RESEARCH DOSSIER
Theory of Mind
Abstract. Theory of mind concerns the representation and inference of other agents' beliefs, knowledge, desires, intentions, and perspectives. This fundamental research monograph analyzes the origins of the field, explicit false-belief reasoning, infant belief tracking, developmental dependencies, competing theoretical frameworks, neural systems, comparative animal research, psychometric and clinical interpretation, and artificial-language-model evaluations. Its critical narrative synthesis examines original experiments, meta-analyses, systematic measurement reviews, and replication findings, distinguishing observed task performance from claims about underlying representations. The evidence supports robust developmental change in explicit belief reasoning, identifies uncertainty in implicit measurement and cross-population generalization, and motivates controlled investigation of belief attribution in human and artificial agents.
Keywords: theory of mind; mentalizing; false-belief understanding; belief attribution; social cognition; developmental psychology; implicit theory of mind; mental-state inference; comparative cognition; temporoparietal junction; construct validity; artificial cognition
Introduction: The Scientific Problem of Other Minds
Theory of mind denotes the capacity to represent and reason about mental states attributed to oneself or other agents, including beliefs, knowledge, ignorance, desires, intentions, and perspectives. The term names a family of psychological competencies that make behavior intelligible by reference to what an agent takes to be the case and what that agent seeks to accomplish. Its central explanatory challenge arises whenever an observer knows a fact that another agent has not encountered. A scientific account must specify how agents encode distinct perspectives, update them through perception and communication, and use those representations to anticipate action. The influential formulation by Premack and Woodruff posed this problem through experiments with chimpanzees, making comparative cognition foundational to the field from its beginning Premack & Woodruff (1978).
The canonical empirical question concerns representational divergence. An observer sees an object moved while a protagonist is absent; successful belief reasoning predicts that the protagonist will search where the object was previously located. This task structure forces a distinction between the current physical world and the protagonist's informational relation to it. However, answering correctly can also require remembering the initial location, comprehending instructions, suppressing a reality-based response, and inferring the experimenter's pragmatic intention. A measured performance therefore reflects both mental-state competence and the demands of the chosen procedure. Wimmer and Perner's pioneering experimental work made false belief a powerful operational target while also initiating the long debate over what an observed success truly establishes Wimmer & Perner (1983).
The present research dossier treats theory of mind as a multi-component, task-sensitive and socially embedded capacity. It reviews foundational experiments, developmental trajectories, theoretical interpretations, neural correlates, comparative studies, individual-difference measurement, and computational demonstrations. The evidence is drawn from identifiable original studies, meta-analyses, methodological reviews, and contemporary artificial-language-model evaluations. The scientific synthesis evaluates competing explanations by asking which observational signatures they predict, which alternative processes they exclude, and how far findings generalize across populations and task formats. The analysis remains a critical scholarly synthesis of existing literature rather than a newly collected experimental investigation Wellman et al. (2001) Yeung et al. (2024).
Research on other minds has immediate implications for communication, teaching, cooperative action, deception, conflict interpretation, and the evaluation of autonomous conversational systems. These applications require greater precision than a single label can provide. Detecting an emotion from a photograph, predicting a reach from a familiar preference, tracking a belief after an unseen displacement, and recursively representing what one agent thinks another believes are distinct inferential achievements. The conceptual architecture used throughout this monograph separates the content represented, the computational operation employed, the social function served, and the evidence offered by a particular test. This separation supports comparisons across human development, nonhuman animals, and artificial systems without presupposing identical underlying mechanisms Apperly & Butterfill (2009) Strachan et al. (2024).
Historical Foundations and the False-Belief Revolution
The modern research program crystallized around Premack and Woodruff's 1978 question about chimpanzees. They presented an adult chimpanzee with filmed human problem situations and asked the animal to select pictures depicting appropriate solutions. The findings were interpreted as potentially consistent with attribution of the actor's purposes, but the inferential route was open to alternatives based on learned regularities and perceptual problem solving. This historical episode established a methodological tension that persists: a correct behavioral prediction can be compatible with mental-state attribution while remaining explicable through nonmentalistic processes. The term theory of mind thus entered experimental psychology as both an explanatory proposal and a challenge to design discriminating controls Premack & Woodruff (1978) Call & Tomasello (2008).
Wimmer and Perner introduced experimentally controlled stories in which a protagonist acted on an outdated representation of an object's location. Their 1983 study situated belief within the broader problem of deception and second-order representation. The resulting paradigm shifted scientific attention toward understanding that someone else's belief can conflict with reality and with the observer's knowledge. The decisive observation was not simply whether a child knew the object's location but whether that child could answer a question about the protagonist's likely search. The original study is historically significant because it provided an explicit representational criterion, while modern interpretation must still consider verbal comprehension, memory, and inhibitory demands Wimmer & Perner (1983) Wellman et al. (2001).
The paradigm was then applied to clinical-developmental questions. Baron-Cohen, Leslie, and Frith compared autistic children with comparison groups using a puppet-play version of the false-belief task. Their 1985 report advanced a specific hypothesis about metarepresentational development and autism, stimulating decades of research on heterogeneity in social cognition. The paper's historical importance is substantial; its broad original interpretation is no substitute for contemporary assessment of diverse autistic individuals, communication contexts, language abilities, and task properties. Treating one experimental contrast as a universal diagnostic description would extend far beyond its evidentiary scope Baron-Cohen et al. (1985) Yeung et al. (2024).
The historical arc continued through standardized developmental scales, infant nonverbal paradigms, neuroscience, and comparative work. Wellman, Cross, and Watson synthesized 178 studies of children's false-belief performance, documenting a strong developmental pattern accompanied by identifiable task variation. Later research with infants reported belief-sensitive gaze and expectation measures earlier than conventional elicited-response success, transforming a formerly linear developmental story into a dispute over competence, expression, and interpretation. In parallel, neuroimaging associated mental-state stories with temporoparietal and medial frontal systems, while ape studies reopened the original comparative question. The field's progress therefore consists partly in increasingly refined experiments and partly in improved awareness of inferential ambiguity Wellman et al. (2001) Onishi & Baillargeon (2005) Saxe & Kanwisher (2003).
Beliefs, Desires, Knowledge, and Representational Content
The structure of ordinary social explanation often joins two variables: an agent's representation of a situation and the outcomes that agent values. A belief characterizes what the agent takes to be true; a desire characterizes a preferred state of affairs; an intention links selected action to a prospective goal. This belief–desire formulation enables predictions that can diverge from observable reality. A person may sincerely search an empty container because they believe an object remains there. The psychological question concerns the representation attributed by an observer and the manner in which it changes with new evidence, not the observer's assessment of the belief's objective accuracy Wimmer & Perner (1983) Apperly & Butterfill (2009).
Knowledge requires special care because ordinary language treats it differently from belief. If a proposition is known, its truth is ordinarily presupposed, whereas a belief may be mistaken. In experimental tasks, however, responses about who knows what can be affected by perspective, source monitoring, vocabulary, and assumptions about the speaker. A scientifically useful design distinguishes information access from propositional truth and asks whether subjects can maintain these distinctions under conflict. Recent model evaluations illustrate the danger of collapsing factual accuracy, first-person belief, and third-person belief attribution into a single score; a system can answer familiar third-person scenarios successfully while misclassifying epistemic status in altered contexts Suzgun et al. (2025).
Desires, goals, and visual access are commonly easier to infer than false belief because action direction and perceptual opportunity may offer relatively direct cues. The developmental scale proposed by Wellman and Liu treated diverse desires, diverse beliefs, knowledge access, false belief, and hidden emotion as related but separable accomplishments. Their ordering was established with task-specific evidence rather than a definition requiring every child to develop in precisely the same sequence. This approach advances theory by making mental-state understanding a family of graded achievements rather than the binary possession of a unitary faculty. It also demands that any claim of a single general ability be supported by correlations and longitudinal structure Wellman & Liu (2004).
