of cognitive psychology. Our approach may also be used in psychological applications that move away from modelling the average participant and towards capturing subjective concept space.

Automation transformed various aspects of our human civilization, revolutionizing industries and streamlining processes. In the domain of scientific inquiry, automated approaches emerged as powerful tools, holding promise for accelerating discovery, enhancing reproducibility, and overcoming the traditional impediments to scientific progress. This article evaluates the scope of automation within scientific practice and assesses recent approaches. Furthermore, it discusses different perspectives to the following questions: Where do the greatest opportunities lie for automation in scientific practice?; What are the current bottlenecks of automating scientific practice?; and What are significant ethical and practical consequences of automating scientific practice? By discussing the motivations behind automated science, analyzing the hurdles encountered, and examining its implications, this article invites researchers, policymakers, and stakeholders to navigate the rapidly evolving frontier of automated scientific practice.

Multimedia inputs have been often used in second language (L2) vocabulary learning; however, the effective elements in multimedia inputs for L2 vocabulary learning have hardly been established. This study aims to identify the effective element(s) and further clarifies the meaning of different “domains” in multimedia L2 vocabulary learning. Considering that the learning target (L2 vocabulary) belongs to the verbal domain, the meaningful inputs are then constructed as within-domain (i.e. L1-based), cross-domain (i.e. Picture-based), and mixed-domain (i.e. L1+picture-based) learning conditions. The present study firstly conducted an omnibus analysis and then three meta-analyses: (a) within-domain vs. cross-domain (20 studies, 51 effect sizes), (b) within-domain vs. mixed-domain (21 studies, 55 effect sizes), and (c) cross-domain vs. mixed-domain (9 studies, 27 effect sizes), for a total of 2056 participants. Both immediate and available delayed tests were used as dependent variables. Four moderators were used to identify potential predictors. The results indicate that, at the immediate tests, the mixed-domain condition consistently outperforms the within-domain condition (g = 0.334, p < 0.001) and cross-domain condition (g = 0.350, p = 0.001) in facilitating L2 vocabulary learning. However, at the delayed tests, the advantageous effect of the mixed-domain condition only marginally outperformed the cross-domain condition (g = 0.271, p = 0.067). No significant difference was found between the within-domain and cross-domain conditions in L2 vocabulary learning. No significant moderator was detected at either the immediate tests or the delayed tests. These findings highlight the benefit of incorporating both L1 words and pictures into L2 vocabulary learning.

Artificial Intelligence (AI) is transforming many aspects of daily life and improving efficiency and outcomes in various sectors, ranging from healthcare to education. However, this rapid integration also introduces significant anxiety among users, as it undermines their sense of autonomy, identity, meaning, and purpose. The present study examines the associations between key factors and AI anxiety, using self-reported online questionnaires completed by 221 participants. The factors examined included existential reflection, existential anxiety, curiosity and exploration, AI literacy, general anxiety, alongside sociodemographic variables such as gender, age, and job exposure to AI. The measures used were the Existential Reflection Scale (ERS), Existential Concerns Questionnaire (ECQ), Curiosity and Exploration Inventory (CEI), General Anxiety Disorder (GAD), AI Literacy Scale (AIL), and AI Anxiety Scale (AIAS). A multiple regression analysis showed that existential anxiety was the best predictor of AI anxiety, with higher existential anxiety predicting higher AI anxiety (β = .26, p = .004). Higher existential reflection also predicted higher AI anxiety (β = .13, p = .037). Notably, higher AI literacy was found to predict lower AI anxiety (β = −.18, p = .014). Age, gender, job exposure, curiosity and exploration, and general anxiety did not predict AI anxiety. The findings highlight the importance of addressing existential concerns and improving AI literacy to reduce AI anxiety. The study provides a foundation for future research, providing novel insights into AI anxiety and guiding organisations and policy makers to alleviate it.

When individuals are asked to keep in mind arbitrary sequences of items such as words, letters, numbers or images, they spatialize them in working memory forming a horizontal mental line. This study is the first meta-analysis of this phenomenon known as SPoARC (Spatial Positional Response Codes) effect or OPE (Ordinal Position Effect). For this purpose, we had access to the raw data of 21 of the 24 behavioral studies ever published on this topic. A multilevel meta-analysis was performed with participants nested within experiments, both used as levels. After confirming the existence of the SPoARC effect, we analyzed it as a function of four features: the size and nature of the memoranda, the pace of presentation of the memoranda and the type of classification of the probes. Results showed that (a) the SPoARC effect varied as a function of the nature of the memoranda, which we suggest highlights the importance of phonological processes in WM spatialization, (b) the SPoARC effect was the largest when the presentation pace was around 3 s per item or above and (c) the SPoARC effect increased when participants were asked to pay attention to the ordinal structure of the memoranda (whenever a temporal classification task is used), confirming the link between order information and WM spatialization.

