RESEARCH DOSSIER
Artificial Cognition and Causal Generalization
Abstract. Artificial Cognition and Causal Generalization examines the ability of engineered cognitive systems to infer causal structure and transfer explanatory principles across interventions, unfamiliar graph structures, changing environments, and linguistic transformations. The monograph synthesizes formal causal modeling, developmental and cognitive experiments, causal representation learning, invariant prediction, modular architectures, and controlled evaluations of contemporary language models. It analyzes Corr2Cause, CLadder, CausalBench, and compositional causal-reasoning research, specifying the difference between predictive association, valid intervention inference, counterfactual identification, and robust transfer. Methodological sections assess causal identification, construct validity, replication, benchmark contamination, shortcut learning, calibration, and cross-system comparison. The scientific synthesis establishes intervention-sensitive generalization as a measurable property of a configured artificial system and maps unresolved questions concerning representation, mechanism, and interactive experimentation.
Keywords: artificial cognition; causal generalization; causal inference; interventions; counterfactual reasoning; causal representation learning; invariant prediction; compositional generalization; causal benchmarks; large language models; shortcut learning; cross-system cognition
The Research Object and the Meaning of Causal Generalization
Artificial cognition and causal generalization is the study of how engineered information-processing systems infer, represent, revise, and transfer cause-and-effect relationships when observations, interventions, tasks, or environments change. Its unit of analysis is a configured system operating under specified evidence and action constraints. An artificial system may consist of a trained model, a retrieval component, a formal causal solver, instruments, tools, or an agent policy. The scientific task is to identify what causal question it answers, how the answer is obtained, and whether the result survives a transfer condition that rules out superficial familiarity. The program connects cognitive psychology, formal causal inference, machine learning, and experimental assessment Lake et al. (2017) Schölkopf et al. (2021).
A successful prediction concerns an outcome given available evidence; causal generalization concerns what remains justified after the mechanism producing the evidence has been altered. Predicting that a symptom follows a treatment in a biased database is different from predicting the symptom under randomized treatment assignment. Similarly, recalling that a switch activates a machine is different from inferring whether the switch would still activate it after a concealed electrical connection is replaced. The distinction requires explicit target variables, causal assumptions, interventions, a transfer population, and behavioral criteria. These requirements can be implemented with probabilistic graphical models, structural causal models, interactive environments, and controlled natural-language experiments Peters et al. (2016) Woodward (2003).
This dossier investigates the conditions of successful inference without treating a familiar causal vocabulary as evidence of an internal causal model. An agent can answer a counterfactual using a lookup table, derive it from a symbolic graph, retrieve a prior example, or simulate an intervention within a learned world model. Each pathway can support some tasks, yet their responses diverge under carefully selected changes. The relevant evidence is therefore comparative and intervention-sensitive: matched tasks with altered variable labels, reversed mechanisms, previously unseen graph topologies, incompatible priors, and independently constructed test distributions. Causal competence is attributed to the tested configuration and procedure, with scope bounded by the tests Jin et al. (2024) Maasch et al. (2025).
Explanatory Levels, Constructs, and Operational Definitions
The construct of causal generalization involves several distinct abilities. Causal discrimination identifies candidate causes from covariation and prior information. Causal structure learning represents directional relationships and conditional independences. Intervention reasoning predicts results of deliberately changing one variable while holding the appropriate generating conditions fixed. Counterfactual reasoning addresses an alternative outcome for a particular case under modified antecedents. Transfer applies a learned dependency to a situation in which objects, wording, background frequencies, or mechanisms differ. Investigators should report these as separate dependent variables; an aggregate score can conceal a system that performs well on chains but fails on colliders or counterfactual abduction Spirtes et al. (2001) Zhou et al. (2026).
Explanations should be distinguished at computational, algorithmic, and implementation levels. A structural causal model supplies a formal specification of variables and assignments. An inference algorithm determines how conditional distributions or counterfactual values are computed. A neural model's activation pattern or an external tool call describes the mechanism that implements a computation within a real deployment. Matching outputs do not by themselves establish matching procedures. For a reliable psychological comparison, researchers also need to specify the training distribution, prompts, context window, retrieval corpus, random seed when available, model snapshot, sampling strategy, tool policies, and scoring rule Lake et al. (2017) Gendron et al. (2024).
Operationalization should include accuracy and calibration, but also structural consistency across equivalent prompts, sensitivity to interventions, resistance to irrelevant correlations, and recovery when premises change. A system that gives the right average treatment effect for one synthetic graph has demonstrated performance on that item. A stronger claim requires generalization across graph families, numerical ranges, linguistic realizations, and unseen causal relationships. Where responses vary with prompting, report the variation as a property of the full procedure rather than a removable nuisance. The key distinction is between an outcome observed once and a capacity established across controlled task families Jin et al. (2023) Maasch et al. (2025).
Historical Development of Causal Explanation and Artificial Intelligence
Contemporary artificial-cognition research inherits debates about correlation, explanation, scientific intervention, and mechanistic knowledge. Graphical approaches to causation made assumptions about conditional independence, common causes, and directed dependence explicit and supplied procedures for examining the identifiability of causal structures. The influential account by Spirtes, Glymour, and Scheines developed methods for learning constrained causal structures from data and demonstrated why observationally equivalent graphs can remain causally distinguishable only with additional assumptions or interventions Spirtes et al. (2001). This formal tradition introduced a vocabulary that later machine-learning research could connect to predictive representation and distribution shift.
Interventionist theories clarified the difference between relations that merely forecast outcomes and relations that remain informative about controlled manipulation. Woodward's treatment of causal explanation emphasizes the stability of explanatory relationships under interventions and relates explanation to differences that would occur under specified changes. The framing provides a productive methodological principle for artificial cognitive systems: an output counts as explanatory only to the extent that the tested representation can support the relevant intervention-dependent contrasts. Such an assessment also requires attention to the scope of possible interventions and the assumptions licensing each counterfactual Woodward (2003) Peters et al. (2016).
