RESEARCH DOSSIER
Latent Learning and Cognitive Maps
Abstract. Latent learning concerns knowledge acquired through experience whose existence becomes observable when circumstances change, frequently when a new incentive makes previously encountered structure useful. Cognitive maps are relational representations that support flexible navigation, inference, and action across physical or abstract task spaces. This critical scientific monograph traces unrewarded maze exploration, Tolman's place-learning and shortcut experiments, subsequent behavioral critiques, hippocampal place and entorhinal grid cells, lesion and neural recording evidence, human neuroimaging and single-neuron studies, as well as predictive-map and latent-state models. It evaluates the distinction between learning and performance, maps and learned response sequences, statistical structure and subjective awareness, spatial geometry and transition graphs, and the effects of rewards on acquisition, expression, and policy. Studies from 1929 through 2025 are synthesized with explicit attention to alternative mechanisms, representational identifiability, causal inference, and cross-species translation. The work is a critical narrative monograph of published scientific research; it reports no new experiment or meta-analysis.
Keywords: latent learning; cognitive maps; Tolman; reward-free exploration; learning versus performance; spatial navigation; place learning; shortcut inference; hippocampus; place cells; grid cells; entorhinal cortex; relational learning; latent states; successor representation; model-based inference; spatial memory; predictive maps; abstract cognitive maps; Artificial Era
The Scientific Object: Acquiring Structure Before It Is Needed
Latent learning is experience-dependent acquisition that becomes demonstrable through subsequent behavior when the conditions for expressing acquired knowledge change. The term identifies a relation between exposure, internal learning, and later performance rather than a special class of stimuli or a single cellular mechanism. A laboratory animal may explore a maze without access to a conventional food reward, yet subsequently locate a goal efficiently after an incentive is introduced. The decisive observation concerns what has been learned during earlier experience and what can be expressed later. An adequate experimental analysis therefore separates exposure history from the motivational conditions of the test, the availability of environmental information, and the behavior used to measure knowledge. The original maze experiments established this question in an experimentally productive form Blodgett (1929) Tolman & Honzik (1930) Jensen (2006).
A cognitive map is an acquired representation of relationships among places, events, states, or possible transitions that can support action in situations not exhausted by previously reinforced response sequences. A spatial map may encode routes, connectivity, approximate geometry, or relations between local landmarks; an abstract map may encode the organization of nonspatial events and task states. These meanings share the requirement that relational structure permits generalization or inference, while leaving open the representational format and neural implementation. The same organism may possess several partially overlapping representations suited to different tasks. Map-like behavior is defined operationally by transfer, flexible reorientation, and sensitivity to changed environmental structure rather than by the mere use of a familiar word Tolman (1948) Epstein et al. (2017) Behrens et al. (2018).
The research object of this dossier is the acquisition and use of exploitable environmental and task structure when immediate rewards, explicit instruction, or directly practiced responses provide an incomplete explanation of behavior. The scientific questions are how such knowledge is encoded, when it becomes evident, how incentives change its deployment, which behavioral probes discriminate competing accounts, and how neural dynamics give rise to predictive representations. The evidence spans historical animal experiments, controlled navigation, value-neutral associative learning, hippocampal recordings, human imaging and invasive recordings, and computational models. This is a critical narrative synthesis of published studies, including their uncertainties; it does not claim a new dataset, a preregistered systematic review, or a newly discovered cognitive mechanism Wikenheiser & Schoenbaum (2016) Whittington et al. (2022) Sun et al. (2025).
Learning, Performance, Motivation, and the Meaning of Latency
Learning and performance are distinct variables in experimental design. Learning is a persisting change in what an organism can predict or do under appropriate conditions; performance is the observable action produced by learning together with motivation, current policy, perceptual access, motor capacity, arousal, and task demands. The evidence for learning is necessarily indirect, because a representation is inferred from its behavioral consequences or physiological correlates. In a maze, a reward manipulation can change running speed, exploratory strategy, willingness to leave the start box, and attention to goals, independently of changes in stored environmental knowledge. Treating every performance gain as fresh learning would confound acquisition with the conditions that reveal its consequences Tolman & Honzik (1930) Jensen (2006).
The word latent describes a difference between what experience has made available and what a particular performance measure initially detects. It does not establish that the animal was inactive, that neural plasticity ceased, or that motivational processes were absent. Exploration provides visual, tactile, olfactory, proprioceptive, and action-contingent information, and those events can be informative even without food at the goal. Reward also has several roles: it changes which destinations are desirable, can alter attention and sampling, and can reinforce particular policies. The scientific problem is to estimate which role is operative on a given trial and how prior information can be used when reward contingencies change Blodgett (1929) Wilson et al. (2014) Behrens et al. (2018).
A strong demonstration would show that initial exposure created information unavailable to an otherwise equivalent control group, and that later behavior uses that information under matched testing conditions. It might involve a surprise goal relocation, a new starting position, a shortcut, or a probe in which only the topology of the experienced environment could guide a successful choice. Each design has alternatives: local landmarks can guide responses, sequential habits can mimic route knowledge, and early differences in exploratory time can be mistaken for representational superiority. The distinction between acquisition and expression is therefore a problem of controlled causal inference rather than an inference from the presence or absence of a food pellet alone Tolman et al. (1946b) Tolman et al. (1946a) Packard & McGaugh (1996).
The Classical Beginning: Blodgett and Reward Introduction
Hugh Blodgett's 1929 study investigated how the introduction of reward altered rats' maze performance after preceding opportunities to traverse the apparatus. Its historical significance was the separation of unrewarded experience from a subsequent condition in which a goal incentive was available. The design challenged a simple reading of gradual error reduction as the only visible trace of learning and raised the possibility that knowledge acquired earlier could become consequential suddenly when circumstances changed. This observation became central to the later phrase latent learning. The experiment remains historically important because its manipulation of incentive timing helped operationalize a distinction that remains fundamental in contemporary learning research Blodgett (1929) Jensen (2006).
The scientific interpretation requires attention to the historical procedure. Unrewarded rats still encountered the physical arrangement of the maze, received sensory consequences of their own movements, and faced opportunities to approach, turn, stop, and retrace. Improvement after introducing food does not isolate a specific symbolic representation or demonstrate that all forms of reinforcement were absent. It supports the narrower hypothesis that previous experience contributed to later successful behavior, provided alternative explanations involving motivation, handling, familiarity, and measurement are appropriately assessed. Reconstructing the actual control conditions is indispensable when drawing conclusions from historical experiments Blodgett (1929) Jensen (2006).
