The specialization of human face recognition for upright own-race faces is well-established. While experience is thought to play a key role in face specialization, establishing its direct causal contribution in humans is difficult, as natural experience cannot be systematically controlled. Recent advances in deep learning algorithms offer a solution: these algorithms were shown to generate human-like face specialization effects, including the face inversion and other-race effects. Critically, deep neural networks allow precise manipulation of their training experience, allowing us to test its sole contribution to human-like face expertise in artificial systems. In the present study, we systematically manipulated the amount of face experience provided to deep neural networks and examined its effect on the face inversion, the other-race and other-age effects. Mirroring human development, the magnitude of the other-group and face inversion effects increased with greater own-group upright face experience. These effects were primarily driven by a steep improvement in recognition of upright, own-group faces, with much shallower gains for other-group or inverted faces. These findings demonstrate that increased exposure to upright own-group faces selectively improves performance for this category, establishing that experience alone is sufficient to produce human-like face specialization effects in artificial systems.
A substantial recent literature has investigated readers' failure to detect transposed words in otherwise grammatical sentences. This phenomenon may reflect a process of rational inference, whereby a reader's prior over word strings overrides the bottom-up perceptual evidence. In three experiments, we tested the prediction emerging from the rational inference framework that failure to detect a transposition should be most common when the prior probability of the untransposed string is high. We found that the frequency of the trigram consisting of the two transposed words and the preceding word predicts failure to notice the transposition: Readers are more likely to miss a transposition of the final two words of a high-frequency trigram such as close the gate than a lower-frequency trigram such as close the pool. We also found a weaker role for the conditional probability of the transposed words based on the entire preceding sentence context. The results support a role for rational inference in 'correcting' errors, and also contribute further evidence that multi-word units play an important role in language processing.
This study examined visual statistical learning using EEG-based steady-state visual evoked potentials (SSVEP). Fifty-one adults were exposed to image sequences organized into triplets across three conditions (n = 17 per condition) in which the alignment of category-level and exemplar-level information was manipulated. Neural entrainment at the triplet frequency (1.11 Hz) differed significantly across conditions (η p 2 = .13), with stronger responses in the Single-Category and No-Category conditions than in the Mixed-Category condition. There were no differences at the image frequency (3.33 Hz; η p 2 = .05). Behavioral reaction times mirrored this pattern, showing faster responses to the last exemplar in the triplet in the Single-Category (η p 2 = .71) and No-Category (η p 2 = .22) conditions, but not in the Mixed-Category (η p 2 = .10) condition. Both signal-to-noise ratio (SNR) and inter-trial coherence (ITC) captured neural entrainment across fronto-central and parietal-occipital electrode clusters. These findings validate SSVEP as an online measure of visual statistical learning and demonstrate that category-exemplar mismatch interfered with statistical learning.
Can you perform an action intentionally if there was only a small chance your action would succeed? Researchers studying the folk concept of intentional action have documented complex patterns of intuitions in these cases. We propose to explain the data in terms of a causal theory of intentional action, according to which people consider an action intentional if there was a strong causal link between the agent's desire for the outcome and the outcome itself. When the outcome of an action is left to chance, people judge that the agent's desire was a weak cause of the outcome; this results in corresponding low intentionality judgment. Supporting our proposal, we find that in situations where people deny that a lucky outcome was brought about intentionally, they also deny that the outcome happened because of the agent's desire for that outcome. Intentionality judgments track causal judgments to a larger extent than they track other factors previously proposed to explain the luck effect. Our findings support the idea that a robust causal link between desire and outcome is a key component of the folk concept of intentional action.