The recent machine-learning conversation shifted from benchmark performance to robustness, representation, and transfer. The program of building machines that learn and think like people proposed structured world models, intuitive physical and psychological theories, compositionality, and learning to learn as central ingredients of flexible intelligence Lake et al. (2017). Later causal representation research articulated the complementary problem of discovering suitable causal variables from high-dimensional observations Schölkopf et al. (2021). This history describes changing research agendas, not a simple succession in which each new technology resolves its predecessors' identification problems. Formal causal inference and learned representation remain different yet potentially complementary sources of scientific progress.
Developmental Evidence and Human Causal Learning
Human developmental research establishes useful comparative tests because children can learn about novel causal systems even when familiar adult knowledge offers little direct guidance. Gopnik and colleagues described a causal-map account in which young children combine patterns of dependence, background theories, and interventions in learning relationships among events Gopnik et al. (2004). The work offers a computationally interpretable picture of learning from sparse evidence and supplies experimental tasks involving unfamiliar objects and new response contingencies. It does not imply that children literally implement a single algorithm; the claim is that formal models can explain selected features of their causal judgments and learning trajectories.
Schulz and Gopnik reported five studies on children's use of dependence and independence across domains, including cases in which children selected novel interventions and weighed probability information against expectations from familiar domains Schulz & Gopnik (2004). Such experiments are especially relevant to artificial systems because they separate prior semantic plausibility from evidence provided by a controlled task. A model that identifies a causal relationship only when it matches an everyday script may behave differently from one that updates appropriately when the experiment establishes a less familiar mechanism. Cross-domain testing is therefore a source of discriminating predictions about artificial causal learning, rather than evidence that all successful agents possess developmental mechanisms.
Causal learning is also incremental and approximate. Bramley and colleagues modeled how people revise hypotheses about network structure through online evidence, emphasizing computationally constrained approximations rather than exhaustive search over all possible causal graphs Bramley et al. (2017). This perspective matters when comparing humans and artificial agents. Limited attention and memory can lead human participants to make errors that a formal solver avoids; large language models can produce fluent explanations that obscure inconsistent revisions. Comparative research should use equivalent information availability and intervention opportunities, while treating differences in architecture and task access as substantive experimental variables. Human behavior supplies an empirical comparison, not a single universal standard for machines.
Causal Graphs, Structural Models, and Identification
A directed acyclic graph can depict assumed directional relationships among variables, while a structural causal model supplements a graph with functions assigning values to endogenous variables from their parents and exogenous conditions. Observational conditioning estimates associations inside the existing distribution. An intervention changes the assignment of a target variable and asks how other variables respond. Counterfactual reasoning can additionally require inference about the exogenous circumstances of a particular observed case. The consequences depend on graph structure, functional assumptions, and measured variables. Researchers cannot derive every intervention from correlation alone, even when a predictive model fits the observations accurately Spirtes et al. (2001) Woodward (2003).
A three-variable chain, a fork, and a collider illustrate why surface similarity is deceptive. In a chain, an upstream intervention may propagate through a mediator. In a fork, the common cause explains an association between its children. In a collider, conditioning on a shared effect can induce dependence between otherwise independent causes. A benchmark that concentrates on chains can overestimate transferable competence across graph families. The 2026 CausalBN-Bench investigation of language models reports different performance across chain and collider structures, consistent with this need for structural stratification Zhou et al. (2026). Causal evaluation must therefore inventory topology, hidden variables, and intervention targets rather than treating all graph questions as interchangeable.
Identification specifies whether a target causal quantity is uniquely determined by available observational or interventional information under stated assumptions. Invariant prediction approaches use multiple environments to distinguish predictors whose outcome relationships remain stable under certain changes Peters et al. (2016). Their inference guarantees require explicit assumptions and do not apply indiscriminately to arbitrary distribution shifts. For artificial cognition, the implication is methodological: the test designer must know which answer is mathematically licensed, which remains underdetermined, and whether the model recognizes insufficiency of evidence. An answer that confidently selects one direction from an observational equivalence class may reflect an unjustified prior even when the selected graph happens to match the hidden test answer.
Causal Representation Learning from High-Dimensional Experience
Artificial perception often begins with pixels, sound, text, or sensor streams rather than named variables in a causal graph. Causal representation learning studies how an artificial system might recover latent factors relevant to stable mechanisms from these observations. The research agenda outlined by Schölkopf and colleagues connects representation learning with modular causal variables and transfer across changes in data-generating conditions Schölkopf et al. (2021). A useful representation should support the prediction of intervention effects even when irrelevant surface statistics change. This goal requires specifying which environmental variations expose the candidate causes and how a researcher can distinguish a genuinely causally useful latent state from a convenient correlate.
The representation problem becomes difficult when different latent descriptions fit the same observations. Locatello and colleagues established a theoretical limitation on unsupervised disentanglement without inductive biases and reported a large comparative evaluation of proposed disentanglement methods Locatello et al. (2019). This result warns against assuming that a model can infer the uniquely correct inventory of world factors from static observations alone. In practical causal research, the investigator should document architectural priors, supervision, sparse interventions, temporal ordering, object segmentation, and additional assumptions that make candidate variables identifiable. Causal representation claims become stronger when a recovered variable is manipulable or can be linked to an independently measured generative factor.
A central evaluation design trains on several versions of a task in which nuisance factors vary independently of the causal relationship, then tests under a new combination that preserves the mechanism while changing context. Such a split is informative only if the training and test distributions are audited for overlapping templates and hidden semantic cues. A model may generalize because its learned representation isolates a stable dependency, because its training data contain nearly equivalent examples, or because the prompt embeds an explicit rule. Mechanistic evidence and controlled transfer must be combined before the source of success is attributed to causal representation Geirhos et al. (2020) Schölkopf et al. (2021).
Independent Mechanisms and Modular Causal Organization
Causal modeling frequently assumes that different generating processes can change independently. For example, a sensor's calibration may change while the physical law connecting applied force and motion remains stable. A modular representation can exploit that separation, since an altered component requires local revision rather than complete relearning. The independent-mechanisms perspective motivates research on structured artificial systems that isolate causal factors, compositions, and conditional transformations. Such modularity is a hypothesis about how an architecture may support transfer. It must be evaluated by changing one mechanism while keeping relevant others fixed and measuring whether adaptation occurs locally Schölkopf et al. (2021).