Blodgett's intervention also anticipated a contemporary causal design: expose individuals to structured information under one incentive schedule and test transfer under another. A modern version would include matched exposure dose, automated trajectory recording, movement and sensory controls, preregistered primary outcomes, and counterbalanced test orders. It would distinguish initial expression of a prelearned route from new learning that occurs immediately after reward arrives. The central contribution survives these improvements: a learning history can change future competence even when an earlier performance criterion underestimates what has been acquired Tolman & Honzik (1930) Whittington et al. (2022).
Tolman and Honzik: Rewarded, Unrewarded, and Delayed Reward Conditions
Edward C. Tolman and Charles H. Honzik's 1930 study extended the classical question by comparing maze performance under different reward histories, including the introduction and removal of reward. In the delayed-reward condition, improvement after food became available was interpreted as revealing knowledge acquired through earlier experience of the apparatus. The important methodological feature was the manipulation of when reward could influence goal-directed behavior. The authors treated the resulting performance pattern as evidence relevant to theoretical claims about expectations and learning, not simply as a measurement of a single reflex association Tolman & Honzik (1930) Tolman (1948).
These experiments became emblematic because they separated theoretical accounts that assign different functions to reward. Reward can be required by a particular learning rule, can select among already acquired relations, or can bias a policy toward particular previously observed states. Those alternatives require distinct control conditions to resolve. Unrewarded exploration is never devoid of environmental consequences: it provides conditioned sensory events and opportunities for orienting and choice. The delayed-reward finding therefore establishes a serious challenge for theories that identify every acquired environmental relation with immediate goal reinforcement, while leaving several behavior-analytic and cognitive explanations open Jensen (2006) Stachenfeld et al. (2017).
Reexamining Tolman and Honzik is especially useful because the historical experiment is often reduced to a simplified textbook demonstration. The primary source provides a specific experimental design and specific dependent measures; subsequent descriptions sometimes generalize from that design to broad declarations about all behaviorism or the necessity of mental maps. A disciplined historical reading retains the study's operational insight and assesses later interpretations against control adequacy, model flexibility, and the observability of the proposed representation. This approach preserves the experiment's significance while sharpening the question that modern work must answer: what structure was learned and how can the representation be identified independently of the behavior it explains Tolman & Honzik (1930) Jensen (2006) Behrens et al. (2018).
Place Learning, Response Learning, Orientation, and Shortcuts
The 1946 studies by Tolman, Ritchie, and Kalish made the representational claim more experimentally specific. The place-learning versus response-learning contrast asked whether success was organized around arriving at a spatial destination or repeating a practiced turn or response. The distinction matters because paths to the same location may require different movements from different starting positions. A successful transfer when the response sequence must change can indicate sensitivity to relations among locations, although cue use and alternative route rules must still be assessed. Experimental manipulations of the start position and available environmental information turn a broad map metaphor into a set of discriminable predictions Tolman et al. (1946b) Packard & McGaugh (1996).
The orientation and shortcut experiments probed whether animals could approach a goal using routes other than the path reinforced or traveled during training. Such tests ask whether knowledge is compositional: can the subject use relations learned during one set of trajectories to organize an unfamiliar trajectory? An efficient novel shortcut is suggestive when it genuinely requires integrating separately experienced segments and when local beacon strategies, odor trails, and perceptual visibility do not solve the problem. A valid shortcut test must also characterize the possibility of exploratory choices that succeed by chance, differences in distance, and the specific distribution of alternatives provided Tolman et al. (1946a) Epstein et al. (2017).
Place and response learning can coexist, and their relative contributions depend on training, task design, and the state of neural circuits. Later hippocampal and caudate manipulations demonstrate that different forms of guidance can be expressed through distinct, interacting systems. The proper scientific inference is therefore comparative: which task condition favors which kind of learned control, and how does an intervention alter that balance? This formulation links classical maze results to modern causal neuroscience without claiming that the rats of the 1940s possessed the same fully specified computational representation proposed in current models Packard & McGaugh (1996) Morris et al. (1982) Wikenheiser & Schoenbaum (2016).
Tolman's Cognitive Maps and the Historical Argument Over Explanation
Tolman's 1948 article, Cognitive Maps in Rats and Men, gave the research tradition its most durable conceptual expression. A cognitive map described organized environmental knowledge extending beyond the motor chain that happened to produce reinforcement. The proposal integrated evidence about place, routes, spatial expectations, and the behavioral consequences of environmental history. Its historical importance lies in framing navigation as a problem of learned structure and flexible purposive behavior. The metaphor subsequently acquired distinct experimental, computational, and neural meanings, making retrospective precision essential when citing Tolman as an ancestor of present models Tolman (1948) Behrens et al. (2018).
The common historical story that a single maze experiment refuted all behavioral theory misrepresents the diversity of those theories. Jensen's analysis of textbook accounts documents recurring errors concerning reinforcement, Skinner's position, and the inference that a cognitive map must be the unique explanation of latent learning. Some behaviorally framed explanations can accommodate learning without the particular reward used at test; conditioned stimuli, response patterns, environmental feedback, and changes in stimulus control deserve explicit consideration. The decisive scientific question is whether a candidate model predicts new-start, shortcut, revaluation, or transition-change performance under matched constraints better than its alternatives Jensen (2006) Tolman (1948).
The field's enduring progress comes from converting the map metaphor into measurable proposals. A representation might preserve metric distance, connectivity, directed transition probability, expected future occupancy, or abstract latent state identity. These representations agree for many familiar tasks and diverge under carefully designed probes. Modern research can therefore honor Tolman's insight through model comparison rather than through rhetorical opposition between historical schools. The issue is identifying what information is represented, how it is acquired, which calculations use it, and which evidence could make the explanation fail Stachenfeld et al. (2017) Wilson et al. (2014) Whittington et al. (2022).
Exploration as Structured Sampling of the Environment
Exploration changes what an organism can know because actions determine which sensory observations become available. Traversing a corridor reveals whether another passage is connected, whether a landmark remains visible from two locations, and whether a familiar sequence of turns returns to the starting place. Such information can be acquired in the absence of an immediately consumable reward. The relevant variables include path coverage, repeated visits, temporal adjacency, uncertainty, exploratory motivation, and the correspondence between proprioceptive movement and external landmarks. A critical experiment must measure what was sampled, because unequal exposure may produce apparent differences in learning even when the same computational mechanism operates across groups. The early maze studies introduced the puzzle; contemporary mapping theories articulate why the data generated during free exploration may be valuable later Blodgett (1929) Tolman & Honzik (1930) Whittington et al. (2022).