Phonesthemes are sound clusters that recur in words with related meanings. For example, glow, gleam, and glitter begin with the phonestheme /gl/. Experiments with pseudowords provide evidence that phonesthemes affect how people retrieve and represent meanings, which suggests that phonesthemes capture language evolution in action, namely the emergence of compositionality. However, there is little evidence that phonesthemes affect how people process familiar words with established meanings. We therefore investigated whether phonesthemes facilitate visual word recognition and affect semantic representations in English. In two lexical decision experiments, people recognized words that contain phonesthemes faster and more accurately than control words, and phonesthemes amplified N170 ERPs; and in four lexical decision megastudies, phonesthemes predicted faster visual word recognition (but not faster auditory recognition). In two semantic relatedness experiments, we found only limited evidence that phonesthemes affect interpretation: Decisions were slower and less accurate for words containing incongruent phonesthemes (e.g., /gl/ misled people about the meaning of glove), but only in the online experiment, with a marginal effect of accuracy in the lab-based experiment and an amplified N400 ERP in frontal sites (but not the expected centroparietal sites). Our findings demonstrate that people process words which contain phonesthemes differently than they process controls, consistent with the hypothesis that phonesthemes inhabit a middle ground between morphemes and meaningless strings of letters, but our findings also demonstrate that, unsurprisingly, these quasi-meaningful constituents have only small effects in how people represent the meanings of familiar words. We discuss implications for morphology, language evolution, and natural language processing.
When someone acts generously, what expectations does this create? Classic theories and experiments emphasize reciprocity: the idea that generosity will be returned by the recipient. Yet daily-life and ethnographic observations often show a different pattern: generosity sets a precedent, leading people to expect the same person to be generous again. Across six online behavioral experiments (total N = 599 U.S. adults) using third-party vignette judgments and first-person incentivized economic games, we test when generosity creates expectations of reciprocation versus a precedent. We found that people expected reciprocity only in equal or symmetric relationships. Otherwise, they expected generosity to continue from the same actor. These expectations generalized across roles, contexts, and cost structures. Taken together, the results suggest that classic evidence on reciprocity and turn-taking may capture expectations among strangers or equals, but not the wider set of relationships that structure much of social life.
Vague quantifiers like many, few, and several may vary considerably with respect to the quantity they denote. Depending on the context, many may indicate different quantities in "many students" versus "many cups of coffee." The vagueness and context sensitivity of such quantifiers pose a challenge for semantic theories that aim to formally characterize quantifier meaning. We address this challenge by extending and experimentally testing a Bayesian model proposed by Schöller and Franke (2017), which represents quantifiers as cumulative density thresholds over probability distributions of expected values. We hypothesized that each quantifier has a stable semantic threshold, with contextual variability arising from differences in expected value distributions. To test this, we conducted two experiments: one eliciting contextual expectations, and another collecting cardinality judgments for quantified utterances. We then fit five hierarchical Bayesian models and used model comparison (via WAIC and DIC) to evaluate whether thresholds generalize across contexts. Our results reveal conflicting evidence. While estimated thresholds are highly overlapping across contexts-suggesting some stability-models with individualized thresholds consistently outperform the context-stable alternative. Moreover, semantically motivated bounds do appear more stable than pragmatically motivated ones, as expected.
Plato's Republic, Einstein's Theory of relativity, Vivaldi's Four Seasons are all remarkable examples of humans' unique ability to create and manipulate complex abstract structures, whether in language, mathematics, or music. Yet the mechanisms by which children develop such abstract thinking, and the role of education and structured experiences such as musical practice in shaping these abilities remains unclear. To explore these questions, we conducted cross-sectional behavioral experiments with 528 children aged 4 to 8, spanning four educational grades, half of whom participated in a violin training program, from age four. Two experiments examined how children encode, process, and compress auditory sequences and visual patterns, while a third examined their sensitivity to geometric regularities. Our results reveal the emergence of symbolic reasoning as early as the start of formal schooling, with deeper abstraction as a function of grade. By first grade, children encoded complex auditory sequences within a Language of Thought (LoT) similar to adults. Additionally, when confronted with quadrilaterals, children showed increasing sensitivity to geometric regularities, suggesting a developmental transition from perceptual to symbolic reasoning. However, we did not observe significant impact of musical practice on abstraction abilities across any of the domains tested. We discuss whether and how the impact of education and extracurricular activities such as music could be enhanced.
Bilingualism is a complex experience that may enhance executive functions (EF). This potential benefit could be particularly relevant for autistic children, given that they tend to present EF difficulties. Some studies find bilingualism benefits in attention, inhibitory control, cognitive flexibility, and memory, in neurotypical and autistic children while others do not. These inconsistencies may arise from oversimplifying the operationalization of bilingualism, obscuring the mechanisms underlying its effects on EF. This study used continuous measures to operationalize bilingualism and test its links with EF. Bayesian multilevel modeling showed that higher second language proficiency in 168 autistic (M = 8;3) and 262 neurotypical (M = 7;8) children was associated with better attention and working memory in neurotypical children and better attention, short-term memory, working memory, and shifting abilities in autistic children.