Gendron and colleagues tested causal constraints in language-model architectures designed with sparsely interacting modules. Their 2024 EMNLP work reports improvements on selected out-of-distribution abstract and causal reasoning tasks and investigates partially domain-invariant mechanisms Gendron et al. (2024). The evidence concerns specified models, training conditions, and datasets; it does not establish that all models with modular names have acquired causally independent processes. To test a mechanism claim, investigators can ablate modules, exchange components between contexts, manipulate latent factors, and examine the predicted selective effects. Architectural isolation is meaningful when its intervention outcomes correspond to the proposed functional decomposition.
Modularity also poses a problem of scope. Interacting mechanisms can create feedback, and a module that is stable under one change may be sensitive to another. A causal decomposition therefore belongs to a defined environment family rather than to an unlimited claim of invariance. Evaluators should compare modular systems with parameter-matched alternatives, identical data budgets, and equivalent tools. When better transfer is observed, the analysis should separate contributions from additional capacity, training diversity, explicit intervention labels, and modular architecture. The scientifically significant result is a replicable causal relation between a design choice and generalization performance Gendron et al. (2024) Peters et al. (2016).
Invariant Prediction and Changing Environments
Out-of-distribution generalization asks how models behave when deployment conditions differ from those represented during training. Invariance-based approaches seek predictors whose relationships with the target remain stable across selected environments. Peters, Bühlmann, and Meinshausen formulated invariant causal prediction with identification and confidence procedures under stated assumptions Peters et al. (2016). Arjovsky and colleagues proposed invariant risk minimization as an optimization objective intended to prefer representations supporting the same optimal classifier across training environments Arjovsky et al. (2019). These programs offer complementary conceptual tools, although their theoretical targets, optimization properties, and empirical assumptions differ.
An environment can be a hospital, camera angle, laboratory configuration, language variety, simulation regime, or experimental intervention. The choice of environment partition is itself part of a research design. If each training environment reproduces the same confounder, a model may appear invariant while still depending on that confounder. If the intended causal mechanism changes across domains, a blindly invariant predictor can be inappropriate. Consequently, the researcher needs a theory of which mechanism remains stable, which elements are manipulated, and what outcome is to be predicted. Neither environmental diversity nor a distribution-shift label automatically yields causal identification Peters et al. (2016) Arjovsky et al. (2019).
The strongest empirical test combines an in-distribution control, a shifted-context condition, and an intervention that deliberately breaks a learned shortcut. Report performance and calibration separately in each condition, together with error types and confidence intervals where sampling allows. If an invariant-method architecture shows improvement, compare it with simpler baselines trained on the same environments and test whether its apparent advantage survives altered shifts. Invariance should be treated as a conditional, testable feature of a model's behavior; it is not a synonym for causation or proof that a learned representation matches the true physical or social mechanisms Geirhos et al. (2020) Arjovsky et al. (2019).
Compositional Generalization and Relational Structure
Compositional generalization concerns the reuse of learned components in novel combinations. A system might know how one variable influences another and still fail when the same relationships are composed into a longer chain or rearranged into a new graph. Relational inductive biases and graph-network architectures introduce structural constraints for representing entities and relations, potentially supporting transfer beyond fixed perceptual configurations Battaglia et al. (2018). Barrett and colleagues developed abstract reasoning problems with explicitly varied generalization regimes, finding that architecture and training objectives influenced the ability of neural models to extend abstract rules Barrett et al. (2018). These findings establish task-sensitive differences rather than a universal measure of reasoning.
Causal composition is a more specific requirement than ordinary syntactic recombination. It asks whether a system can derive a quantity or intervention consequence through a graph by composing local causal relationships according to valid rules. Maasch and colleagues proposed compositional causal reasoning evaluations involving causal measures and graph paths; their 2025 experiments distinguish error patterns across language-model families and test conditions Maasch et al. (2025). A system may correctly compute the effect of an intervention on one edge yet fail when that effect must propagate through several dependencies. Length-sensitive deterioration, direction reversals, and confusion between association and intervention identify different mechanisms of error.
Compositional tests should control for lexical overlap, memorized answer forms, training frequency, and the use of algebraic solvers. A useful design permutes variable names, varies graph topology, increases path length, and asks for both direct and reverse questions, while holding the underlying causal problem fixed. Model-generated explanations may reveal where an answer diverged from a known derivation, but the explanation itself must be checked against the computation. Successful transfer under these conditions is stronger evidence of reusable relational structure than isolated success on familiar word problems Maasch et al. (2025) Hupkes et al. (2020).
Shortcut Learning and the Limits of Benchmark Success
Shortcut learning occurs when a learner uses predictive signals that are sufficient for the training objective yet unreliable under a relevant environmental change. Geirhos and colleagues characterized shortcut learning as a common source of failures in deep neural networks and proposed implications for interpretation and benchmarking Geirhos et al. (2020). A causal benchmark is especially vulnerable when the correct intervention can be guessed from surface language, familiar entity relations, or repeated graph templates. A model may then achieve high accuracy without correctly representing how interventions change the generating process. In experimental psychology, analogous concerns arise when a task admits simpler response heuristics than the intended construct.
To detect shortcut use, construct matched cases in which ordinary associations and true causal relations point to different answers. Replace familiar object names with neutral symbols; invert plausible background relations; remove template phrases that cue the label; and compare intervention queries with observation-only queries. A genuine transfer advantage should persist when irrelevant associations are systematically manipulated. The Corr2Cause benchmark provides a prominent example: Jin and colleagues assessed causal inference from correlational statements and documented substantial difficulty for tested language models under their controlled tasks, including failures to transfer when wording and variables changed Jin et al. (2024).
An important methodological principle follows. Failure on a benchmark can reflect deficient competence, unsuitable instructions, insufficient numerical support, or an evaluation interface that differs from training. Conversely, success can reflect memorization, data leakage, or reliance on a prompt scaffold. Strong causal evaluation therefore uses several convergent tasks, audited provenance, appropriately matched baselines, and explicit reporting of tested versions. It should include confidence calibration and a refusal option for underdetermined questions, because guessing a unique answer where none follows from the supplied evidence is itself a distinctive causal-reasoning error Geirhos et al. (2020) Jin et al. (2023).