Cognitive scientists often represent theories of cognitive behavior in the form of computer programs which simulate and model the performance of humans in experimental settings. Earlier work has demonstrated that evolutionary techniques, specifically genetic programming (GP), can be used to generate a pool of candidate models in the form of executable computer programs. However, previous work has not considered the impact of changes to hyper-parameter values, such as those controlling the behavior and timing of operators or those controlling the operation of the GP process. In this paper, we develop and use a cluster analysis technique based around the Silhouette index to investigate the impact of hyper-parameter changes on the composition of evolved populations of programs. Our metrics support visualizations and enable a user to assess both qualitatively and quantitatively the diversity of candidates from different populations. In this way, a cognitive scientist can analyze the output of the evolutionary system in order to uncover or inspire potentially novel theories of human behavior.

Chunking theory is among the most established theories in cognitive psychology. However, little work has been done to connect the key ideas of chunks and chunking to the neural substrate. The current study addresses this issue by investigating the convergence of a cognitive CHREST model (the computational embodiment of chunking theory) and its neuroscience-based counterpart (based on deep learning). Both models were trained from raw data to categorise novel stimuli in the real-life domains of literature and music. Despite having vastly different mechanisms and structures, both models largely converged in their predictions of classical writers and composers - in both qualitative and quantitative terms. Moreover, the use of the same chunk/engram activation mechanism for CHREST and deep learning models demonstrated functional equivalence between cognitive chunks and neural engrams. The study addresses a historical feud between symbolic/serial and subsymbolic/parallel processing approaches to modelling cognition. The findings also further bridge the gap between cognition and its neural substrate, connect the mechanisms proposed by chunking theory to the neural network modelling approach, and make further inroads towards integrating concept formation theories into a Unified Theory of Cognition (Newell, 1990).

Ce commentaire aborde trois questions (apprentissage, unification et validation) soulevées par la proposition audacieuse de Rey. Tout d’abord, il suggère d’intégrer un autre type d’apprentissage à son unification: le chunking , défini dans un sens plus large que dans son article et englobant non seulement les informations verbales et séquentielles, mais aussi les informations visuelles. Deuxièmement, il propose qu’une approche explicative à plusieurs niveaux – en particulier les niveaux cognitif et neuronal – puisse offrir une stratégie plus efficace que l’approche réductionniste de Rey. Enfin, il souligne plusieurs défis posés par le développement de modèles informatiques unifiés. La question de savoir si l’associationnisme radical unifié de Rey réussira ou si une approche explicative à plusieurs niveaux est plus prometteuse est une question fascinante, importante non seulement pour la psychologie et les neurosciences, mais aussi pour l’épistémologie.

AI-guided scientific discovery will become an indispensable tool for scientists in the not too distant future. This half-day tutorial begins by introducing the area of computational scientific discovery within the cognitive sciences along with an overview of recent models of cognitive behaviour developed using these tools. The second half of the tutorial presents a detailed walk-through of our methodology and computational system, where participants will learn how to design experiments, evolve and analyse candidate models using the GEMS (Genetically Evolving Models in Science) system. Although specialised towards the cognitive sciences, many of the principles of model definition and discovery can be more broadly applied, and so the tutorial should be of interest to the wider Cognitive Machine Intelligence community.

This study extends an existing cross-linguistic model of verb-marking error in children’s early multi-word speech (MOSAIC) by adding a novel mechanism that defaults to the most frequent form of the verb where this accounts for a high proportion of forms in the input. Our simulations show that the resulting dual-factor model not only provides a better explanation of the data on typically developing (TD) children, but also captures the cross-linguistic pattern of verb-marking error in children with Developmental Language Disorder (DLD), including the tendency of English-speaking children to show higher rates of Optional Infinitive (OI) errors and the tendency of Dutch-, German- and Spanish-speaking children to show higher rates of agreement errors. The new version of MOSAIC thus provides a unified cross-linguistic model of the pattern of verb-marking error in TD children and children with DLD.