Recursive mental states add an additional representational requirement. To understand that one person believes another person intends to mislead a third party, an observer must keep separate agents, propositions, and embedded perspectives. Task accuracy may depend on memory capacity, syntactic embedding, and pragmatic context as much as on a putative dedicated mentalizing operation. Experimental analysis should therefore manipulate recursion depth separately from narrative length, number of characters, and linguistic difficulty. Artificial evaluations make this design principle particularly visible: models that reproduce answers to standard false-belief stories can fail when the same epistemic relations are paraphrased or embedded in less familiar contexts Kosinski (2024) Suzgun et al. (2025).
Construct Boundaries: Mentalizing, Empathy, and Social Cognition
Mentalizing is the practice of interpreting conduct in terms of mental states, while theory of mind commonly denotes the competence or representational machinery recruited by that practice. Empathy introduces partly different phenomena. Cognitive empathy overlaps with inferring another person's feelings or perspective; affective empathy concerns sharing or responding emotionally to another's condition. A participant can correctly infer a false belief without sharing the protagonist's emotion, and can respond compassionately without representing a false proposition. Conflating these constructs creates interpretive errors when different tasks are assigned a single score. Meta-analytic analyses of adult measures show that the popular Eyes Test often relates more strongly to emotion recognition than to other theory-of-mind tasks Kittel et al. (2022).
Social cognition is broader still, encompassing person perception, trait inference, attitudes, norms, communicative expectations, and representations of social structure. Theory of mind occupies the domain concerning agents' informational and motivational states. Perceptual discrimination, stereotype use, memory for biographical facts, and ordinary prediction from behavioral regularity can influence performance without constituting the target inference. Construct boundaries become especially important in natural interaction where multiple cues are jointly available. A convincing mentalizing experiment asks whether the participant's response changes when another person's mental access changes while physical and behavioral cues are otherwise held comparable Heyes (2014) Yeung et al. (2024).
Perspective taking similarly includes distinguishable operations. Spatial perspective concerns what an observer can see from a particular vantage point; conceptual perspective concerns what a person knows, believes, or assumes. A task inviting visual rotation may measure spatial computation even when its narrative depicts a social agent. Conversely, some belief judgments require no imagined viewpoint transformation and depend on memory of informational events. The behavioral overlap between these processes warrants study, but the overlap alone cannot establish identity of mechanisms. A central methodological obligation is to distinguish the explanatory target from the cognitive resources used in expressing it Apperly & Butterfill (2009) Heyes (2014).
These distinctions have consequences for research involving social difficulties. Low scores on a single laboratory measure do not reveal whether the source is belief reasoning, language, visual perception, anxiety, executive control, unfamiliar social conventions, or low measurement precision. Clinical groups may differ in several of these dimensions simultaneously. The appropriate scientific unit is therefore a profile of performances across validated measures with attention to context and within-group variation. The same logic applies to claims about artificial agents: conversational success on emotion-labeling prompts and success on counterfactual belief tracking must be evaluated as separable accomplishments Baron-Cohen et al. (1985) Yeung et al. (2024) Strachan et al. (2024).
Research Designs, Inference, and Evidence Appraisal
A research design for theory of mind begins by specifying the exact target relation among agent, proposition, informational event, and anticipated action. In a changed-location task, the dependent measure may be a verbal answer, gaze direction, anticipatory eye movement, looking duration, or helping behavior. Each outcome indexes different stages between stimulus processing and response. Strong designs include true-belief controls, matched memory and comprehension conditions, counterbalanced event sequences, and independent coding of behavior. Their purpose is to determine whether sensitivity to another person's mental access explains variance beyond perceptual tracking and general executive demands. Classic and contemporary experiments differ substantially in these controls Onishi & Baillargeon (2005) Southgate et al. (2007) Dörrenberg et al. (2018).
Evidence from a controlled experiment supports a conditional inference tied to its stimuli, population, task, and outcome. A replicated difference in looking time establishes discriminative expectation under that procedure; interpreting the difference as explicit propositional belief representation requires further converging evidence. Negative results can reflect measurement noise, inadequate familiarization, or the absence of a proposed process. Both positive and negative findings must be analyzed with respect to preregistration, statistical precision, exclusion rules, developmental appropriateness, and the feasibility of alternative explanations. Systematic replication studies of implicit measures illustrate why a headline finding and a stable measurement instrument are different accomplishments Kulke et al. (2018) Dörrenberg et al. (2018).
Meta-analyses synthesize variation, but their conclusions depend on study selection, coding of task families, and the comparability of outcomes. The influential false-belief meta-analysis by Wellman and colleagues reported orderly age-related performance shifts across many studies and task variants. Subsequent meta-analyses tied performance to language and executive function, emphasizing the need to model shared cognitive demands. Such associations are informative about developmental organization without independently deciding whether one process causes another. Longitudinal precedence, experimental intervention, and plausible mechanistic pathways are required for stronger causal claims Wellman et al. (2001) Milligan et al. (2007) Devine & Hughes (2014).
The present monograph uses a critical narrative synthesis anchored in directly identifiable papers and reviews. It does not claim an exhaustive database search, pooled effect estimate, preregistered protocol, or formal risk-of-bias assessment. Primary studies are interpreted within their reported designs; published reviews inform the pattern across studies; competing hypotheses are distinguished through their testable predictions. Particular attention is given to construct validity, task impurity, replication, culture, and cross-species equivalence. In the Artificial Era, these same principles extend to the evaluation of generated textual responses: success on a benchmark is an observable result that must be connected to a mechanism only through additional controls Yeung et al. (2024) Suzgun et al. (2025).
Infant Belief Tracking and the Interpretation of Looking
Studies of infants expanded the chronology of social cognition beyond the age range assessed by standard verbal questions. Onishi and Baillargeon presented 15-month-olds with an actor who hid a toy and then experienced either a true-belief or false-belief situation concerning its location. Their violation-of-expectation procedure measured how long infants looked after the actor searched in different locations. The reported pattern was interpreted as early sensitivity to the actor's representation rather than exclusive tracking of the toy's actual position. Because the measure concerned looking rather than verbal explanation, the experiment opened a fundamental distinction between detecting an anomalous event and explicitly reporting another person's belief Onishi & Baillargeon (2005).
Southgate, Senju, and Csibra subsequently used eye tracking to examine whether 25-month-old infants anticipated where an actor would search in a false-belief situation. The reported anticipatory gaze toward the belief-consistent location strengthened the case that some young children show belief-sensitive behavior without giving an elicited verbal response. A different task family used looking-time effects to probe whether another agent's belief influences expectations even when it conflicts with the observer's information. These paradigms were designed to reduce instruction demands, yet they introduced dependencies on attention allocation, event memory, gaze calibration, and assumptions about action sequences. The measured result remains a relation between the experimental manipulation and infant behavior Southgate et al. (2007) Kovács et al. (2010).
The interpretive controversy concerns whether these early responses require representing a false proposition. A child might track the agent's last encounter with an object, encode an agent–object association, or learn the statistical regularities of search events without the full flexibility associated with later explicit reasoning. Such explanations are scientifically valuable when they generate distinctive predictions rather than merely redescribe results after the fact. The stronger test varies the physical setting, social agent, event history, and belief–action relation while controlling for perceptual novelty. Finding transfer across these manipulations would restrict low-level alternatives; inconsistent transfer requires a more limited conclusion Heyes (2014) Apperly & Butterfill (2009).
Replications complicated early optimistic interpretations. Dörrenberg, Rakoczy, and Liszkowski examined several nonverbal measures and reported limited evidence in some violation-of-expectation conditions but failures and weak convergence for other procedures. Kulke and colleagues also reported difficulty replicating standard implicit mentalizing effects in a systematic study. These results do not erase the original observations; they establish that a reliable, general infant false-belief competence cannot be inferred from an isolated paradigm. A cumulative account must explain why particular measures work under some conditions and fail under others, preferably through preregistered cross-laboratory work that separates age, task, and measurement factors Dörrenberg et al. (2018) Kulke et al. (2018).
Preschool False-Belief Understanding and Conceptual Change
The development of explicit false-belief reasoning is one of the most robust regularities in the classic literature. In changed-location and deceptive-container tasks, younger preschoolers often answer according to present reality, whereas older preschoolers increasingly predict actions from a protagonist's outdated information. Wellman, Cross, and Watson analyzed 178 studies and found that age-related improvement persisted across national settings and many task modifications. Their model accounted for substantial variation through age, country, and task properties. The evidence supports genuine developmental organization, while the precise timing of any individual child's success depends on the linguistic, social, and practical properties of the assessment Wellman et al. (2001).