French number words provide a unique window into the relationship between numerical cognition and language, because numbers above 60 follow a vigesimal (base-20) word structure (e.g., 72 = "60-12"). In a two-digit magnitude comparison task with sixty French native speakers, we replicated the classic unit-decade compatibility effect (UDCE; slower responses when unit and decade comparisons conflict) and within-decade effect (faster responses when decades are identical), reflecting the place-value structure of Arabic numerals. Given the French vigesimal system, we expected not only the classic UDCE and within-decade effect but also their vigesimal counterparts driven by magnitudes of number words: a unit-vigintade compatibility effect (UVCE) and a within-vigintade effect, in which pairs sharing the same decade word (e.g., "soixante" for the 60s and 70s) are processed faster than other between-decade pairs. Linear mixed models revealed both a UDCE for numbers larger than 60 and a UVCE, indicating that number words were accessed during processing. Participants also responded faster to within-vigintade items (86 vs. 95) than to between-vigintade items (76 vs. 85) and as fast as to within-decade items (82 vs. 85), indicating a verbal equivalent of the within-decade effect. This effect is unaffected by decade distance and can only be explained by access to number words so that the decades were identical ("80-6" vs. "80-15"). Overall, our data indicate that verbal representations can shape basic numerical judgments and that number processing may be more closely tied to language than previously assumed.
Semantic roles, namely the agent ('doer') and patient ('undergoer') roles, are fundamental to language acquisition, as they enable learners to map meaning onto syntactic structure. Grammars typically impose a binary classification on how these roles map onto syntactic functions. These functions are encoded through features such as agreement, case marking and word order, which children gradually acquire through exposure to their linguistic environment. Notably, child-directed speech constitutes a primary source of linguistic input during acquisition (Hart & Risley, 1995; Weisleder & Fernald, 2013). At present, little is known about the effects of child-directed speech on the learning of semantic roles and even less in languages with diverse grammatical features. Here, we investigate whether child-directed speech facilitates the learning of semantic roles, specifically examining whether it enhances semantic role interpretation compared to adult-directed speech. We examine English and Russian, two languages, which differ fundamentally in how they encode semantic roles, thereby presenting distinct challenges for the language-learning child. We use artificial neural language models to analyse the statistical properties of naturalistic child-directed and adult-directed speech, testing which register more effectively facilitates semantic role learning. In Study 1, we examine whether semantic roles are more easily classified in naturalistic utterances from child-directed speech than adult-directed speech. In Study 2, we test which register better supports learning and generalising semantic roles by evaluating language models trained on either register on the same controlled test set. Study 1 shows that semantic roles are more easily classified in child-directed speech than adult-directed speech, with a more pronounced effect in Russian than in English. This suggests that child-directed speech may be optimised more strongly in a language where semantic roles are expressed in more varied forms and positions, as is the case in Russian. Study 2 shows that the knowledge of semantic roles can be generalised by the models to structures that do not frequently occur in either child-directed speech or adult-directed speech, and that, on the whole, this is more successful based on input from child-directed speech than adult-directed speech in both languages. Our results provide first evidence that child-directed speech is tailored to the language-specific needs of children, facilitating the acquisition of semantic roles, which are a prerequisite for the acquisition of syntax. These findings show that child-directed speech actively supports the acquisition of of semantic roles, helping children map meaning onto syntactic structure.
This paper reexamines a recent claim that Large Language Models lag behind humans in language comprehension on what were described as minimally complex statements. We argue that human performance was overestimated and LM performance, underestimated. Moreover, both people and lower-performing LMs are disproportionately challenged by queries involving potentially appropriate inferences, suggesting shared pragmatic sensitivity rather than model-specific deficits. Analysis of more sensitive log probabilities of Llama-2-70B demonstrate ceiling-level accuracy and pragmatic sensitivity. A separate group of LM grammaticality judgments previously characterized as incorrect are shown to correlate with human judgments, while certain reasoning models approximate idealized judgments when prompted to respond as an expert generative syntactician. Overall, the findings suggest that apparent deficits in LM performance may reflect task design, evaluation choices, and assumptions about human performance, rather than deficiencies in current models.