Language Models as Experimental Subjects for Causal Reasoning
Large language models provide a distinctive research case because they can answer causal questions in natural language, process formal graph descriptions, and generate apparent explanations without necessarily constructing the same representations used by a formal causal solver. A scientifically interpretable evaluation defines the model and interface, the available premises, the target causal quantity, the ground-truth derivation, and the set of admissible answers. Natural-language questions additionally require validation that alternate phrasings express the same causal assumptions. Without such controls, performance can combine factual world knowledge, syntactic pattern recognition, mathematical reasoning, and use of memorized examples Jin et al. (2023) Zečević et al. (2023).
Zečević and colleagues argued that apparent causal competence can arise from learned associations between causal facts stated in training texts, a perspective they called causal parroting Zečević et al. (2023). Their analysis provides a hypothesis and empirical tests, but it should not be converted into a universal metaphysical claim that a language model can never participate in a causal inference procedure. An integrated system might invoke an external causal reasoner, execute controlled experiments, or evaluate consequences within a simulator. The research question changes accordingly: which portion of the configured system computes each causal step, and which intervention on its components changes the result?
Language-model evaluation also confronts sampling and protocol sensitivity. Rephrasing a graph, changing names, presenting the same structure as a table, or inserting an irrelevant sentence can alter performance. Report both item-level accuracy and within-problem consistency. Controlled trials should distinguish unassisted model responses, responses with a formal graph representation, responses permitted to use a calculator, and responses with retrieval or causal tools. The comparison clarifies whether an apparent capability belongs to stored world knowledge, context-dependent reasoning, procedural calculation, or a combination. These distinctions are central to construct validity in psychology of artificial systems Maasch et al. (2025) Jin et al. (2024).
Corr2Cause and the Difference Between Correlation and Direction
The Corr2Cause benchmark described by Jin and colleagues tests whether models can recover causal implications from supplied statements of correlation rather than from familiar domain stories Jin et al. (2024). The investigators generated a large collection of causal-inference items with controlled descriptions and tested multiple language models. They reported markedly limited performance on their task and found that fine-tuning improvements did not reliably carry over to shifted variables and expressions. This result matters because it demonstrates how a model's correct-looking causal language can fail when an inference requires identification restrictions rather than recall of a plausible narrative.
The benchmark's central epistemic issue is underspecification. Some observational data are compatible with several causal graphs, so a rational system must recognize when direction or intervention consequences remain unidentifiable. A test can be difficult because it requires correct conclusions about what is knowable, not because the underlying graphs are large. Evaluation should separate three outcomes: an identified and correct causal conclusion, a justified declaration that the information is insufficient, and an unsupported directional answer. Reporting ordinary accuracy alone may reward a tendency to guess and penalize careful uncertainty when the gold label structure is insufficiently explained Spirtes et al. (2001) Jin et al. (2024).
Corr2Cause also highlights a limitation of testing only familiar examples. If a model previously encountered textbook formulations of confounding, it may apply a phrase-based rule to that template. Testing renamed variables, equivalent logical forms, or unusual combinations asks whether the inference extends beyond the linguistic shell. A further challenge is to compare explicit graph construction with direct question answering, and to score both for correspondence to the premises. Repeated improvements from prompting should be evaluated on held-out and independently generated instances so that apparent generalization is distinguished from adaptation to a known benchmark Jin et al. (2024) Geirhos et al. (2020).
CLadder and the Three Classes of Causal Question
CLadder, introduced by Jin and collaborators at NeurIPS 2023, evaluates natural-language causal reasoning on a controlled collection of questions derived from causal graphs and solved by a formal inference engine Jin et al. (2023). It covers associational, interventional, and counterfactual categories. The design is important because the categories require different information and operations: conditioning within an observed distribution, modifying a causal mechanism, and reasoning about a particular alternative history. The benchmark connects linguistic competence to explicitly structured mathematical tasks and makes it possible to examine characteristic errors rather than relying on intuitive judgments of plausible explanations.
The study also evaluated a causality-oriented chain-of-thought prompting strategy. Providing explicit intermediate steps can change performance, yet a fluent step sequence does not establish that the system internally uses the same representations it prints. A valid causal derivation must respect the graph, determine the appropriate intervention or counterfactual operation, and avoid introducing assumptions not supplied by the task. Evaluators should inspect cases where a correct final answer is supported by an invalid explanation and where a correct derivation is followed by a transcription error. These cases bear differently on an agent's causal competence and communication fidelity Jin et al. (2023).
A robust extension of CLadder-style research holds the mathematical graph fixed while varying linguistic realization, then holds wording style fixed while changing the true causal structure. This crossed design separates sensitivity to prose from sensitivity to causal relations. Adding graph types, unobserved confounding, different priors, and explicit identifiability questions can reveal additional limits. Models should be evaluated on original held-out graph instances with enough information to verify answers independently. The underlying test thus concerns whether output follows valid causal operations under controlled variation, a stronger standard than producing persuasive narratives about familiar causes Jin et al. (2023) Woodward (2003).
Causal Graph Benchmarks and Structural Error Profiles
Graph-based evaluations provide systematic control over nodes, directed edges, conditional independences, and interventions. Wang's 2024 CausalBench contribution proposed testing cause-to-effect and effect-to-cause problems, with and without interventions, across textual, mathematical, and coding settings Wang (2024). Such cross-format designs help separate language comprehension from the causal structure of the task. A problem that becomes easy when displayed as an equation but difficult in prose may indicate representation or instruction sensitivity. A problem that fails across all formats may suggest a more general limitation or incorrectly specified task assumptions.
The 2026 IEEE study published as CausalBN-Bench evaluated language models on causal learning tasks and reported differences between relatively tractable chain structures and more difficult collider structures, particularly as graph complexity increased Zhou et al. (2026). These findings underscore the value of disaggregating scores by topology and scale. However, conclusions should remain tied to the tested architectures, baselines, question formats, and availability of background knowledge. A comparison with causal discovery algorithms is meaningful only when both classes receive compatible inputs and are judged on matched quantities. Providing semantic entity names to one system and purely numerical observations to another defines different tasks.