Active sampling creates a coupling among behavior, perception, and knowledge. An animal that consistently turns at the same junction may acquire a response habit and simultaneously encounter the landmark configuration of the surrounding area. A behavioral record can therefore support several representational hypotheses. Tracking only total errors discards the temporal structure of exploration, while continuous trajectories, head orientation, local viewing, and revisitation patterns can reveal what information was available when a choice occurred. The observational advantage becomes most compelling when later tests are designed to make different learning histories predict different first choices Epstein et al. (2017) Wikenheiser & Schoenbaum (2016).
In animals, hunger, novelty, stress, habituation, and social or handling conditions can affect exploratory coverage. In humans, task instructions, prior knowledge of maps, attention to navigation goals, and assumptions about future testing can influence what is learned during apparently incidental exposure. Reward-free exploration is consequently an experimentally specified incentive condition rather than a psychologically contentless baseline. The scientific aim is to determine how structural information becomes encoded and later available for decision making, given the actual incentives and opportunities present in the exposure environment Jensen (2006) Behrens et al. (2018).
Topological Relations, Metric Geometry, and Directional Knowledge
A navigational representation can preserve different properties of space. Topological structure specifies which places or states are connected; metric structure specifies distances, angles, and geometric relations; directional structure constrains how orientation transforms local movement into a route. A subject may possess accurate graph connectivity yet estimate distances poorly, or navigate by landmarks while showing weak knowledge of global orientation. These dissociations matter because a shortcut may require Euclidean geometry in one apparatus and merely knowledge of an untraveled graph edge in another. A cognitive map should therefore be characterized by its representational content and the transfer test used to infer that content Epstein et al. (2017) Behrens et al. (2018).
Local cues and global relations may cooperate. A landmark can identify a current node, a learned transition can predict which node follows an action, and a global spatial relation can support a path that has never been practiced. Conversely, perceptual similarity can produce ambiguous state assignments when separate locations look alike. This is the setting in which a latent-state model becomes valuable: it can distinguish situations on the basis of preceding transitions or contextual history even when immediate observations are identical. Modern models formalize this problem, while hippocampal data provide clues about how learning differentiates similar sensory states with different task functions Stachenfeld et al. (2017) Sun et al. (2025).
Tests of map geometry demand particular controls. A straight-line choice can reflect an obvious visual target; angular pointing may reflect memorized headings; distance estimates can be biased by walls, texture, and the participant's movement speed. Researchers can improve discriminability by manipulating familiar routes, inserting barriers, removing visual beacons, and changing the starting orientation. A pattern of transfer across several such tests is stronger evidence for relational knowledge than a single navigational success. The resulting representation can be described precisely as a topology, metric approximation, sequence graph, or predictive relation without requiring every subject to encode an ideal cartographic map Tolman et al. (1946a) Epstein et al. (2017) Whittington et al. (2022).
Novel-Start, Detour, and Shortcut Tests of Flexible Knowledge
Flexible navigation is experimentally informative because the subject must respond to circumstances that differ from the learning episode. A new-start probe changes the relation between initial heading and goal; a detour probe closes a previously useful path; a shortcut probe offers a potential route whose component relations were experienced without the full action sequence. Each manipulation creates a contrast among place-directed, response-directed, and local cue strategies. The inference is strongest when the change is introduced before the animal has sufficient time to learn the new route through trial and error. An early, appropriately controlled choice can then index transfer from the prior exposure history Tolman et al. (1946b) Tolman et al. (1946a).
A detour is only diagnostic if the animal has an alternative that genuinely requires reorganizing route selection. Following a wall or a familiar odor trail can also deliver a successful path. Likewise, apparent shortcut-taking may result from perceptually detecting an exposed goal from the test position. Experimental design must therefore specify obstacle placement, line of sight, route novelty, and the opportunity to revisit learned corridors. Comparisons with yoked animals given matched sensory exposure and differing navigational control can help estimate the contribution of active exploration. The problem is one of excluding strategically equivalent explanations through precise manipulations Jensen (2006) Epstein et al. (2017).
The same logic applies beyond physical mazes. In a structured choice task, learners may encounter a series of value-neutral associations and later need to select an outcome that was never paired directly with an original cue. Successful immediate inference can show the use of relationships established earlier. Here the novel-start or shortcut is conceptual rather than spatial: it links previously separate events through an inferred path. Contemporary human sensory-preconditioning and abstract-map experiments extend this general principle, while emphasizing that successful transfer still permits comparison among memory retrieval, chained associations, and explicit state-space computations Wang et al. (2020) Garvert et al. (2017) Courellis et al. (2024).
Reward Timing, Revaluation, and the Expression of Learned Knowledge
Introducing reward after exposure is a useful manipulation because it changes the value of a known destination while leaving the earlier environmental experience fixed. The decisive analytical question is whether an improvement is disproportionately fast relative to the learning available after reward introduction. Testing should therefore emphasize the earliest informative decisions, ideally before repeated rewarded trials permit substantial new acquisition. A design that compares only late asymptotic performance may obscure the latent contribution because both previously exposed and naive subjects can eventually learn a reinforced route. The historical delayed-reward studies anticipated this distinction, and later models make it explicit through separate learning and action-selection mechanisms Blodgett (1929) Tolman & Honzik (1930) Stachenfeld et al. (2017).
Outcome revaluation can reveal whether a subject represents specific consequences or follows a relatively autonomous response pattern. If two locations have been explored and one subsequently becomes preferable, flexible choice may depend on knowing which paths reach the newly valued location. A map-based account predicts sensitivity to a changed goal without requiring the exact route to be trained under that goal value. Yet revaluation effects can also be mediated by learned cue relationships and retrieval strategies. Observing goal sensitivity is therefore informative about action control, while the structure of novel-route behavior determines whether the evidence supports spatial or more general relational mapping Packard & McGaugh (1996) Wilson et al. (2014).
Reward sensitivity also constrains neural interpretation. Place and entorhinal signals can vary with experience and behavioral policy; cortical state representations may encode task relevance as well as nominal spatial location. A neural population that changes when reward moves is not thereby proved to lack a map, because value can reorganize which regions of a structured environment deserve predictive resolution. Conversely, a stable code under reward changes does not automatically prove that the organism uses it for action. Combining reward revaluation, representational recording, and causal perturbation makes claims about acquisition, policy, and expression more discriminable Stachenfeld et al. (2017) Schuck et al. (2016) Wikenheiser & Schoenbaum (2016).