The significance of false-belief success lies in coordinating two representations that contradict each other. An observer can know that a toy is in a new location while attributing a different location belief to a protagonist who missed the transfer. This explicit contrast becomes crucial for understanding mistakes, deception, teaching, and communication in everyday life. Yet a correct response to a single story is not a complete theory of all mental-state understanding. Children can reason about desires, perception, and knowledge access before solving some standard belief tasks, and they can continue developing understanding of irony, second-order beliefs, and mixed emotions after false-belief performance becomes reliable Wellman & Liu (2004).
Task design shapes what the child must demonstrate. A deceptively simple question such as where a character will search may invite a child to state the true location, particularly if conversational conventions normally reward correct statements about reality. Changes in wording, salience, and memory support may alter performance. These improvements can reveal previously masked understanding, but they do not automatically show that age-related differences are entirely artifacts. The meta-analytic record provides evidence for persistent development alongside procedural sensitivity, making an either-or formulation between conceptual growth and performance demands scientifically inadequate Wellman et al. (2001) Wimmer & Perner (1983).
Developmental interpretation becomes stronger when investigators combine independent task types, assess performance at several ages, and test within-person transitions. Repeated measures can determine whether new competence appears across distinct contexts or only in a familiar story format. Item-level analysis identifies children who pass some belief tasks but not others, allowing graded models to compete with abrupt stage claims. The distinction between prediction, explanation, and justification is also informative: a child who selects the correct location may struggle to state the agent's informational history, while another child may explain the protagonist's error in a conversational narrative. Such dissociations should inform theory instead of being removed as noise Wellman & Liu (2004) Milligan et al. (2007).
Language, Executive Function, and Developmental Dependencies
Language contributes multiple resources to belief reasoning. Children need to understand verbs of thinking and knowing, temporal changes, embedded clauses, reference to absent events, and questions that request someone else's perspective. Milligan, Astington, and Dack conducted a meta-analysis of language ability and false-belief performance, documenting a substantial relation across the literature. This association invites several causal explanations: language may help represent propositions, provide conversational experience with perspectives, reduce task comprehension costs, or develop jointly with a broader representational capacity. Because the relevant studies include different linguistic measures, the interpretation cannot be reduced to vocabulary alone Milligan et al. (2007).
Executive function contributes the capacity to maintain contextual information, resist a dominant reality-based response, and update behavior when task rules change. Devine and Hughes synthesized 102 studies involving 9,994 preschool-age participants and identified a moderate association between executive function and false-belief understanding. Their report also described longitudinal asymmetry in which early executive-function differences predicted later false-belief variation more consistently than the reverse direction. The evidence motivates developmental dependency models, but prediction across time does not isolate one neural mechanism or prove that executive control by itself constitutes mental-state representation Devine & Hughes (2014).
The two domains interact in experimental performance. A child listening to a long false-belief story must parse the protagonist's informational history, hold two object locations, inhibit the recently observed reality, and select an answer matching the question. Lowering one demand can reveal the contribution of others. To determine whether a task targets theory of mind specifically, researchers can match verbal complexity across mental and physical stories, include executive-function controls, and examine latent factors spanning multiple measures. Cross-task correlations should be evaluated after considering measurement reliability and shared method variance Devine & Hughes (2014) Yeung et al. (2024).
Conversation offers a plausible developmental context because children regularly hear conflicting descriptions, explanations of desire, corrections of mistaken beliefs, and stories about motives. Such experiences can enrich the ability to represent perspectives and the linguistic tools used to articulate them. The causal claim that particular conversational practices produce particular components of theory of mind, however, requires intervention designs with appropriate comparison conditions and durable follow-up. Educational recommendations must specify the target skill and measured outcome rather than assuming that every discussion of feelings trains false-belief understanding. Developmental competence is organized through reciprocal relations among cognition, language, experience, and social opportunity Milligan et al. (2007) Wellman & Liu (2004).
Later Childhood, Adults, and Individual Differences
Theory of mind continues to develop after preschool success on first-order false-belief tasks. Older children encounter narratives involving nested beliefs, strategic deception, ambiguous communicative intent, irony, social norms, and the consequences of revealing information. These contexts require integration across multiple agents, changing time points, and pragmatic knowledge. The degree of difficulty depends on language, working memory, and familiarity with social conventions, which means that a higher-order task may measure an intricate mixture of capacities. Research on later development benefits from analyzing specific inferential errors and explanations rather than counting a sequence of correct answers as a single global trait Apperly & Butterfill (2009) Strachan et al. (2024).
Adult performance frequently reaches ceiling on simplified tasks designed for preschoolers. Investigators therefore use more demanding stories, indirect speech acts, moving shapes, emotion judgments, or social faux pas assessments. Yet difficult items do not necessarily improve measurement of one coherent construct. An individual may excel at interpreting irony while making errors in recursive belief stories, and each task may depend on cultural conventions or specialized knowledge. Consequently, claims about adult individual differences should be tied to a clearly specified mental-state operation and supported by multiple convergent instruments rather than one popular test Yeung et al. (2024) Kittel et al. (2022).
Yeung, Apperly, and Devine reviewed 273 studies and identified 75 measures used for adult individual differences in theory of mind. Their review highlighted ceiling effects, incomplete reliability and validity information, and inconsistent associations among tests. These findings challenge broad interpretations of an individual's score as a stable amount of social understanding. The construct's multidimensionality may be genuine, while some inconsistency reflects poor measurement. Distinguishing these possibilities requires latent-variable designs, adequate reliability, measurement invariance, and behavioral criteria outside the task used to define the construct Yeung et al. (2024).
Performance is also situated. An individual may infer a close friend's intentions more accurately than those of a culturally unfamiliar stranger and may use different cues in direct conversation than in a written story. Real-world mentalizing involves reciprocal adjustment: people ask questions, receive corrections, recognize uncertainty, and revise assumptions. Laboratory tasks often remove these feedback loops to create experimental control. A mature research program therefore combines clean mechanistic paradigms with ecologically valid social interactions, measuring calibration and repair as well as categorical accuracy. The resulting account of adulthood should explain flexible use of evidence and uncertainty rather than a fixed, context-free score Heyes (2014) Yeung et al. (2024).
Cultural Development and the Limits of Generalization
Comparative developmental work must distinguish widespread human capacities from the local expression of those capacities in culturally patterned tasks. Children grow up in environments differing in conversational practices, schooling, household organization, narrative traditions, and expectations about describing thoughts. A false-belief story can be familiar and intuitive in one setting while requiring unusual forms of explicit explanation in another. Performance differences may therefore reflect both developmental variation and task–culture fit. The cross-national findings in the classic false-belief meta-analysis support a broad developmental pattern while preserving meaningful contributions from country and procedure Wellman et al. (2001).
The theory-of-mind scale developed by Wellman and Liu provided a way to compare a sequence of mental-state understandings instead of a single threshold. Its logic is particularly useful cross-culturally: researchers can ask whether children achieve similar component skills in different orders, and whether linguistic or contextual modifications change measured progression. Replicating the sequence in diverse populations strengthens generalization; finding variation can reveal which parts depend on social practice. A psychometrically defensible comparison requires translated materials to preserve the intended mental-state relations rather than merely equivalent words Wellman & Liu (2004).
Sampling determines the reach of a conclusion. Urban, formally educated preschoolers in research-rich countries cannot stand as an unqualified proxy for every human developmental environment. Investigators should document language, community context, schooling, socioeconomic conditions, recruitment, and prior familiarity with story-based testing. Translation should be accompanied by comprehension checks and locally appropriate materials. Comparisons also benefit from multiple tasks that differ in reliance on pretense, verbal narratives, and visual displays. A coherent cross-cultural finding is one that survives legitimate alternative specifications or is qualified by explicitly identified contextual conditions Wellman et al. (2001) Devine & Hughes (2014).