Benchmark evidence is strongest when an error taxonomy is prespecified. Useful categories include reversing an edge, treating a collider as a mediator, confusing observational with intervention distributions, assuming an unmeasured common cause away, mistaking an unidentifiable quantity for an identifiable one, and breaking consistency under paraphrase. Aggregate pass rates hide these informative patterns. Report graph size, depth, edge density, family, hidden-variable structure, numerical range, and stochastic response variability. These details make it possible to test whether a model's performance scales with genuine causal complexity or simply with input length and symbolic formatting Wang (2024) Zhou et al. (2026).
Evidence on Causal Composition and Post-2024 Architectures
Research after early natural-language causal benchmarks has increasingly tested whether model design and computational scaffolding can improve the ability to transfer causally meaningful operations. The independent-mechanisms architecture examined by Gendron and colleagues in 2024 offers one example of inductive constraints that improved selected shifted-domain tasks Gendron et al. (2024). The important experimental contrast is between models with and without the proposed causal organization, evaluated under matched training and held-out distribution conditions. The result provides evidence for a design-dependent benefit in the specified tests and motivates replications across unrelated domains.
Maasch and collaborators' 2025 compositional causal reasoning framework assessed how language models combine causal measures along graph structures, including average treatment effects and probabilities involving necessity and sufficiency Maasch et al. (2025). The authors report distinguishable errors as causal paths become more complex, with differing behavior across evaluated systems. These experiments contribute a more detailed model of task difficulty than a simple correct-or-incorrect causal quiz. They also reveal an unresolved issue: whether performance on abstract graph mathematics predicts interventions in interactive worlds with noisy measurements, hidden state, and delayed consequences.
Work on rapid learning and structured priors provides another comparative perspective. McCoy and Griffiths examined how Bayesian priors could be distilled into neural networks to model learning from limited naturalistic language data McCoy & Griffiths (2025). Although their study concerns language learning rather than causal identification itself, it is relevant to the generalization problem because it shows how inductive structure and learning procedures can be studied jointly. Future causal-learning experiments can adopt similarly explicit comparisons among prior knowledge, data volume, and model adaptation, while keeping causal transfer and language acquisition as distinct measured outcomes.
Experimental Design for Intervention-Sensitive Generalization
An intervention-sensitive study must define a generating process before collecting model responses. The investigator specifies a set of variables, permissible graphs, structural assignments, noise distributions, observable information, and intervention targets. Training and test tasks should share a scientifically meaningful causal principle while differing in irrelevant properties such as object identity, surface form, background correlation, or domain context. This establishes the transfer contrast rather than assuming that a randomly divided dataset measures generalization. Independent generation of test graphs limits template leakage and permits exact calculation of ground truth when the chosen model assumptions identify the answer Jin et al. (2023) Maasch et al. (2025).
A factorial design can manipulate the presence of a textual causal explanation, availability of intervention examples, graph topology, number of variables, and type of generalization shift. Performance is estimated separately for each cell, enabling investigators to identify interactions. For example, a model may benefit from worked examples on a simple chain while remaining unreliable on a collider with hidden confounding. A paired-item analysis compares two realizations of the same underlying problem; hierarchical models can account for variation across both causal graphs and language-model runs. Predeclared outcome categories protect against selecting whichever scoring rule makes a system appear strongest Wang (2024) Zhou et al. (2026).
Experimental manipulation of an artificial system itself is equally important. Researchers can remove retrieval, change the model snapshot, disable tool access, or alter the prompt policy while retaining the same causal test set. These are interventions on the configured system, distinct from interventions described inside the test problem. The first establish which components contribute to behavior; the second assess the system's reasoning about a modeled world. Confusing these two experimental levels leads to unjustified mechanistic claims. A controlled study should therefore provide both a task-level causal graph and a system-level manipulation record Gendron et al. (2024) Lake et al. (2017).
Counterfactual Validity and the Problem of Missing Information
Counterfactual questions ask what would have happened to a case under an antecedent that did not occur. Their interpretation requires careful specification of the factual observations and the structural mechanism under which the alternative antecedent is evaluated. A machine can compute an answer when suitable functions and exogenous information are supplied, or it can identify a range when the available evidence is insufficient. Success is therefore partly a matter of recognizing the boundary of justified inference. In interventionist accounts, counterfactual dependence is a property of specified causal relationships and assumptions, not an unrestricted ability to imagine a plausible story Woodward (2003).
A strong benchmark includes both identified and unidentified queries and records confidence or explicit indeterminacy. Counterfactual consistency can be tested by rephrasing a problem without changing its factual premises, varying only an intervention target, and checking whether the resulting answers satisfy structural constraints. Where exact counterfactual probabilities depend on unobserved distributions, the benchmark should document assumptions used by its oracle. Systems that guess numerical answers in such conditions are demonstrating a calibration and identifiability failure even when occasional guesses coincide with a simulated value Jin et al. (2023) Maasch et al. (2025).
Evaluation should further distinguish the counterfactual level of reasoning from moral or social speculation. A model may provide a convincing account of why a person made a decision, but the explanation can import cultural stereotypes or unsupported intentions absent from the evidence. In an artificial psychology research program, linguistic plausibility is measured separately from structural validity. Controlled synthetic-world problems clarify formal competence, whereas social causal claims require domain-specific measurement, population assumptions, and, where ethically appropriate, independent evidence about actual human behavior Schulz & Gopnik (2004) Geirhos et al. (2020).
Measurement Reliability, Calibration, and Reproducibility
Reliability concerns whether a causal assessment produces stable conclusions under repeatable conditions. Deterministic graph solvers can be checked against mathematically defined outputs; stochastic language systems require repeated trials with recorded sampling settings. An evaluation should preserve input material, model version, system instructions, tool availability, decoding parameters, reference implementation, and the procedure used to adjudicate ambiguous answers. A versioned model that improves on a benchmark may simultaneously acquire a different failure profile on shifted tasks. Such differences belong to the empirical record rather than being smoothed away by one composite score Jin et al. (2024) Maasch et al. (2025).