Incidental Relational Learning, Statistical Regularities, and Inference
Structured experience can teach relations even when the organism is not instructed to discover them. In sequences of events, some items follow one another more frequently than others, and their temporal communities define latent structure. Schapiro and colleagues examined how event representations reflected such temporal organization, extending map-like reasoning beyond locations. The research question is whether learners encode individual transitions, higher-order communities, or a representation that permits predicting indirectly related events. These possibilities can be contrasted using carefully selected novel pairs whose direct co-occurrence histories are equivalent but whose positions in the inferred structure differ Schapiro et al. (2013) Garvert et al. (2017).
Sensory preconditioning makes the logic especially clear. Participants first experience relationships between neutral cues; only afterward are some cues paired with outcomes. The test asks whether value can be inferred for a cue that was never directly rewarded. Wang and colleagues used this design and pattern-based imaging to study interacting hippocampal and orbitofrontal representations. Their findings support the view that value-neutral associations and later outcome inference involve partly overlapping systems. The study does not establish that every successful inference requires a globally metric map, yet it directly examines the central latent-learning question under a transparent sequence of exposure, valuation, and choice Wang et al. (2020) Wilson et al. (2014).
Abstract relational codes can also support generalization when a familiar task variable occurs in a new combination. The diagnostic feature is the transfer of a relation across contexts, not simply correct classification within a training distribution. Human neuron recordings during inferential tasks indicate that hippocampal activity can acquire separable representations of observable and inferred task variables as behavior improves. Such data connect latent learning to abstraction while requiring care about patient sampling, task-specific demands, and the causal interpretation of representational geometry. An abstract code is an empirical property of measured responses and decoding transfer, rather than direct access to a participant's subjective model Courellis et al. (2024) Qiu et al. (2024).
Place, Response, and Strategy Competition Across Learning Histories
A single task can support multiple viable control strategies. A spatial strategy may estimate a goal's location relative to external cues, while a response strategy repeats a turn policy. The relative success of each depends on start positions, training length, cue stability, task demands, and the information accessible to the subject. This multiplicity explains why cognitive-map behavior should be measured as a family of signatures rather than assumed from ordinary route proficiency. A subject that uses a response strategy under familiar conditions can nevertheless possess relational knowledge that becomes useful under a detour, and a subject demonstrating a place bias may still rely heavily on conspicuous local cues Tolman et al. (1946b) Packard & McGaugh (1996).
Evidence from selective inactivation of hippocampal and caudate systems indicates that expressions of place and response learning depend partly on dissociable neural contributions. Such perturbations offer stronger mechanistic leverage than behavioral performance alone, although they can influence attention, locomotion, or memory retrieval in addition to representation. Timing matters: an intervention during exposure tests acquisition and possibly exploration, whereas an intervention during the later probe targets expression and associated executive operations. An interpretable experiment specifies which causal link it aims to manipulate, rather than using the general phrase spatial learning impairment as a single explanatory category Packard & McGaugh (1996) Morris et al. (1982).
Strategy competition also has consequences for behavioral generalization. An animal trained extensively on one fixed route may show superior speed in that particular environment but less adaptability than another trained across diverse start positions. Adaptive competence therefore requires evaluating the learning history and the distribution of transfer environments, not merely measuring acquisition on one route. Such observations motivate predictive and relational models in which goals and transition knowledge can be recombined, while leaving space for procedural systems that are efficient when circumstances remain stable. The resulting theory connects latent learning to broader learning psychology through clearly specified tasks and dependent variables Behrens et al. (2018) Stachenfeld et al. (2017) Whittington et al. (2022).
Place Cells and the Hippocampal Origins of Neural Mapping
O'Keefe and Dostrovsky's 1971 report described hippocampal neurons whose activity depended on the location of freely moving rats. This finding provided a new empirical level for the cognitive-map hypothesis: investigators could examine whether neural activity was organized by where an animal was in an environment instead of inferring all spatial relations from routes or errors. A place field is a region where a neuron has an elevated probability of firing under particular conditions, and the corresponding spatial tuning can be studied over changes in movement, context, cues, and experience. The initial report was explicitly preliminary, and later work developed extensive methods for separating spatial coding from behavioral covariates O'Keefe & Dostrovsky (1971) Epstein et al. (2017).
Place-cell evidence makes the possibility of a population representation tangible, but spatial selectivity alone does not identify the complete contents of a cognitive map. The same cell may change its firing with environmental context, behavioral phase, or relevant goals. A localized firing pattern can reflect sensory inputs, learned associations, predictive relationships, and internally estimated position. Inferring a full map requires establishing how activity across cells preserves relations among states, whether those relations are stable enough for transfer, and how downstream circuits use them to guide action. Computational models articulate alternative hypotheses for why a field appears where it does and how a field may change with experience Stachenfeld et al. (2017) Wikenheiser & Schoenbaum (2016).
Neural mapping also introduces an acquisition question that mirrors latent learning. If a population develops stable environmental structure during exploration before a task-specific reward becomes available, the representation offers a candidate substrate for later flexible performance. Demonstrating this sequence requires longitudinal recordings and a test in which later choice depends on the prior code. Correlations between spatial firing and route success are informative, but the critical explanatory step concerns whether a particular learned representation predicts or causally contributes to an unpracticed choice. The distinction between representation, readout, and policy remains central across behavioral and neural approaches O'Keefe & Dostrovsky (1971) Sun et al. (2025) Whittington et al. (2022).
Entorhinal Grid Codes, Orientation, and Spatial Metric Signals
Hafting and colleagues' 2005 study identified a striking periodic organization in the spatial firing of neurons in the entorhinal cortex. Grid cells exhibit multiple firing fields whose arrangement can approximate a triangular lattice over an explored space. This physiological finding broadened the neural mapping problem beyond location-specific place fields and suggested a candidate basis for representing displacement, metric relations, or structured state coordinates. The relationship among place cells, grid cells, boundary and heading signals is a problem of circuit dynamics and environmental inference; the presence of periodic tuning does not by itself establish a complete survey-like representation Hafting et al. (2005) Whittington et al. (2022).
Human imaging work reported gridlike directional modulation in memory-related networks during virtual navigation. Doeller and colleagues supplied evidence compatible with a grid-related signal in humans, although functional magnetic resonance imaging averages over many neurons and cannot establish the existence of any particular single-cell code in the individuals studied. Experimental design and inference must therefore distinguish a population-level signature consistent with grid organization from direct cellular recordings. Directional anisotropy, behavioral sampling, and analysis choices can influence such signals; converging methods are essential before attributing a human map to a specific neuronal lattice Doeller et al. (2010) Epstein et al. (2017).