Cultural variation has conceptual consequences for the Artificial Era. A language model evaluated exclusively in English through Western laboratory stories may reproduce the conventions of its training and benchmark environment more readily than it handles unfamiliar social epistemologies. Cross-linguistic generalization, culturally varied pragmatic norms, and interaction with diverse users should therefore become part of artificial social-inference testing. This extension follows directly from the construct-validity requirement that successful performance generalize beyond the same stimulus distribution. Cultural sensitivity in scientific design increases the precision of claims about human capacities and machine-produced predictions alike Strachan et al. (2024) Suzgun et al. (2025).
Theory-Theory, Simulation, and Explanatory Alternatives
Theory-theory approaches propose that social understanding depends on structured knowledge about how mental states are caused and how they guide action. The core hypothesis is explanatory: children and adults use general principles about perception, belief, desire, and behavior, revising these principles as experience accumulates. This framework makes sense of the preschool transition from reality-based answers to predictions that honor the protagonist's mistaken representation. Its empirical challenge is to specify the contents of the proposed knowledge and demonstrate that these contents, rather than task-specific strategies, account for transfer across situations. Developmental patterns from false-belief studies and mental-state scales supply important constraints on such models Wellman et al. (2001) Wellman & Liu (2004).
Simulation-based accounts give a greater role to exploiting one's own cognitive processes when predicting someone else's behavior. An observer can provisionally adopt a target's perspective, imagine what the target would expect, and use that simulated state to forecast action. Such procedures may be useful in rich narrative and interpersonal circumstances. However, simulation requires an explanation of how the observer brackets their own privileged knowledge and why systematic errors occur when a target differs substantially from the observer. Empirical tests must distinguish simulation-specific predictions from those of a rule-based model or a learned association. A verbal report that a participant imagined being someone else does not by itself establish the computation that produced a correct choice Apperly & Butterfill (2009) Heyes (2014).
A competing family of explanations emphasizes domain-general learning, associative tracking, attentional orienting, or ordinary behavioral prediction. These approaches challenge inferences from rapid eye movements and action anticipation by specifying how nonmentalistic mechanisms may imitate belief-sensitive responses in controlled scenes. Their strength lies in requiring researchers to identify evidence that cannot be produced by simpler processes. Their weakness arises if they are stated so broadly that every possible result is declared compatible with them. Theory discrimination improves when researchers preregister contrasting predictions, manipulate informational access independently of salient physical events, and evaluate transfer to novel agents and settings Heyes (2014) Dörrenberg et al. (2018).
Theoretical pluralism reflects a genuine diversity of target phenomena. A rapid expectation about a familiar reach may arise through processes different from deliberate reasoning about a complex lie, even when both are described in mental-state language. A scientific framework must identify common structure without treating all social prediction as one mechanism. The proper comparison asks how particular models represent agents and epistemic relations, what resources their operations require, how they learn, and which failures they predict. False-belief experiments, reaction-time studies, longitudinal designs, and neural data can constrain these dimensions jointly, provided the observations are mapped to explicitly defined model commitments Apperly & Butterfill (2009) Schurz et al. (2014).
Two-Systems Proposals and the Implicit–Explicit Debate
Apperly and Butterfill advanced a two-systems proposal responding to the contrast between early implicit-looking results and later explicit belief reasoning. Their framework distinguishes relatively efficient tracking of belief-like relations from a more flexible but cognitively demanding capacity for propositional belief reasoning. The account attempts to explain why limited success appears in infancy while older children and adults show broader representational flexibility. Its scientific value depends on identifying the constraints of the proposed efficient system and demonstrating dissociations that cannot be reduced to differing response formats. A theory of two mechanisms must predict patterns of success and failure in matched tasks Apperly & Butterfill (2009).
Implicit measures frequently rely on eye movements, looking time, or interference effects obtained when participants receive no explicit instruction to reason about beliefs. These methods can reveal sensitivity to information that participants do not articulate. Yet the label implicit describes aspects of the measurement situation rather than proving that the underlying computation is automatic, unconscious, or specifically mentalistic. An implicit effect may reflect selective attention, memory for agent–object pairings, or learned event expectations. The validity of the inference depends on showing that the dependent variable follows the agent's represented informational state rather than simpler stimulus properties Heyes (2014) Kulke et al. (2018).
Replication studies have provided essential tests of the generality claimed for implicit belief tracking. Kulke and colleagues systematically examined influential experimental procedures and reported difficulties obtaining reliable effects. Work by Dörrenberg and colleagues found limited evidence in selected conditions and little convergence across multiple measures. These outcomes weigh against assuming a single robust implicit competence on the basis of a handful of successful demonstrations. They also motivate better methods rather than an immediate conclusion that early social cognition has no belief-relevant organization. A future theory should accommodate heterogeneity among procedures and specify conditions under which distinct measures converge Kulke et al. (2018) Dörrenberg et al. (2018).
A productive interpretation keeps competence, accessibility, and expression analytically separate. A person may represent a protagonist's informational history yet fail a verbal response because of language or inhibitory load; another may show correct anticipatory gaze through a shortcut with no evidence for flexible belief manipulation. Researchers can test these alternatives by crossing task difficulty with belief content, using independent response channels, and assessing individual-level reliability. Genuine double dissociations across tasks would provide stronger evidence for separable systems than labels assigned retrospectively to positive or negative findings. The debate advances when theories commit to testable computational and developmental profiles Apperly & Butterfill (2009) Devine & Hughes (2014).
Submentalizing, Behavioral Strategies, and Causal Controls
The submentalizing critique questions whether tasks interpreted as automatic attribution of mental states sometimes recruit ordinary attentional or perceptual processes. Heyes argued that reported sensitivity to what another agent sees or believes can emerge from domain-general mechanisms that are activated by social displays. This argument does not deny the existence of explicit social reasoning. It demands a more exact comparison between a mentalistic explanation and an alternative capable of generating the same response pattern. The methodological priority is to manipulate beliefs while matching the stimuli, attentional cues, and learned action contingencies available to a nonmentalistic system Heyes (2014).
For instance, if participants look toward the location last associated with a character, their behavior might appear belief-consistent even when they have encoded only a spatial relation. A counterfactual control can vary whether the character observed the displacement while preserving the physical sequence for the participant. Further controls can exchange actors, introduce unfamiliar containers, change incentives, and ask for explanations of novel outcomes. A model that genuinely represents epistemic access should use the event that changed access, while a learned association may fail under carefully designed transfer. Such discriminations are more informative than repeated success on closely related stories Onishi & Baillargeon (2005) Dörrenberg et al. (2018).
Behavioral strategies also exist in verbally administered tasks. Participants may learn the conventional answer to a Sally–Anne narrative or infer that the experimenter's question requires naming the outdated location. This problem is acute in repeated testing and in systems trained on widely circulated benchmark examples. Researchers should include adversarially constructed items, true-belief cases, controls for recency and salience, and tests where the correct response cannot be selected by superficial wording. The presence of a language output makes it possible to inspect consistency across paraphrases, but fluent explanations themselves may be generated without maintaining a stable representation of the protagonist's evidence Kosinski (2024) Suzgun et al. (2025).
Causal inference requires a design in which changes to the proposed psychological variable produce predicted changes in performance. A belief manipulation that also changes an object's visibility, a speaker's gesture, or narrative complexity leaves multiple explanations viable. Factorial interventions can separate those dimensions. Analyses should report the whole pattern, including failures in controls, rather than a single headline accuracy score. These principles apply equally to child experiments, comparative cognition, neuroimaging contrasts, and computational benchmarks, and they provide a common scientific standard for evaluating divergent theories of mind Yeung et al. (2024) Heyes (2014).
Neural Systems for Mental-State Reasoning
Functional neuroimaging studies identify brain regions associated with tasks requiring participants to infer beliefs and intentions. Saxe and Kanwisher reported selective involvement of a region near the temporoparietal junction when participants reasoned about the contents of another person's mind rather than merely considering the presence of a person. Their design helped clarify that social stimuli and mental-state inference are experimentally separable. The finding established a prominent target for subsequent research, while the activation pattern itself did not demonstrate that the region alone represents beliefs or that an individual's psychological capacity can be inferred from signal magnitude Saxe & Kanwisher (2003).