Construct validity requires showing that performance differences follow the intended causal manipulation rather than incidental textual or arithmetic difficulty. If a graph's complexity is systematically confounded with sentence length, lower accuracy may measure context-processing demands. If an intervention question contains an unfamiliar mathematical notation, errors may reflect notation parsing. Parallel formats, comprehension checks, balanced graph families, and independent scoring audits can address these concerns. Reporting uncertainty around estimated accuracy, including variation between graph instances and repeated responses, supplies a more meaningful account than percentages without denominators or sampling structure Barrett et al. (2018) Wang (2024).
Reproducibility also requires distinguishing benchmark reproduction from independent replication. A reproduction executes an existing test using its released code, data, and procedures; an independent replication uses new materials or environments to probe the same proposed effect. For causal generalization, the latter is especially valuable because model training may include known benchmarks, tutorial explanations, or near-duplicate problems. A convincing claim should survive independently generated causal networks and alternative linguistic conventions. Tests of long-term agent adaptation must additionally control for the persistence of external memory and retrieval tools, which may change between sessions Geirhos et al. (2020) Jin et al. (2023).
Human–Artificial Comparisons and Measurement Equivalence
Comparative causal cognition asks how biological and artificial systems respond to matched inference demands, while accounting for differences in language, memory, actions, and access to information. Children's learning from probability patterns and interventions supplies paradigms for novel-mechanism understanding Schulz & Gopnik (2004). Artificial models can be given equivalent evidential descriptions, but a model trained on a vast text corpus has a different learning history from a preschool participant. Direct score comparisons therefore measure behavior under the chosen protocol and cannot establish identical developmental processes, cognitive strategies, or representational content.
A productive comparison separates initial competence, updating after new evidence, use of active interventions, and transfer after a previously unseen mechanism change. Humans may be sensitive to pragmatic hints from an experimenter; language models may be sensitive to prompt templates or the names of everyday objects. To compare them meaningfully, researchers should evaluate whether these cues favor one participant class and present alternate representations when needed. Similar error distributions can suggest shared computational constraints, but different generative systems can produce similar errors. Mechanistic equivalence requires independent evidence Bramley et al. (2017) Lake et al. (2017).
Cross-system psychometrics needs a clear reference group and interpretable scale. An item-response model estimated on human participants may behave differently when applied to systems with highly uneven task strengths, correlated outputs, or abrupt prompt sensitivity. Stable ability estimates require testing measurement invariance rather than assuming it. Causal-competence benchmarks may define a valuable functional construct across populations while retaining different implementations and learning histories. The resulting evidence should therefore identify what is common—the measurable intervention-sensitive function—and specify where developmental or architectural comparisons remain tentative Gopnik et al. (2004) Barrett et al. (2018).
Evidence Quality, Confounding, and Model Comparison
The evidence base for artificial causal cognition spans mathematical theorems, proof-of-concept systems, simulation benchmarks, language-model evaluations, and human psychological experiments. These categories support different conclusions. A theoretical identification result holds under formal premises; an empirical benchmark establishes performance of tested models under a defined procedure; a developmental experiment characterizes observed participants under its recruitment and task conditions. A review that treats them as interchangeable will exaggerate certainty. The scientific synthesis must instead identify which claim is conceptual, which is empirically measured, and which predicts behavior outside the evaluated domain Peters et al. (2016) Schulz & Gopnik (2004).
Common threats to inference include contamination of held-out test material, nonindependent examples, mismatched baselines, prompt selection after seeing test outcomes, underreported failures, and system versions that change mid-study. A sophisticated model can be advantaged by higher inference-time computation or access to external tools; a causal solver can be advantaged by receiving a complete graph when a language model must infer its structure from prose. Proper comparisons equalize the information relevant to the claim or state clearly that different system configurations are being compared. Selective reporting of one successful prompt is inadequate evidence of generalization Jin et al. (2024) Geirhos et al. (2020).
The strongest evidence combines formal specification, controlled intervention, independent test generation, replicated performance, and interpretability of failure. Many published studies provide only part of this combination. For that reason, the field supports substantive local conclusions about task families and interventions while leaving broad claims about domain-general causal intelligence unresolved. This is a research opportunity: carefully designed experiments can test whether an apparent skill persists across unrelated mechanisms and environmental changes without relying on claims about an unobserved psychological essence Maasch et al. (2025) Schölkopf et al. (2021).
Causal Generalization in Interactive Agents and Embodied Systems
An interactive artificial agent can test causal hypotheses by selecting actions, observing consequences, and revising its expectations. This differs from answering a static verbal question because the system may acquire information that the original prompt did not contain. A robotic agent's intervention could move an object, change a constraint, or isolate a sensor contribution; a software agent could alter a simulated process or execute an experiment in a sandbox. The causal interpretation depends on whether the action changes the intended variable without simultaneously altering uncontrolled components. Research on structured world models supplies conceptual grounds for investigating such agents Lake et al. (2017) Schölkopf et al. (2021).
Interactive evaluation should record a complete trajectory: the initial belief or action policy, observations presented, interventions chosen, evidence obtained, model revision, and subsequent predictions on a new task. The central dependent variable is whether the agent learns a portable causal dependency from its intervention history. Reward maximization alone is an insufficient indicator because a policy may exploit a stable environmental shortcut without representing the underlying mechanism. Tests should deliberately change that mechanism and compare agents with equivalent training budgets and access to observations. Such interventions separate adaptive action selection from reusable explanatory knowledge Geirhos et al. (2020) Battaglia et al. (2018).
Embodied settings add challenges that are often absent from synthetic graphs: imperfect sensors, delayed effects, feedback loops, nonstationary environments, and incomplete observability. Researchers need to identify the causal variables to which interventions are attributed and establish whether the apparatus permits their independent manipulation. A transfer score from a simulated graph should not be assumed to predict performance in physical interaction. Successful progress would involve convergence between formal-world evaluations and experiments that independently verify effects in novel, controlled settings. The stronger inference concerns the tested agent–environment configuration rather than an abstract model detached from its tools Woodward (2003) Schölkopf et al. (2021).