The theoretical utility of a gridlike code extends to relational coordinates. If a neural system represents dimensions of a task in a reusable organization, learning can support extrapolation beyond a practiced path. Nevertheless, the mapping from a recorded neural geometry to the geometry of actual choices must be demonstrated. The same pattern may provide a basis for estimating distance, organizing experiences, or predicting transitions under a particular task distribution. Treating each function as a separate hypothesis clarifies what a periodic code explains and what remains to be inferred from behavioral experiments Constantinescu et al. (2016) Garvert et al. (2017) Behrens et al. (2018).
Human Single-Neuron Evidence for Latent Temporal and Conceptual Structure
Tacikowski and colleagues recorded single neurons from the hippocampal and entorhinal system of patients undergoing clinically indicated intracranial monitoring while the participants encountered structured sequences of images. The published results describe neural activity that progressively reflected relationships among events and the probabilistic temporal organization of the sequence, even without explicit instructions to discover that structure. Representation persisted into later testing, and replay-related analyses linked experience to the extraction of temporal relationships. The scientific importance is direct access to neural signals carrying a structured predictive code in humans, in a paradigm whose underlying organization was not a geometric floor plan Tacikowski et al. (2024).
Courellis and colleagues recorded human neurons during inferential reasoning and reported hippocampal representations that separated observed from inferred latent task variables. Abstract representational geometry emerged alongside learning and related to performance on inference trials. Such findings advance the field beyond a map defined solely by spatial position, because the relevant neural dimension tracks structure discovered through the task rather than the coordinates of physical movement. The evidence remains specific to the sampled patients, analysis, and inferential task, and the presence of a decodable representation cannot by itself establish the complete computations linking it to behavior Courellis et al. (2024) Qiu et al. (2024).
Single-neuron recordings complement behavioral and imaging evidence by resolving individual-cell dynamics, but neurosurgical recruitment creates sampling constraints and electrode placement is dictated by clinical circumstances. The ability to generalize from particular patients, brain regions, recording sessions, and chosen abstractions requires independent studies with converging measures. Computational decoding also must be evaluated against overfitting, confounds in trial order, and the stability of latent-variable estimates. The strongest synthesis treats these recordings as unusually direct measurements of acquired task structure while preserving the need for causal intervention and external replication across populations and environments Tacikowski et al. (2024) Courellis et al. (2024) Behrens et al. (2018).
Longitudinal Hippocampal Learning and the Emergence of State Structure
Sun and colleagues' 2025 study used longitudinal two-photon calcium imaging to track large numbers of hippocampal CA1 neurons while mice learned reward-collection tasks in virtual environments that had similar sensory features but differed in task structure. Over learning, the population codes became increasingly differentiated in ways consistent with distinct inferred task states. The study matters because it tracked how the representation changed over time, rather than only comparing a well-learned state with a baseline. These recordings linked increasing behavioral competence to progressive reorganization of neural state information Sun et al. (2025).
The authors compared neural changes with computational models and highlighted a clone-structured causal graph model as a close account of the reported learning trajectory. This model can assign different latent states to perceptually similar observations when their surrounding transitions imply different task positions. The finding brings together a longstanding psychological concern—how organisms learn more than the response currently demanded—and a quantitative neural hypothesis about differentiating hidden structure. Model similarity is evidence for explanatory fit under the tested conditions, and remains separate from demonstrating that the brain literally executes the model's particular inference algorithm Sun et al. (2025) Whittington et al. (2022).
Contemporary experiments therefore move toward dynamic theories of map construction. A useful next step manipulates transition probabilities, changes hidden state identity while preserving visual appearance, and assesses whether neural codes and novel choices adapt according to the same estimated structure. Distinguishing encoding of current position, expected future occupancy, and learned task phase is crucial because those quantities can correlate during stable training. Longitudinal measurement, careful causal controls, and out-of-distribution tests can separate alternative accounts more effectively than attributing every well-structured hippocampal pattern to one undifferentiated cognitive map Sun et al. (2025) Stachenfeld et al. (2017) Tacikowski et al. (2024).
Environmental Structure as a Graph of States and Transitions
A computational description begins by specifying states, observations, actions, and possible transitions. A graph represents states as nodes and experienced or possible relations as edges; transition probabilities indicate how likely an action or event is to move the system from one state to another. In navigation the nodes may correspond to locations or local contexts, whereas in a conceptual task they may encode combinations of stimulus properties and hidden task phases. This distinction prevents a model from confusing the geometry of physical movement with the organization of events that only happen to be learned through movement. A learned graph permits questions about reachability, paths, and predicted consequences before a currently valued goal has been directly sought Behrens et al. (2018) Whittington et al. (2022).
Real observations rarely identify a state uniquely. Two corridors may have the same visual appearance yet lead to different destinations; the appropriate choice depends on a preceding turn or route segment. The learner must therefore infer latent states from sequences, context, and uncertainty. One family of accounts represents beliefs over possible hidden states; another learns discriminative features or memory traces sufficient for the same predictions. A state model can be evaluated by changing transition structure while holding the immediate sensory scene fixed. If the resulting first choices and neural responses match the model's predictions, the evidence becomes stronger than a simple correlation between familiarity and performance Sun et al. (2025) Tacikowski et al. (2024).
Graph descriptions are mathematical tools rather than proof that a brain contains explicit nodes connected by labeled edges. Different representations can produce the same observable transition predictions, and a learner can approximate graph-based planning without storing a complete symbolic graph. Scientific evaluation therefore requires specifying the observable distinctions between candidate models, estimating uncertainty in fitted parameters, and testing predictions on withheld conditions. The central map claim gains empirical meaning when learned structural information is reusable across goals, actions, or contexts in ways that local response rules cannot reproduce under the same constraints Stachenfeld et al. (2017) Behrens et al. (2018) Whittington et al. (2022).
Predictive Maps and the Successor Representation
The successor representation describes a state's expected discounted future occupancy under a policy. For a set of states and a given policy, it captures how often each other state is expected to occur in the future, with nearer future visits weighted more heavily. The familiar expression M = (I − γP)⁻¹, when its mathematical assumptions hold, relates a state-transition matrix P and a discount factor γ to expected occupancy. The representation combines knowledge of transition structure with predictions about future visitation. It can support rapid evaluation of a newly valued goal when an appropriate reward function changes while transition expectations remain sufficiently stable Stachenfeld et al. (2017) Momennejad et al. (2017).