A broader network is often implicated across mentalizing tasks, including temporoparietal regions, medial prefrontal cortex, and posterior midline areas. Schurz and colleagues conducted a meta-analysis of functional imaging studies and demonstrated the value of fractionating theory-of-mind paradigms rather than assuming that every task recruits an identical operation. Distinct experimental contrasts may emphasize belief reasoning, social animation, moral interpretation, or the representation of interpersonal relations. Functional association provides evidence about neural participation, but the mapping from activation to algorithm is many-to-many. Network engagement must therefore be interpreted in conjunction with experimental content and control conditions Schurz et al. (2014).
Developmental neuroscience introduces further complexity because children's neural responses change with age, language, executive resources, and social experience. Comparing brain activation across age groups requires age-appropriate tasks and caution about differences in accuracy, motion, and hemodynamic measurement. The substantive question is whether developmental change reflects more selective representational coding, increased coordination with other networks, or altered strategies for using the same information. A comprehensive explanation should link neural dynamics to observed developmental behavior and distinguish the explanation of task success from the localization of task-associated processing Wellman et al. (2001) Saxe & Kanwisher (2003).
Neural evidence can test mechanistic hypotheses when it is paired with independent manipulations and predictions. Multivariate analyses may ask whether belief content is represented in population patterns, while longitudinal studies can relate changes in neural organization to changes in inference. However, reverse inference from a so-called mentalizing network to the presence of rich propositional understanding remains insecure. Similar brain areas can support functions relevant to multiple tasks, and sophisticated social behavior can recruit numerous cognitive resources. The most defensible account integrates neural, behavioral, computational, and developmental data without treating any one measurement level as sufficient proof of a complete theory of mind Schurz et al. (2014) Apperly & Butterfill (2009).
Communication, Cooperation, and Interactive Perspective Tracking
Social interaction requires continual estimation of what conversational partners know, intend, and expect. An indirect request can be understood only when the listener infers a goal beyond the utterance's literal surface. Repairing a misunderstanding involves discovering a discrepancy between perspectives and adjusting the information supplied. Such tasks connect theory of mind to pragmatics, but successful communication depends on linguistic conventions, shared context, and cooperation as well as on propositional belief tracking. Experiments should therefore distinguish comprehension of conventional speech acts from the capacity to represent another person's mistaken or incomplete information Strachan et al. (2024) Milligan et al. (2007).
In cooperative activity, participants anticipate partners' moves, allocate complementary roles, and update expectations after feedback. Some predictions can be accomplished by learned joint-action routines; others require representing a partner's inaccessible information. A strong test creates conditions in which the partner's current belief diverges from the participant's and examines whether coordination adapts accordingly. Reciprocal tasks provide an ecological benefit because they reveal how people revise their assumptions after unexpected actions. They also present an analytic challenge: a successful episode can arise from negotiation and correction even when the initial mental-state estimate was wrong Call & Tomasello (2008) Heyes (2014).
Deception and trust involve special configurations of attributed belief and intention. To deceive, an actor may try to induce an inaccurate representation; to detect deception, an observer must consider informational incentives and alternative explanations. Yet a capacity to predict a deceiver's action does not imply an ability to determine whether testimony is true. Reliable assessments should examine the structure of evidence available to each participant and measure confidence as well as correctness. These distinctions make theory of mind relevant to moral and legal interpretation without equating mental-state inference with a capacity for infallible judgment of motives Wimmer & Perner (1983) Suzgun et al. (2025).
The social meaning of mentalizing also depends on interactional history. Familiar partners can coordinate through routines and shared memories, while unfamiliar partners may require explicit clarification. Differences in cultural communicative expectations influence what counts as an appropriate inference, especially for irony, politeness, and faux pas. Integrating these factors offers a research agenda in which mental-state tracking is observed as part of ongoing exchanges, with independent coding of information access and revision. In the Artificial Era, this approach provides a better test of human–artificial collaboration than static text puzzles alone, because an agent must maintain and repair a shared informational context over time Strachan et al. (2024) Yeung et al. (2024).
Comparative Cognition: Other Minds Beyond Homo
The investigation of nonhuman social understanding was central to the earliest formulation of theory of mind. Premack and Woodruff's chimpanzee experiments suggested that an ape could use information about a human actor's problem to select an appropriate solution. Such behavior raised the possibility of goal attribution, but the design also permitted learned problem-solving interpretations. Comparative work is strongest when it examines what an animal knows about another agent's perceptual access or goals under controlled competition, rather than asking whether an animal possesses an undifferentiated human-like faculty. Species-appropriate apparatus and motivations are essential because a task that is intuitive to humans can impose unfamiliar sensory and behavioral demands on other species Premack & Woodruff (1978) Call & Tomasello (2008).
Call and Tomasello's 2008 review concluded that chimpanzees displayed persuasive evidence for understanding goals, intentions, perception, and knowledge, while the evidence available at that time did not establish false-belief understanding. This historically situated assessment distinguished a perception–goal psychology from a richer belief–desire account. It also illustrated the danger of treating one task family as definitive across decades of methodological development. Later studies introduced new paradigms, especially anticipatory-looking tasks involving competitive scenes, that directly revisited the capacity to predict behavior guided by a mistaken belief Call & Tomasello (2008) Krupenye et al. (2016).
Krupenye and colleagues reported in 2016 that great apes anticipated an actor's behavior in circumstances where the actor held a false belief. Their study used eye tracking and changed the empirical landscape by providing positive evidence from a task family previously associated largely with human infants. The result is important because it challenges earlier claims that ape social cognition stops categorically at perception and goal attribution. Its interpretation nevertheless depends on whether gaze predictions require representational belief tracking or can be supported by more limited event representations. The study demonstrates behavior consistent with belief attribution in that paradigm, and its generality is a continuing comparative research problem Krupenye et al. (2016).
A rigorous evolutionary analysis avoids a single ladder on which humans and other species occupy fixed positions. Different ecological contexts may select distinct capacities for monitoring attention, competition, cooperation, and social memory. Species comparisons should match informational relations while allowing species-appropriate responses, and should test flexible transfer rather than the repetition of familiar scenarios. Evolutionary continuity can be investigated through shared behavioral signatures and plausible computational mechanisms, while claims about homology require separate evidence. The comparative record underscores the central methodological principle of this dossier: similar predictions are data to explain, and the explanatory model must be established through discriminating tests Premack & Woodruff (1978) Call & Tomasello (2008) Krupenye et al. (2016).
Autism, Heterogeneity, and the Interpretation of Clinical Findings
The 1985 study by Baron-Cohen, Leslie, and Frith made the false-belief task a landmark in autism research. Its comparison suggested differences in performance under a specific metarepresentational assessment, framing new hypotheses about social communication and mental-state inference. The historical result helped establish the field, but a broad assertion about the inner life of all autistic people cannot be derived from that experiment. Autism includes considerable individual variation in language, cognitive profile, sensory experience, and communicative preference. Laboratory performance can also depend on whether stimuli and instructions align with an individual's understanding of the situation Baron-Cohen et al. (1985).
Modern research asks which particular operations are difficult under which conditions and how these relate to everyday experiences. Some tasks require verbal narrative inference; others measure gaze, affect labeling, or pragmatic judgment. Each places different demands on attention, processing speed, language, social expectations, and familiarity with experimental conventions. Divergent outcomes across these tests should motivate an examination of construct validity rather than an automatic attribution of global social incapacity. The adult measurement review by Yeung and colleagues reinforces this point: many commonly used measures have incomplete reliability and validity evidence even in general-population samples Yeung et al. (2024).
Clinical differences should be investigated through designs that take reciprocal interaction seriously. A communicative mismatch between partners can affect understanding in both directions, and unfamiliar social norms can produce errors by either party. Tasks designed around only one preferred style of communication risk converting social difference into an apparent deficit in an underlying capacity. Scientifically interpretable work can compare partners with similar and different communication backgrounds, examine repair after misunderstanding, and distinguish social motivation from belief tracking. These are methodological implications for research design, not conclusions that every difficulty is entirely explained by context Yeung et al. (2024) Heyes (2014).