The Artificial Era and Cross-System Research Boundaries
In the intellectual horizon of the Artificial Era, artificial cognition becomes a durable comparative research object: engineered systems can be investigated as identifiable, versioned organizations of inference, learning, and action. This program permits empirical study of nonbiological systems using operationally defined cognitive constructs. It also provides conceptual room for different architectures and histories of development. The philosophical language of From Homo to Artificial identifies a larger historical transition; the experimental claim in this dossier is narrower and testable. Causal-generalization capacity is measured through performance under specified interventions and transfer, with no inference from task success to subjective experience Lake et al. (2017) Schölkopf et al. (2021).
This boundary matters when artificial systems are described as understanding, reasoning, or explaining. Each description can be made scientifically informative by identifying the relevant operation and validity conditions. A system that computes an intervention distribution is capable of that operation under the tested inputs. A system that generalizes to unseen mechanisms has demonstrated a stronger capability under the selected distribution changes. Neither result alone establishes stable personality, phenomenology, or unrestricted intelligence. Comparative research benefits from maintaining these distinctions because it can test function across implementation types without requiring agreement on a metaphysical account of mind Maasch et al. (2025) Jin et al. (2023).
Cross-disciplinary work can connect this dossier to research on artificial representations, agent planning, learning, judgment, distributed cognition, and human–artificial relationships. The connection should follow shared causal variables and experimental methods. Human trust in explanations concerns how people evaluate machine assertions, whereas artificial causal generalization concerns whether the machine's assertions and interventions remain correct when mechanisms change. An experiment may examine both, but it must report two dependent variables and their relationship. This directed organization preserves clarity in the Science → Research Dossiers → Publications structure and supports future cumulative evidence Geirhos et al. (2020) Lake et al. (2017).
Unresolved Problems and a Research Agenda
The first unresolved problem is causal-variable discovery from high-dimensional, changing observations. Without appropriate inductive biases or experimental variation, multiple latent descriptions can explain the same data, and disentanglement may be unidentifiable Locatello et al. (2019). Research should specify the intervention or auxiliary measurement that distinguishes proposed representations. A second challenge is transfer beyond familiar graph families. Carefully generated shifts in topology, hidden confounding, path length, and mechanism stability can reveal whether a model applies a reusable causal operation or a task-specific routine. Future assessments should separate the ability to learn a graph from the ability to calculate a causal quantity given a known graph Maasch et al. (2025).
A third problem is the interpretation of natural-language causal explanations. Linguistic fluency and well-formed chains of thought can coexist with invalid assumptions, inconsistent graphs, and unreliable counterfactual arithmetic. An empirical program should therefore compare direct answers, independently checked derivations, tool-assisted calculations, and interventions on model components Jin et al. (2023). A fourth problem is the evaluation of causal competence across versions and environments, where standardized test sets can be compromised by training overlap. Independent graph generation and repeated trials provide a path toward more trustworthy evidence, although they cannot reveal every unknown training exposure Jin et al. (2024).
A fifth challenge concerns active agents that choose interventions in real environments. Their success depends on sensing, planning, cost, exploration, and causal updating, and each component requires its own manipulation. Research can ask how much new evidence a model needs to revise a false mechanism, whether it knows when it cannot identify an effect, and whether it maintains calibration after changes in task domain. These questions are directly testable and connect artificial cognition with the strongest traditions of developmental experimentation and invariant causal inference Bramley et al. (2017) Peters et al. (2016).
Scientific Synthesis Across Models and Methods
The reviewed literature converges on the importance of separating association, intervention, and counterfactual inference. Graphical methods articulate the assumptions under which causal quantities can be identified, while psychological experiments reveal how organisms learn from patterns of evidence and from actively changing their environments Spirtes et al. (2001) Gopnik et al. (2004). Machine-learning research extends the program to high-dimensional representations and shifting domains, where stable mechanisms may support transfer but do not emerge automatically from an accurate classifier Schölkopf et al. (2021). The shared research problem is to determine when a system's learned structure licenses predictions about new circumstances.
Experimental studies of language models show heterogeneous competence rather than one uniform ability. Corr2Cause probes transfer from correlations to identifiable causal implications; CLadder distinguishes formal causal question types; graph benchmarks examine topology and intervention direction; compositional tests probe how causal quantities propagate through increasingly complicated relationships Jin et al. (2024) Jin et al. (2023) Wang (2024) Maasch et al. (2025). Because these studies differ in input representations, model generations, and causal operations, their numerical scores should not be collapsed into a single scientific ranking. Their combined value is the converging map of error mechanisms and research controls.
The resulting methodological synthesis is a staged program of evidence. Define causal variables and assumptions, test basic identification, vary graph structure, break shortcuts, introduce intervention and context shifts, inspect compositional reasoning, and finally examine interactive transfer in new environments. At every stage, record the configuration of the artificial system and the observations that distinguish competing explanations. This approach can establish progressively stronger forms of functional causal generalization while respecting the known limits of causal identification, measurement equivalence, and representation learning Peters et al. (2016) Locatello et al. (2019) Geirhos et al. (2020).
Conclusion
Artificial cognition and causal generalization constitute a coherent research domain concerned with the ability of engineered systems to infer and reuse causal relationships when relevant conditions change. Its scientific subject is defined by observable performance under interventions and by the generative procedures that produce that performance. The historical foundations in graphical models and cognitive developmental studies, contemporary work on causal representation learning, and experimental evaluations of language models provide complementary ways of studying the phenomenon. Together they establish that meaningful assessment requires more than a model's ability to state a plausible causal explanation Gopnik et al. (2004) Schölkopf et al. (2021).
The evidence supports several specific conclusions. Formal causal knowledge depends on assumptions about structure and identifiability; representation learning faces constraints that observational prediction alone cannot remove; invariant and modular methods provide research strategies whose advantages must be independently tested; and language-model performance varies with graph topology, wording, intervention type, and compositional complexity Peters et al. (2016) Locatello et al. (2019) Gendron et al. (2024) Maasch et al. (2025). Results from one benchmark or architecture establish bounded conclusions about that tested system. Broad generalization must be demonstrated through independent challenges that separate stable mechanisms from accidental correlations.