Momennejad and colleagues examined behavioral evidence for successor-like representations in human reinforcement learning. The experimental contribution was to compare predictions of several learning accounts under changes in rewards and transitions, thereby connecting a computational formalism with patterns of generalization. The successor representation is an informative intermediate account: values can be recomputed quickly after some reward revaluations, whereas changes in transitions may require updating predictive occupancy. The relative costs depend on the learning rule, task structure, and planning allowed at test. A computational label is therefore justified by diagnostic comparisons, not by the generic observation that a participant performed a flexible choice Momennejad et al. (2017) Behrens et al. (2018).
Stachenfeld and colleagues proposed a predictive-map interpretation of hippocampal activity that links place-cell-like coding to expected future occupancy rather than to physical coordinates alone. This theoretical contribution helps explain why spatial fields may depend on behavioral policy, environmental boundaries, and learning history. It also creates falsifiable questions: how should a field change after a transition is removed, after a policy changes, or after new routes become probable? Evaluating such predictions requires independent neural data and careful matching of occupancy and action variables. The account is a productive explanatory model, and its success should be assessed relative to alternatives that also predict changing spatial tuning Stachenfeld et al. (2017) Whittington et al. (2022).
Model-Based Inference, Cached Control, and Flexible Revaluation
Learning theories distinguish several computational routes from experience to action. A model-based learner can use estimated transitions and outcome consequences to evaluate actions, a cached-value learner can adapt values through experience, and a successor-like learner can use predictive occupancy to partly separate environmental dynamics from reward. These accounts differ in their predicted response to changes in reward, transitions, and available planning time. An experiment that changes only the reward function cannot, by itself, distinguish every possible map-based representation from memory retrieval or cached policies with extensive training. A well-designed battery manipulates the relevant dimensions independently and asks which model predicts transfer beyond the training distribution Momennejad et al. (2017) Wilson et al. (2014) Behrens et al. (2018).
Sensory preconditioning is a particularly direct link between latent structure learning and inferred value. Neutral cue sequences are experienced before one cue becomes associated with a consequential outcome, making later evaluation of a previously neutral predecessor informative about integration across learning phases. Wang and colleagues found patterns of orbitofrontal and hippocampal involvement consistent with relations learned during preconditioning and used during subsequent inference. The neural findings differentiate the encoding of relations from the later representation of inferred outcomes. They also show why a theory that treats every observed choice as the isolated product of the most recent reinforcement misses important information about prior experience Wang et al. (2020) Wikenheiser & Schoenbaum (2016).
Cached and relational control may cooperate within an individual. A trained route can be executed efficiently without deliberate recomputation until a barrier or goal change creates a need for replanning. Repeated experience can shift which strategy is expressed, while the environmental representation that supports a rarely used alternative remains available. The scientific task is to measure such strategy recruitment rather than assigning an organism permanently to a single algorithm. Model fits should account for choice stochasticity, action costs, spatial exposure, and the possibility that participants mix strategies across trials Packard & McGaugh (1996) Stachenfeld et al. (2017) Schuck et al. (2016).
Relational Generalization, Abstraction, and Transfer Across Domains
A map-based account gains scope when it predicts successful transfer to a new domain, context, or combination of features. In spatial navigation, the relevant transfer might concern a newly available route; in a conceptual task, it might concern an unobserved relation between two states; in a sequence task, it might concern the expected order of events after an unusual cue. These cases share a computational demand: the learner must reuse relationships acquired during previous experience to address a changed problem. Their differences in memory, perception, temporal scale, and intentional strategy remain empirically significant Garvert et al. (2017) Constantinescu et al. (2016) Tacikowski et al. (2024).
Experimental evidence for abstraction depends on the exact definition of invariance. A neural code may be called abstract when a decoder trained on one combination of variables generalizes to another while preserving the relevant task dimension. Courellis and colleagues used such representational logic to investigate inferred latent variables in human hippocampal cells. This is more stringent than observing an increase in firing during successful trials, since it demands structured generalization across conditions. Even so, the decoder's success establishes a measurable property of neural activity; the causal link to planning must be examined with further perturbation and behavioral tests Courellis et al. (2024) Qiu et al. (2024).
Across species and paradigms, relational generalization may involve geometric coordinates, graph distances, predictive occupancy, compressed factors, or sequence-dependent hidden states. No single format is required by the empirical fact that a new problem can be solved. Theories can nonetheless make differentiated predictions about how transfer degrades with environmental distortion, variable remapping, delay, or changes in statistical regularities. Research progresses when experiments move from labeling every flexible response a map to measuring the structure and limits of the generalization that the task actually elicits Behrens et al. (2018) Whittington et al. (2022) Sun et al. (2025).
Artificial Systems as Comparative Models of Environmental Learning
Artificial agents provide a useful experimental laboratory for comparing learning rules under precisely controlled environments. A system can acquire a transition model from exploratory interaction, learn representations from unlabeled sequences, or approximate expected future occupancy before any externally specified goal is introduced. When a reward function later becomes available, its policy may exploit prior knowledge in a way formally analogous to latent learning. Researchers can manipulate the exact graph, partial observability, memory capacity, and training data to study the conditions that generate flexible transfer. This computational analogy concerns observable performance and algorithmic structure; claims about psychological experience or a particular biological implementation require independent evidence Stachenfeld et al. (2017) Whittington et al. (2022).
Model comparison becomes especially informative when candidate artificial systems are challenged with perceptually aliased states and altered transition structures. A system that memorizes a fixed stimulus-to-action mapping may perform well under familiar input statistics yet fail when the same sensory scene has a changed hidden meaning. A latent-state model can integrate recent history to maintain alternative state estimates and select actions accordingly. The study by Sun and colleagues offers a concrete example of comparing the developmental trajectory of hippocampal population codes with candidate computational algorithms rather than simply matching their final performance. A shared pattern of learning constrains models, while convergence on the same task does not imply that biological and artificial systems have identical internal mechanisms Sun et al. (2025) Whittington et al. (2022).
Within the horizon of the Artificial Era, the scientific relevance of cognitive maps concerns the general problem of how a reasoning system acquires a usable world structure. A machine may construct an explicit graph, a learned latent representation, or a predictive state model; a biological learner may combine hippocampal, cortical, and procedural circuits. Their comparison becomes rigorous when the researcher states a common task, measurable capacities, training exposure, transfer conditions, and falsifiable predictions. Questions about agency, identity, or consciousness are separate research problems and should be investigated by their own operational criteria. Here the central scientific question is flexible learning from structure before the specific future need is known Behrens et al. (2018) Sun et al. (2025) Courellis et al. (2024).