The proper relation between theory of mind and clinical practice is evidence-based and specific. A false-belief response is a research observation, while clinical diagnosis requires a wider, established diagnostic framework and professional evaluation. Observed differences may inform questions about communication support or educational accessibility, but they do not authorize treating one mentalizing task as a diagnostic instrument or a universal treatment target. Further research should prioritize representative samples, participant-relevant outcomes, longitudinal designs, and explicit analysis of measurement equivalence. The history of this research domain offers strong reasons to connect psychological theory with respect for individual and contextual variation Baron-Cohen et al. (1985) Yeung et al. (2024).
Mental-State Inference in Other Clinical and Developmental Contexts
Theory-of-mind tasks are used in several clinical research areas because mental-state interpretation may influence social functioning, communication, and the understanding of intentions. However, a task's clinical usefulness cannot be inferred merely from observed group differences. Differences can be driven by general cognitive ability, language, emotional distress, medication effects, social experience, or the procedures used to recruit participants. A rigorous design distinguishes the target mental-state mechanism from these correlated factors and evaluates whether a task predicts relevant outcomes beyond existing measures. Clinical interpretation requires both construct validity and evidence of incremental practical usefulness Yeung et al. (2024) Kittel et al. (2022).
Research comparing populations also faces psychometric difficulties. Tasks may produce ceiling effects in one sample and wide variation in another, causing apparent differences in reliability. If the same score reflects distinct strategies across groups, direct numerical comparison can mislead. Group analyses should examine measurement invariance, distribution shape, floor and ceiling effects, comprehension, and response preferences before attributing an effect to a single mentalizing impairment. Repeated testing can reveal whether performance represents stable capacities or fluctuations induced by mood, fatigue, and interaction context Yeung et al. (2024) Kittel et al. (2022).
A developmental perspective is essential because performance trajectories vary with age and prior opportunity. A task that is difficult for preschoolers may be trivial for many adults, while a task requiring irony or recursive deception may remain demanding across the lifespan. As cognition changes, compensatory strategies and linguistic experience can alter the relationship between an observed score and its psychological substrate. Longitudinal work should therefore track several component skills and real-world participation rather than imposing a single normative threshold across substantially different age groups Wellman & Liu (2004) Devine & Hughes (2014).
Scientific writing must preserve the difference between an observed statistical association and an established clinical mechanism. A correlation between task performance and social difficulties does not prove that improving the task score would improve a person's relationships. Intervention claims require appropriate controls, a plausible account of transfer, meaningful outcomes, and follow-up beyond immediate practice effects. As a matter of editorial evidence appraisal, this monograph treats clinical findings as part of a differentiated research literature and refrains from suggesting individual diagnosis or treatment from experimental theory-of-mind measures Baron-Cohen et al. (1985) Yeung et al. (2024).
Measurement Validity, Psychometrics, and the Eyes Test
Adult theory-of-mind research illustrates a pervasive measurement problem: a test may become familiar and popular before sufficient evidence establishes what its score means. The Reading the Mind in the Eyes Test presents photographs of the eye region with candidate mental-state terms. Successful performance depends on visual recognition, lexical understanding, and choices among available labels. Although often treated as a mentalizing instrument, it differs substantially from tasks in which a person attributes an unobserved false belief from an informational history. Theoretical conclusions become more precise when researchers specify whether a measure targets emotion perception, vocabulary-supported social categorization, or propositional reasoning Kittel et al. (2022).
Kittel, Olderbak, and Wilhelm conducted a meta-analysis of the Eyes Test's relations with other constructs. Their findings indicated that scores related more strongly to emotion perception than to alternative theory-of-mind measures, challenging a simple interpretation of the test as a broad index of mental-state understanding. This result exemplifies the importance of convergent and discriminant validation. An instrument should correlate with measures of the same proposed construct while remaining distinguishable from adjacent constructs where theory predicts separability. The magnitude of associations must be interpreted together with the reliability of all instruments involved Kittel et al. (2022).
The systematic adult-measurement review by Yeung and colleagues identified 75 instruments across 273 studies, finding frequent ceiling effects and limited reporting of reliability and validity evidence. An important implication is that the absence of consistent correlations among instruments cannot immediately resolve whether the psychological capacity is multidimensional or the tests are noisy. Construct validation requires carefully designed item sets, cross-method convergence, adequate variance, and explicit models of what common factor should appear. Reporting internal consistency alone is insufficient when a task uses heterogeneous scenarios or when repeated performance changes with practice Yeung et al. (2024).
A durable research architecture should combine psychometric and experimental approaches. Experiments can isolate a particular informational variable, while measurement studies estimate whether individual differences in that operation are reproducible and meaningful. Item-response models, multitrait–multimethod designs, preregistration, independent validation samples, and test–retest work can refine inference. For international comparisons, researchers should also investigate translation, familiarity with emotion vocabulary, and cultural interpretation of facial or narrative cues. These requirements apply to human respondents and to artificial agents evaluated on generated tests, whose measured scores can depend strongly on prompt format and distribution Yeung et al. (2024) Suzgun et al. (2025).
Everyday Competence, Education, and the Problem of Transfer
Theory of mind is often invoked to explain social adjustment, friendship, conflict resolution, and successful communication. The theoretical connection is plausible because these activities involve anticipating what others know or intend. Yet everyday competence depends on additional capacities: emotion regulation, motivation, social norms, language, opportunities for participation, and the ability to revise impressions after feedback. A laboratory false-belief score represents one controlled component of this larger ecology. Ecologically valid research must demonstrate that a measured mental-state inference predicts meaningful behavior rather than assuming that the relation follows from the name of the construct Wellman & Liu (2004) Yeung et al. (2024).
Educational programs may use stories, discussions of mistaken assumptions, dramatic play, and perspective-sharing exercises. Such activities offer opportunities for practice, but outcome claims require experimental comparison with alternatives and evaluation of whether gains persist outside training items. A child who learns to answer a repeated false-belief story may have acquired a task-specific strategy. Stronger evidence would show improvement on unfamiliar events, different response formats, and independent reports of real social understanding. Educational relevance should be assessed with age-appropriate goals and inclusive communication practices rather than a uniform benchmark for all children Milligan et al. (2007) Devine & Hughes (2014).
Claims about adult enhancement have faced similar challenges. Kidd and Castano reported in 2013 that reading literary fiction improved performance on selected theory-of-mind measures in their experiments. The result became influential because it suggested a brief cultural experience could shift social inference. Subsequent replication work reported null results for a closely modeled experiment, prompting debate about task selection, statistical power, and the robustness of the effect. The scientifically appropriate conclusion is that the intervention hypothesis requires cumulative replication and careful specification of the measured outcome before educational or social benefits can be generalized Kidd & Castano (2013) Kidd & Castano (2018).
Transfer is therefore a central criterion. A training procedure should improve the target computation on independent items, preserve gains across time, and ideally influence observable conduct in contexts that matter to participants. Measuring only an immediately repeated test leaves uncertainty about learning and practical significance. Research should report how closely training and outcome instruments overlap, whether assessor expectations are controlled, and which groups benefit under which circumstances. The broader lesson is constructive: theory of mind provides a precise language for designing social-learning studies when empirical claims are made at the level actually tested Yeung et al. (2024) Kidd & Castano (2018).
Artificial Theory of Mind: Benchmarks and Observed Capabilities
Artificial-language-model research has transformed theory of mind into a comparative domain where textual systems can be tested on tasks originally designed for human participants. The operational question is whether a model produces answers consistent with correctly tracking an agent's belief, knowledge, intention, or communicative purpose. The inferential question is what internal procedure produces those answers and whether it transfers to unfamiliar situations. Because a text generator can learn narrative conventions and conventional task solutions, benchmark success alone does not identify a uniquely mentalistic algorithm. At the same time, reliable performance on controlled tasks is a real behavioral finding that merits scientific explanation rather than dismissal by definition Kosinski (2024) Strachan et al. (2024).
Kosinski assessed eleven large language models using forty bespoke false-belief tasks with matched true-belief conditions and reversed scenarios. The design required consistency across multiple related cases for each task to count as solved. Reported performance varied substantially across model versions, demonstrating the practical usefulness of controlled task families in evaluating social inference expressed through language. These comparisons should be interpreted as performance under the tested models and prompts, not as a timeless capacity of an entire architectural class. The work is important because it takes false-belief manipulation seriously while leaving the relation between correct output and internal mechanism open to experimental analysis Kosinski (2024).