Future progress depends on replicable designs that couple new causal situations with explicit intervention contrasts, reliable calibration, independent test generation, and transparent system-level documentation. In this framework, the Psychology of Artificial can compare biological and nonbiological cognitive functions without collapsing distinct mechanisms into a shared label. Causal generalization becomes an experimentally progressive construct: each successful transfer expands the documented range of a system's capability, while carefully characterized failures reveal the limits and computational requirements of that capability. The enduring scientific task is to explain which structures make accurate transfer possible, under what changes, and for which systems Lake et al. (2017) Woodward (2003).
References
- Arjovsky, M., Bottou, L., Gulrajani, I., & Lopez-Paz, D. (2019). Invariant risk minimization. arXiv preprint arXiv:1907.02893. https://arxiv.org/abs/1907.02893
- Barrett, D. G. T., Hill, F., Santoro, A., Morcos, A. S., & Lillicrap, T. (2018). Measuring abstract reasoning in neural networks. Proceedings of Machine Learning Research, 80, 511–520. https://proceedings.mlr.press/v80/barrett18a.html
- Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., Gulcehre, C., Song, F., Ballard, A., Gilmer, J., Dahl, G., Vaswani, A., Allen, K., Nash, C., Langston, V., ... Pascanu, R. (2018). Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261. https://arxiv.org/abs/1806.01261
- Bramley, N. R., Dayan, P., Griffiths, T. L., & Lagnado, D. A. (2017). Formalizing Neurath’s ship: Approximate algorithms for online causal learning. Psychological Review, 124(3), 301–338. https://doi.org/10.1037/rev0000061
- Geirhos, R., Jacobsen, J.-H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., & Wichmann, F. A. (2020). Shortcut learning in deep neural networks. Nature Machine Intelligence, 2, 665–673. https://doi.org/10.1038/s42256-020-00257-z
- Gendron, G., Nguyen, B. T., Peng, A. Y., Witbrock, M., & Dobbie, G. (2024). Can large language models learn independent causal mechanisms? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 6678–6701). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.emnlp-main.381
- Gopnik, A., Glymour, C., Sobel, D. M., Schulz, L. E., Kushnir, T., & Danks, D. (2004). A theory of causal learning in children: Causal maps and Bayes nets. Psychological Review, 111(1), 3–32. https://doi.org/10.1037/0033-295X.111.1.3
- Hupkes, D., Dankers, V., Mul, M., & Bruni, E. (2020). Compositionality decomposed: How do neural networks generalise? (Extended abstract). In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (pp. 5065–5069). https://doi.org/10.24963/ijcai.2020/708
- Jin, Z., Chen, Y., Leeb, F., Gresele, L., Kamal, O., Lyu, Z., Blin, K., Adauto, F. G., Kleiman-Weiner, M., Sachan, M., & Schölkopf, B. (2023). CLadder: Assessing causal reasoning in language models. Advances in Neural Information Processing Systems, 36. https://papers.neurips.cc/paper_files/paper/2023/hash/631bb9434d718ea309af82566347d607-Abstract-Conference.html
- Jin, Z., Liu, J., Lyu, Z., Poff, S., Sachan, M., Mihalcea, R., Diab, M., & Ha, D. (2024). Can large language models infer causation from correlation? In International Conference on Learning Representations (ICLR 2024). https://proceedings.iclr.cc/paper_files/paper/2024/hash/7b75a7339dfb256ee4b4bec028a6890b-Abstract-Conference.html
- Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40, e253. https://doi.org/10.1017/S0140525X16001837
- Locatello, F., Bauer, S., Lucic, M., Rätsch, G., Gelly, S., Schölkopf, B., & Bachem, O. (2019). Challenging common assumptions in the unsupervised learning of disentangled representations. Proceedings of Machine Learning Research, 97, 4114–4124. https://proceedings.mlr.press/v97/locatello19a.html
- Maasch, J. R. M. A., Hüyük, A., Xu, X., Nori, A. V., & Gonzalez, J. (2025). Compositional causal reasoning evaluation in language models. Proceedings of Machine Learning Research, 267, 42325–42365. https://proceedings.mlr.press/v267/maasch25a.html
- McCoy, R. T., & Griffiths, T. L. (2025). Modeling rapid language learning by distilling Bayesian priors into artificial neural networks. Nature Communications, 16, Article 4676. https://doi.org/10.1038/s41467-025-59957-y
- Peters, J., Bühlmann, P., & Meinshausen, N. (2016). Causal inference by using invariant prediction: Identification and confidence intervals. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 78(5), 947–1012. https://doi.org/10.1111/rssb.12167
- Schölkopf, B., Locatello, F., Bauer, S., Ke, N. R., Kalchbrenner, N., Goyal, A., & Bengio, Y. (2021). Toward causal representation learning. Proceedings of the IEEE, 109(5), 612–634. https://doi.org/10.1109/JPROC.2021.3058954
- Schulz, L. E., & Gopnik, A. (2004). Causal learning across domains. Developmental Psychology, 40(2), 162–176. https://doi.org/10.1037/0012-1649.40.2.162
- Spirtes, P., Glymour, C., & Scheines, R. (2001). Causation, prediction, and search (2nd ed.). MIT Press. https://mitpress.mit.edu/9780262527927/causation-prediction-and-search/
- Wang, Z. (2024). CausalBench: A comprehensive benchmark for evaluating causal reasoning capabilities of large language models. In Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (pp. 143–151). Association for Computational Linguistics. https://aclanthology.org/2024.sighan-1.17/
- Woodward, J. (2003). Making things happen: A theory of causal explanation. Oxford University Press. https://academic.oup.com/book/4324
- Zečević, M., Willig, M., Dhami, D. S., & Kersting, K. (2023). Causal parrots: Large language models may talk causality but are not causal. Transactions on Machine Learning Research. https://mlanthology.org/tmlr/2023/zecevic2023tmlr-causal/
- Zhou, Y., Wu, X., Wu, J., Feng, L., & Tan, K. C. (2026). CausalBN-Bench: A comprehensive benchmark for causal learning capability of LLMs. IEEE Transactions on Artificial Intelligence. Advance online publication. https://doi.org/10.1109/TAI.2026.3703427