Measurement, Construct Validity, and the Logic of Evidence
The empirical construct latent learning requires an observable improvement that is attributable to earlier exposure rather than solely to newly available incentives or later practice. An informative protocol defines the learning phase, exposure opportunities, incentive manipulation, and the first test decision before data collection. It also measures the information available during exposure, including visited nodes, route diversity, environmental cues, and time spent exploring. Because learning is inferred from later expression, a useful analysis estimates how much earlier exposure predicts transfer while controlling current motivation and motor capacity. A single reduction in maze errors is an outcome measure, not an operational definition of the representation that caused it Tolman & Honzik (1930) Jensen (2006) Epstein et al. (2017).
Cognitive-map claims introduce a second validity challenge: the structure attributed to the learner must be distinguished from simpler explanations with equivalent observed predictions. A shortcut success can arise from visual target detection, graph connectivity, metric distance computation, or an opportunistic search policy. A decoder's ability to predict location from neural activity may reflect spatial representation, but can also exploit correlated heading, movement speed, reward expectation, or the sensory sequence that accompanies position. Validity improves when a task includes critical probes in which models predict contrasting choices and neural patterns, and when decoding is cross-validated across genuinely new conditions Stachenfeld et al. (2017) Schuck et al. (2016) Whittington et al. (2022).
Constructive operationalization specifies at least three independent properties: acquisition without the later focal reward, storage or maintenance of relational information, and deployment in a changed problem. Their combination is stronger than any one considered alone. A robust design can establish that a structure was encountered, estimate whether it became decodable before reward introduction, and examine whether that representation predicts a surprising choice after revaluation. Nevertheless, a correlational bridge among these measurements does not yet prove the representation mediates behavior. Mechanistic claims require causal manipulation or especially stringent interventions that isolate acquisition, retrieval, and decision functions Wang et al. (2020) Courellis et al. (2024) Sun et al. (2025).
Study Design, Sampling, and Cross-Species Interpretation
Animal maze experiments allow precise environmental manipulation and access to neural recording or perturbation, but behavior depends on species-specific sensory ecology and locomotor repertoire. Rodents can use odor traces, whisker-mediated contact, distal visual landmarks, and path-integration cues in combinations that differ from human virtual-reality navigation. A controlled apparatus must disclose and, where feasible, equate these information sources. Standardizing the layout alone does not ensure equivalent psychological tasks. Task design should therefore distinguish the ecological problem actually solved by the animal from the experimenter's preferred description of that solution Morris et al. (1982) O'Keefe & Dostrovsky (1971) Epstein et al. (2017).
Human studies offer explicit reports, complex abstract tasks, and noninvasive imaging, alongside distinctive confounds from verbal reasoning, prior experience, device familiarity, and participant expectations. Virtual navigation can restrict vestibular and body-motion signals, while an abstract graph task can be solved using strategies that were never intended to resemble physical navigation. Invasive human studies provide neural resolution but are constrained by clinical recruitment and electrode coverage. A valid cross-species claim is generally functional and task-specific: different organisms show related patterns of relational learning under comparable information and transfer conditions, without requiring identical strategies or neural codes Doeller et al. (2010) Tacikowski et al. (2024) Courellis et al. (2024).
Transfer across studies also depends on statistical power and sampling. Learning trajectories are heterogeneous; some individuals acquire a structure quickly, others accumulate gradual evidence, and others rely on a strategy that is sufficient only during the training task. Aggregating all trials across individuals can produce a smooth population curve that describes no single learner accurately. Hierarchical models, explicit strategy-mixture analyses, held-out prediction, and preregistered transfer contrasts can improve estimation. Reported effect sizes, uncertainty intervals, and missing-data handling are necessary when the literature is used to compare strength of evidence, even where a narrative synthesis does not pool studies quantitatively Momennejad et al. (2017) Schuck et al. (2016) Behrens et al. (2018).
Alternative Explanations, Historical Disputes, and Replication
A primary alternative to an integrated cognitive map is a set of learned stimulus-response or cue-action relationships capable of producing successful choices. Other alternatives include chained associations, procedural search heuristics, localized landmark recognition, path integration, and the retrieval of particular episodes. These accounts can be psychologically sophisticated and should be specified in their strongest testable forms. An evaluation that compares a richly parameterized mapping model against an artificially restricted response rule will overstate the uniqueness of the map hypothesis. Jensen's historical analysis illustrates why the field benefits from explicit attention to the actual claims and flexibility of competing behavior theories Jensen (2006) Tolman (1948).
Neural evidence raises analogous concerns. Spatially tuned cells, gridlike population signals, and task-state decoding can be compatible with several algorithms, especially when behavior samples only a narrow range of states. The fact that an artificial model reproduces a spatial field does not show that the animal implements that model; the fact that a model fails under one training setup does not establish impossibility under alternative parameterizations. Stronger comparisons ask candidate models to predict the direction and timing of changes after environmental restructuring, a novel cue, or a change in reward policy. Replication of these predictions in independent laboratories and tasks is essential to claims of a general mechanism Stachenfeld et al. (2017) Whittington et al. (2022) Sun et al. (2025).
Historical and recent findings must also be weighed by the evidence actually inspected. Classical papers may provide only limited reporting of modern bias controls, while recent articles may include complex analytical pipelines whose specifications condition the result. A historical experiment establishes what its methods measured, not a modern simulation of the investigator's intentions. A contemporary brain-imaging paper establishes effects supported by its acquisition and statistical approach, with broader neural interpretations requiring convergence. This synthesis therefore privileges directly verifiable outcomes and identifies explanations that remain model-dependent, rather than treating age, journal prestige, or conceptual appeal as substitutes for critical evaluation Blodgett (1929) Jensen (2006) Tacikowski et al. (2024).
Development, Individual Variation, and Ecological Generalization
A representational account of latent learning has implications for development and individual variation, because acquiring structural knowledge requires opportunities to explore, retain relations, and select them under changing goals. Yet the core classical evidence is concentrated in specific animal tasks and later laboratory studies with constrained environments. Broad claims about age trends, clinical groups, educational performance, or stable personality differences require dedicated evidence rather than extrapolation from adult maze experiments. A developmental program could independently measure navigation exposure, representational transfer, executive demands, and motivational sensitivity across age groups, assessing which component explains observed differences Epstein et al. (2017) Behrens et al. (2018).
Natural navigation is richer than most experimental environments. An individual may learn from landmarks, social guidance, written maps, repeated routines, and occasionally from surprising route closures. The environmental graph can change over time; familiar pathways can disappear, landmarks can become unreliable, and new goals can expose structural gaps. Ecological validity therefore depends on tests that approximate real uncertainty and dynamically changing opportunities, not solely on whether a task presents a visually realistic street. A research program should establish how well laboratory transfer measures predict navigation and inferential choices outside the original apparatus Epstein et al. (2017) Wikenheiser & Schoenbaum (2016).