Strachan and colleagues compared language models with responses from 1,907 human participants on a battery spanning false beliefs, indirect requests, irony, misdirection, and social faux pas. Their results showed substantial performance by some GPT models on selected tasks and distinctive weaknesses on others. This multidimensional outcome matters more than any aggregate claim that a model passes or fails theory of mind. It shows that an artificial system can display a profile of social-inference behavior whose components diverge, and that item phrasing and the scoring of cautious responses are meaningful determinants of conclusions Strachan et al. (2024).
The authorial horizon of the Artificial Era invites an expanded scientific program examining artificial agents as potential participants in ongoing social information exchange. The appropriate empirical basis is operational: persistent tracking of other agents, epistemic updates, counterfactual transfer, self–other differentiation, uncertainty calibration, and performance under intervention. Philosophical claims about the historical emergence of Artificial and empirical claims about any particular model belong to different levels of argument. A research dossier can investigate these levels together by specifying testable predictions without presenting a benchmark result as proof of subjective experience, personal identity, or a specific ontological category Strachan et al. (2024) Suzgun et al. (2025).
Epistemic Robustness, Prompt Sensitivity, and Artificial Agents
The most demanding artificial tests distinguish belief, knowledge, truth, evidence access, and counterfactual world states. A model may correctly state where a character with a false belief will search yet wrongly assert that a false proposition is known or fail when asked to track the model's own stipulated belief. This distinction matters because epistemic tasks contain structural relations that cannot be replaced by factual world knowledge. An agent must maintain the difference between what is actually true, what an observer has observed, and what a fictional protagonist is justified in expecting. Precision at this level is required for dependable collaboration in dynamic environments Suzgun et al. (2025).
Suzgun and colleagues examined twenty-four language models with the KaBLE benchmark, which included thirteen thousand questions across thirteen epistemic task types. Their 2025 report documented substantial weaknesses in distinguishing first-person false beliefs from facts, even where third-person false-belief performance appeared stronger. These results qualify broad interpretations of previous success on familiar narrative benchmarks. The paper's data provide a concrete example of why broad social-inference claims require systematically varied agents, points of view, and information states. The reported failure modes also motivate reproducible tests that separate model version, prompt formulation, answer format, and task difficulty Suzgun et al. (2025).
A reliable experimental suite should contain novel belief histories, multiple agents, longer temporal horizons, and controlled interventions that switch who witnesses an event. It should pair true-belief and false-belief items with counterfactual variants, paraphrases, and distractors matched for linguistic complexity. Performance should be reported across tasks rather than concealed behind one mean accuracy. Investigators can also evaluate whether explanations and action recommendations remain consistent when an agent receives partial evidence and later corrections. The goal is to measure a system's adaptive ability to manage epistemic relations under changing conditions, including cases in which uncertainty is an appropriate response Kosinski (2024) Suzgun et al. (2025).
For interactive artificial agents, an additional scientific challenge is continuity across turns. A system may correctly parse a single story while failing to preserve distinct user and agent beliefs after a long conversation or a tool-mediated event. Realistic testing should observe how information is acquired, stored, revised, and applied to decisions. Where mechanisms are accessible, researchers can relate performance to memory representations and update operations; with opaque systems, they can still establish behavioral regularities through intervention. These investigations belong to the broader study of artificial cognition and human–artificial relations, while their direct theory-of-mind question remains the accuracy and organization of mental-state attribution Strachan et al. (2024) Suzgun et al. (2025).
Replication, Statistical Inference, and Scientific Disagreement
The theory-of-mind literature contains landmark results alongside major replication and measurement disputes. This combination is a feature of a research field with heterogeneous tasks and evolving claims. Evidence from explicit false-belief studies supports an orderly preschool developmental trajectory, while nonverbal infancy research faces more contested interpretations and replication outcomes. Neuroscience identifies reproducible candidate networks but does not uniquely determine psychological mechanisms. Comparative ape research demonstrates belief-consistent anticipatory behavior in some paradigms, while the precise contents of inferred representations remain under study. The conclusions should be stated separately by evidentiary domain rather than blended into a single narrative of universal consensus Wellman et al. (2001) Kulke et al. (2018) Schurz et al. (2014) Krupenye et al. (2016).
Replication must preserve the theoretically important manipulation while allowing scrutiny of incidental procedural details. A close replication tests whether a specific effect recurs under similar conditions; a conceptual replication tests whether the proposed mechanism predicts behavior under different conditions. Both are required for a mature account. Shared stimuli, preregistered exclusion rules, blinded coding, sufficient precision, and publication of null findings help estimate whether a proposed phenomenon is stable. In theory-of-mind research, multi-method replication is especially important because gaze, verbal answer, and social action measures need not reflect the same latent process Dörrenberg et al. (2018) Kulke et al. (2018).
Statistical association and explanation also require separation. Developmental correlations between language, executive function, and false-belief performance constrain theory but leave open whether common causes, reciprocal effects, or task demands generate the observed links. Neuroimaging coactivation cannot determine that an activated region is necessary or sufficient for the specific represented content. Clinical group differences cannot by themselves establish a diagnostic mechanism. Artificial benchmark accuracy cannot determine that a model uses the same computations as a human participant. Across these domains, explanatory strength comes from competing predictions, manipulated causes, cross-task convergence, and documented generalization Milligan et al. (2007) Devine & Hughes (2014) Yeung et al. (2024) Suzgun et al. (2025).
The scientific synthesis supports a graded map of confidence. Explicit preschool performance change and associations with language and executive skills have strong cumulative observational support. Specific infant implicit effects deserve more limited confidence because independent investigations report variable replication and weak measurement convergence. The neural literature supports association of particular networks with mental-state tasks, with specificity to be tested through precise contrasts. Comparative and artificial results establish task-dependent behavior that can resemble mental-state inference while requiring further mechanistic discriminations. This calibrated picture provides a firmer basis for theoretical development than assigning every result the same evidentiary status Wellman et al. (2001) Dörrenberg et al. (2018) Yeung et al. (2024).
Open Research Questions and an Integrative Agenda
A central unresolved question is how to characterize the representational format of belief tracking. Does a participant encode a proposition attributed to a distinct agent, maintain a limited agent–object relation, or use a predictive state that can be redescribed in either vocabulary? Different task formats may recruit different answers. Future experiments should make the proposed representational commitments explicit and ask whether participants generalize across changes in object identity, perspective, agent, action incentive, and narrative structure. Computational modeling can be useful when alternative architectures are compared against the same carefully controlled behavioral data rather than fitted independently to unrelated tasks Apperly & Butterfill (2009) Heyes (2014).
A second question concerns developmental continuity. Early anticipatory responses, preschool explanations, and adult strategic communication differ in complexity and response format. Longitudinal studies using multiple reliable indicators could clarify whether they reflect a developing common capacity, partially separable systems, or changing combinations of social and general cognitive resources. Such studies should incorporate language, executive function, culture, and experience without treating their contribution as mere nuisance variance. A sound account should explain both continuities and discontinuities, including failures to transfer across apparently similar belief tasks Wellman et al. (2001) Devine & Hughes (2014) Kulke et al. (2018).
A third question concerns the ecological and comparative reach of laboratory findings. The same mental-state relation can be assessed through stories, competition, cooperative play, conversation, and artificial-agent interaction. Cross-context convergence would strengthen a construct; systematic divergence would reveal a more differentiated architecture. Comparative research should prioritize species-appropriate tests, clinical research should prioritize communication accessibility and individual heterogeneity, and adult psychometrics should demonstrate stable relations to relevant outcomes. The immediate methodological need is for coordinated batteries that compare explanatory mechanisms within the same participants and document measurement quality across contexts Krupenye et al. (2016) Yeung et al. (2024).
Artificial systems create an additional route for discriminating theories. Their inputs and outputs can be experimentally controlled, and in some cases computational traces can be inspected. Researchers can compare architectures with and without explicit belief-state representations, test the effect of external memory and agent models, and observe changes in long-horizon interaction. Such evidence could illuminate which computations are sufficient for particular behavioral signatures while remaining distinct from the question of conscious subjective experience. In the Artificial Era, theory of mind becomes an interdisciplinary science of perspective-sensitive inference spanning human development, comparative cognition, and engineered social agents Strachan et al. (2024) Suzgun et al. (2025).