Humans adjust to the frequency of conflicting stimuli so that the detrimental behavioral effects of frequent conflicts become smaller than those of rare conflicts. These adjustments become contingent on the locations where frequent and infrequent conflicts have been encountered. Experimentally such phenomena are studied by means of conflict tasks, a prominent one being the Simon task of the present experiments. Our Simon task involved four stimulus locations (upper-left, upper-right, lower-left, lower-right). Conflict frequency was manipulated for two diagonally opposite locations. For example, conflict frequency was low at the upper-left location and high at the lower-right location. At the remaining (non-manipulated) locations conflict frequency was intermediate. In Experiment 1 participants responded to stimulus colors by pressing a left or right key, in Experiment 2 by pressing a lower or upper key. We observed larger conflict effects for the low-conflict location than for the high-conflict location. Crucially, these adjustments to conflict frequency transferred to non-manipulated locations depending on the response configuration: transfer regions were the left or right hemifield with left-right responses, but the lower or upper hemifield with lower-upper responses. Assuming that transfer regions around manipulated stimulus locations are defined by identical (or similar) spatial codes, the pattern of transfer suggests a process of weighted two-dimensional location coding. According to this notion, the spatial dimension that is relevant for response discrimination has stronger weight in the coding of stimulus locations than a response-irrelevant dimension, and can therefore produce anisotropic transfer of conflict-frequency effects around manipulated locations.
Humans generally posit that contrary mental states are unlikely to co-exist within a singlemind. We tested the early ontogeny of this assumption in two domains: action andcommunication. Studies 1A and 1B tested whether 9-month-old infants assume that agents actcoherently. Infants watched interactions between two hands whose owner(s) were invisible. Inthe contrary goals condition, the hand performed contrary actions-one hand reached for anobject while the other impeded it. Later, during test trials, infants learned that the handsbelonged to one or two people. Looking-time patterns across the contrary goals and a baselineconditions indicated that clear goal conflict led infants to infer two agents, suggesting theyviewed it as unlikely for a single person to thwart their own goal. Study 2 tested whetherinfants assume communicative coherence, testing whether they assume that a single informantis unlikely to entertain and communicate conflicting information while two informants mightdo so. Informants pointed to indicate a toy's location to 15-month-olds. When two differentinformants each pointed to a different place, infants did not follow one pointing gesture morethan the other. However, when a single informant pointed successively to two locations,infants followed the second gesture, implying they viewed it as an updated, not contradictory,message. Thus, infants assumed that a single informant is unlikely to contradict themselves(i.e., by asserting that a toy is simultaneously in two locations). These findings reveal anearly-emerging assumption of psychological coherence in infants'representation of otherminds, across both action and communication contexts
The Gricean model of communication assumes that cooperation is a precondition for successful communication but humans often use language non-cooperatively. How do language users calibrate cooperation when deciphering communicated content? The current work probes how belief alignment shapes the interpretation of underinformative statements attributed to politicians (Donald Trump or Kamala Harris) presented to self-identified Republican and Democratic participants. The results show that communicated content is more likely to be derived when beliefs align between voter group and speaker. This suggests that we may arrive at different conclusions from the same statement, depending on who the speaker is and how much trust we grant them.
Iconicity has increasingly come to be recognized as widespread in language. It plays a particularly important role in bootstrapping new referring expressions in existing languages and in the emergence of new languages and other communication systems. The basis of this role has long been assumed to depend primarily on transparency of the iconic signal for the receiver, who benefits from being better able to identify its meaning. But might there also be producer-side advantages, distinct from this transparency-based comprehension benefit, that support communication? We investigated this using an experimental referential communication game in which dyads used a novel signaling medium to communicate fruits and vegetables. We manipulated both whether the producer could generate iconic signals and whether the receiver saw iconic or arbitrary signals. Results suggested that there was an iconicity benefit via stability in production even if the receiver was unable to perceive the iconicity (and therefore unable to benefit from transparency). However, while this gave dyads a substantial head-start, the lack of a benefit from iconicity for the receiver meant dyads in this condition still performed significantly less well overall than dyads with full iconicity.