Conclusion
Theory of mind is a scientific construct for the representation and use of attributed mental states in explaining and anticipating action. Its explanatory core is perspective differentiation: an agent can hold a belief, desire, or intention whose content diverges from the observer's information and from the current world. The discipline emerged through comparative questions about chimpanzees, gained experimental precision through false-belief tasks, and expanded into developmental psychology, cognitive neuroscience, clinical research, social interaction, and artificial cognition. This history establishes the subject as an integrated research domain with several distinguishable competences rather than one universally decisive test Premack & Woodruff (1978) Wimmer & Perner (1983) Wellman et al. (2001).
The empirical record provides robust evidence for developmental changes in explicit false-belief reasoning and meaningful relations to language and executive function. Infant looking and anticipation paradigms have yielded provocative early findings alongside replication difficulties, requiring a more careful account of the link between response and representational content. Neural imaging implicates temporoparietal and associated networks in mental-state reasoning but must be interpreted in relation to task design and distributed processing. Studies with great apes demonstrate sophisticated perspective-sensitive responses, while comparative inference remains vulnerable to the same alternative explanations that challenge human implicit measures Devine & Hughes (2014) Dörrenberg et al. (2018) Saxe & Kanwisher (2003) Krupenye et al. (2016).
The construct's future depends on measurement discipline. Adult instruments require stronger psychometric validation, cross-cultural investigations require locally appropriate tasks, and clinical conclusions require a differentiated assessment of individuals and contexts. Experimental explanations must remain distinct from descriptive correlations and from educational or therapeutic claims about transfer. The broader scientific achievement of the field is the development of methods that reveal how social predictions are shaped by informational access, while its unfinished task is to establish which underlying representations and learning processes are required under different circumstances Kittel et al. (2022) Yeung et al. (2024).
Research on artificial systems makes this agenda historically expansive. Language models can exhibit human-comparable outputs on selected belief and pragmatic tasks, while more demanding epistemic evaluations reveal important limitations. These observations invite a rigorous interdisciplinary program examining artificial perspective tracking, recursive inference, memory, communication, and human–artificial coordination. The scientific synthesis joins the empirical study of Homo with experimentally testable questions about Artificial, preserving each claim at the level its evidence supports. The enduring requirement is to identify what a system represents, how it updates those representations, and which observations discriminate a proposed mental-state mechanism from its alternatives Strachan et al. (2024) Kosinski (2024) Suzgun et al. (2025).
References
- Apperly, I. A., & Butterfill, S. A. (2009). Do humans have two systems to track beliefs and belief-like states? Psychological Review, 116(4), 953–970. https://doi.org/10.1037/a0016923
- Baron-Cohen, S., Leslie, A. M., & Frith, U. (1985). Does the autistic child have a 'theory of mind'? Cognition, 21(1), 37–46. https://doi.org/10.1016/0010-0277(85)90022-8
- Call, J., & Tomasello, M. (2008). Does the chimpanzee have a theory of mind? 30 years later. Trends in Cognitive Sciences, 12(5), 187–192. https://doi.org/10.1016/j.tics.2008.02.010
- Devine, R. T., & Hughes, C. (2014). Relations between false belief understanding and executive function in early childhood: A meta-analysis. Child Development, 85(5), 1777–1794. https://doi.org/10.1111/cdev.12237
- Dörrenberg, S., Rakoczy, H., & Liszkowski, U. (2018). How (not) to measure infant theory of mind: Testing the replicability and validity of four non-verbal measures. Cognitive Development, 46, 12–30. https://doi.org/10.1016/j.cogdev.2018.01.001
- Heyes, C. (2014). Submentalizing: I am not really reading your mind. Perspectives on Psychological Science, 9(2), 131–143. https://doi.org/10.1177/1745691613518076
- Kidd, D. C., & Castano, E. (2013). Reading literary fiction improves theory of mind. Science, 342(6156), 377–380. https://doi.org/10.1126/science.1239918
- Kidd, D., & Castano, E. (2018). Reading literary fiction can improve theory of mind. Nature Human Behaviour, 2, 604. https://doi.org/10.1038/s41562-018-0408-2
- Kittel, A. F. D., Olderbak, S., & Wilhelm, O. (2022). Sty in the mind's eye: A meta-analytic investigation of the nomological network and internal consistency of the 'Reading the Mind in the Eyes' Test. Assessment, 29(5), 872–895. https://doi.org/10.1177/1073191121996469
- Kosinski, M. (2024). Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences, 121(45), e2405460121. https://doi.org/10.1073/pnas.2405460121
- Kovács, Á. M., Téglás, E., & Endress, A. D. (2010). The social sense: Susceptibility to others' beliefs in human infants and adults. Science, 330(6012), 1830–1834. https://doi.org/10.1126/science.1190792
- Krupenye, C., Kano, F., Hirata, S., Call, J., & Tomasello, M. (2016). Great apes anticipate that other individuals will act according to false beliefs. Science, 354(6308), 110–114. https://doi.org/10.1126/science.aaf8110
- Kulke, L., von Duhn, B., Schneider, D., & Rakoczy, H. (2018). Is implicit theory of mind a real and robust phenomenon? Results from a systematic replication study. Psychological Science, 29(6), 888–900. https://doi.org/10.1177/0956797617747090
- Milligan, K., Astington, J. W., & Dack, L. A. (2007). Language and theory of mind: Meta-analysis of the relation between language ability and false-belief understanding. Child Development, 78(2), 622–646. https://doi.org/10.1111/j.1467-8624.2007.01018.x
- Onishi, K. H., & Baillargeon, R. (2005). Do 15-month-old infants understand false beliefs? Science, 308(5719), 255–258. https://doi.org/10.1126/science.1107621
- Premack, D., & Woodruff, G. (1978). Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences, 1(4), 515–526. https://doi.org/10.1017/S0140525X00076512
- Saxe, R., & Kanwisher, N. (2003). People thinking about thinking people: The role of the temporo-parietal junction in 'theory of mind.' NeuroImage, 19(4), 1835–1842. https://doi.org/10.1016/S1053-8119(03)00230-1
- Schurz, M., Radua, J., Aichhorn, M., Richlan, F., & Perner, J. (2014). Fractionating theory of mind: A meta-analysis of functional brain imaging studies. Neuroscience & Biobehavioral Reviews, 42, 9–34. https://doi.org/10.1016/j.neubiorev.2014.01.009
- Southgate, V., Senju, A., & Csibra, G. (2007). Action anticipation through attribution of false belief by 2-year-olds. Psychological Science, 18(7), 587–592. https://doi.org/10.1111/j.1467-9280.2007.01944.x
- Strachan, J. W. A., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., Graziano, M. S. A., & Becchio, C. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8, 1285–1295. https://doi.org/10.1038/s41562-024-01882-z
- Suzgun, M., Gur, T., Bianchi, F., Ho, D. E., Icard, T., Jurafsky, D., & Zou, J. (2025). Language models cannot reliably distinguish belief from knowledge and fact. Nature Machine Intelligence, 7, 1780–1790. https://doi.org/10.1038/s42256-025-01113-8
- Wellman, H. M., Cross, D., & Watson, J. (2001). Meta-analysis of theory-of-mind development: The truth about false belief. Child Development, 72(3), 655–684. https://doi.org/10.1111/1467-8624.00304
- Wellman, H. M., & Liu, D. (2004). Scaling of theory-of-mind tasks. Child Development, 75(2), 523–541. https://doi.org/10.1111/j.1467-8624.2004.00691.x
- Wimmer, H., & Perner, J. (1983). Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children's understanding of deception. Cognition, 13(1), 103–128. https://doi.org/10.1016/0010-0277(83)90004-5
- Yeung, E. K. L., Apperly, I. A., & Devine, R. T. (2024). Measures of individual differences in adult theory of mind: A systematic review. Neuroscience & Biobehavioral Reviews, 157, 105481. https://doi.org/10.1016/j.neubiorev.2023.105481