The nature of eye movements during visual search has been widely studied in cognitive science. Virtual reality (VR) paradigms are an opportunity to test whether computational models of search can predict naturalistic search behavior. However, existing ideal observer models are constrained by strong assumptions about the structure of the world, rendering them impractical for modeling the complexity of environments which can be studied in VR. To address these limitations, we modeled immersive visual search as a reinforcement learning problem, in which sequential decisions are made over a multidimensional representation of the environment learned by a convolutional neural network. In our formulation, RL agents learned a policy over latent states-effectively solving what is known as a meta-Markov decision process (meta-MDP), where each decision concerns how to allocate attention to information in the environment. Training deep-RL agents on the meta-MDP showed that learned (i.e., optimal) search policies converge to a classic ideal-observer model of search developed for simple (1D) stimuli. We compared the learned resource-rational policy with human gaze data from a visual-search experiment conducted in VR and found qualitative and quantitative alignment between model predictions and human behavior. However, both the model's simulated performance and its correspondence with human behavior depended strongly on the representational features available to the policy. These results suggest that naturalistic visual search behavior can partially be explained by resource-rational allocation of limited cognitive resources, and the choice of representation influences the degree of alignment between model and human behavior.
Adults often hold different goals for children's achievement: Sometimes they want a child to learn and develop their skills as much as possible (i.e., a learning goal), while other times they may forego a child's learning in favor of successful performance (i.e., a performance goal). How do children think these achievement goals influence adults' child-directed behaviors? Across two preregistered experiments (n = 90 adults; n = 160 5- to 8-year-old children), we found that children systematically predict that an adult would select a more difficult task for a recipient child when the adult held a learning (vs. performance) goal, and when the recipient was more (vs. less) competent. Importantly, we found that this pattern matched adults' actual task choices, although adults showed more sensitivity to choosing a task that anchors closely to what a child can reasonably learn from or accomplish. These results suggest children can reason about how adult's achievement goals manifest into observable actions, which may have consequences for children's own goal orientations and task selections.
Generative AI tools are increasingly being used for creative and academic work. How do people morally evaluate plagiarism involving AI-generated content, and do they judge it differently than when the source is a human? Investigating these questions can provide insight into why people condemn plagiarism; for instance, whether this is due to harm to the original creator or the false benefit gained by the plagiarizer. We examined people's moral evaluations of plagiarism involving AI-generated content in five experiments (N = 1705). In each experiment, participants read scenarios about a poet submitting someone else's poem to a contest without credit. We compared three source types: a friend, ChatGPT, and a little-known poetry blog. In Experiments 1-3, participants judged plagiarism from the blog as more immoral than plagiarism from a friend or ChatGPT, with little difference between the latter two. Moral condemnation increased with the amount of content copied and remained stable when compared to other moral transgressions. In Experiments 4 and 5, moral judgments became harsher when human sources (friend or blog) denied permission, but not when ChatGPT did, suggesting that its refusal was not treated as morally meaningful. When all sources granted permission, differences between conditions disappeared. Overall, these findings support both the harm and false benefit accounts of why people condemn plagiarism. The findings also advance knowledge about how, and when, permission from the source affects condemnation of plagiarism.
As demand for problem-solving skills continues to grow, so does the need for efficient training methods. A potential method of achieving efficient training of problem-solving skills is to design a training programme that maximises transfer of learning between problem domains. However, current models of problem-solving make contrasting predictions about the conditions under which cross-domain transfer is possible. While pattern-recognition models using state-action associations allow only limited transfer to tasks with different actions, heuristic search-based models using state-value estimates can reuse learning across tasks, potentially achieving greater transfer. In this study, we consider the case of transformation problem-solving tasks where each task uses a distinct and non-overlapping set of transformation rules. Participants trained using tasks from one taskset over four consecutive days and completed pre- and post-training probes on the untrained set. Training improved participants’ solution rate, efficiency, and decision speed, but there were no reliable improvements for the untrained tasks beyond what could be explained by direct practice during the probe blocks. There were no consistent trends in participants’ self-reported strategies across sessions, but use of a strategy involving explicit deliberation over which actions to take was consistently associated with better decisions. We note that the absence of transfer contrasts with the predictions of the heuristic search account of problem-solving in which a learned rule-independent state-value heuristic is reused across tasksets to decide actions. We discuss potential reasons for this absence of transfer and its possible implications for future models of problem-solving and for the training of problem-solving skills.