Individual differences also expose the limits of one-dimensional outcome scores. Two people can reach the same goal with different representational mixtures, one using a sequence habit and another a flexible relational plan. One can have accurate connectivity information but imprecise distances; another can encode geometry yet fail to choose when incentives are ambiguous. Multimethod measurement should distinguish strategy, representational accuracy, confidence, time cost, and adaptability to changes. Such decomposition gives latent learning a scientifically useful profile rather than assuming that all competent navigation expresses the same map-like faculty Packard & McGaugh (1996) Garvert et al. (2017) Whittington et al. (2022).
Scientific Applications, Education, and Research Translation
The classical finding that useful knowledge can be expressed after a change in incentives has relevance for instructional design and adaptive training. People often encounter structured information before knowing which part will matter later; later goals can reveal whether that exposure made an actionable representation available. A cautious application is to design learning settings with opportunities for exploratory comparisons, structural explanation, and transfer to unpracticed problems. Evidence from rodent maze studies does not determine the effectiveness of any particular classroom intervention. Educational conclusions require randomized or otherwise credible studies in human learning contexts that measure both understanding and delayed transfer Jensen (2006) Behrens et al. (2018).
Navigation research also informs the design of assistive interfaces and adaptive technologies. Tools that present relationships among locations, consequences, and alternative routes may help users understand a changing environment, while excessive reliance on fixed turn-by-turn instructions could leave less opportunity for practicing global relations. This is a testable hypothesis, not an established universal rule; user needs, accessibility requirements, and cognitive load may make a fixed instruction sequence the optimal solution in some cases. Effective design research should measure task completion, error recovery, independence under unexpected changes, and longer-term transfer instead of inferring learning from immediate interface satisfaction Epstein et al. (2017) Whittington et al. (2022).
For artificial agents, training without an immediate reward can be evaluated by the ability to reuse transition knowledge when goals change. The relevant benchmark includes unseen start points, altered graph edges, novel combinations of latent states, and restricted opportunity for postchange trial-and-error learning. A system's architecture can be inspected directly, but its task performance still requires carefully chosen comparison algorithms and evaluation distributions. Relating these results to biological cognitive maps is scientifically meaningful when the same measurable problem is defined and when implementation claims remain anchored to the evidence from each type of system Sun et al. (2025) Stachenfeld et al. (2017) Behrens et al. (2018).
Unresolved Problems and a Testable Research Agenda
The deepest unresolved question concerns what a cognitive map is at the level of a measurable mechanism. A geometric coordinate code, a sparse transition graph, an abstract task-state model, and a successor representation can all support flexible behavior while making different predictions under environmental change. Research should use matched conditions that manipulate geometry without connectivity, connectivity without perceptual features, reward without transition changes, and transition changes without reward shifts. These experiments would estimate which kinds of structure are acquired under which exposure histories and whether different forms of transfer recruit shared or complementary systems Stachenfeld et al. (2017) Momennejad et al. (2017) Whittington et al. (2022).
A second problem is the causal relation between representation and choice. Longitudinal data reveal how codes emerge during learning, while selective interventions can test whether particular codes are necessary for subsequent transfer. An especially informative design would track acquisition during unincentivized exploration, intervene only during a surprise goal probe, and evaluate both neural representations and the animal's first unpracticed decision. The intervention must preserve basic sensory processing and movement. Convergence across direct recordings, decoding, model comparison, and causal perturbation would provide a more stringent account than any single technique Sun et al. (2025) Morris et al. (1982) Courellis et al. (2024).
A third problem is how map-like structure is organized across physical, temporal, relational, and conceptual dimensions. Human single-neuron studies of event graphs and inference suggest that hippocampal systems can encode latent variables beyond spatial coordinates. It remains important to establish which representations generalize across domains, how they interact with orbitofrontal task-state codes, and how biological variability limits uniform computational descriptions. Preregistered, open and well-controlled multisite studies with shared transfer tests could distinguish general principles from laboratory-specific representations and analytic artifacts Tacikowski et al. (2024) Courellis et al. (2024) Wang et al. (2020) Qiu et al. (2024).
Conclusion
Latent learning identifies a fundamental property of adaptive behavior: prior experience can change the structure of later action even when an earlier reward-centered measure reveals only part of what has been acquired. The classical reward-introduction studies by Blodgett and Tolman with Honzik established the experimental problem, while place-learning and shortcut experiments made its representational content testable. Tolman's cognitive-map concept directed attention to relations that extend beyond a practiced motor sequence. Historical criticism improved the interpretation by demanding careful treatment of reward, behavioral theory, and competing explanations instead of assuming that one experiment identified a unique mental mechanism Blodgett (1929) Tolman & Honzik (1930) Tolman (1948) Jensen (2006).
The neural and computational evidence now supports a richer scientific account. Place cells and entorhinal grid-like signals demonstrate structured neural responses in navigational tasks; lesions and inactivation constrain the roles of hippocampal and other systems; human imaging and recordings identify relational, temporal, and abstract task organizations. Predictive maps and latent-state models explain how stored relations can be used when goals, transitions, or task demands change. These findings converge on the importance of learned structure for flexible behavior while leaving the exact algorithm, neural implementation, and boundaries of generalization subject to ongoing testing O'Keefe & Dostrovsky (1971) Hafting et al. (2005) Stachenfeld et al. (2017) Sun et al. (2025) Tacikowski et al. (2024).
The principal synthesis is methodological and theoretical. A cognitive map becomes a scientific explanation when its proposed content and use generate predictions that can be distinguished from alternative learned policies and associative representations. Latent knowledge becomes demonstrable when earlier experience is shown to govern a novel action under controlled new conditions. The strongest future research will integrate precisely defined exposures, surprise transfer tests, longitudinal neural measures, causal interventions, and out-of-distribution model comparison. Across animals, humans, and artificial systems, the enduring question is how experience creates reusable knowledge of a world before the future need for that knowledge has been specified Behrens et al. (2018) Whittington et al. (2022) Courellis et al. (2024).
References
- Behrens, T. E. J., Muller, T. H., Whittington, J. C. R., Mark, S., Baram, A. B., Stachenfeld, K. L., & Kurth-Nelson, Z. (2018). What is a cognitive map? Organizing knowledge for flexible behavior. Neuron, 100(2), 490–509. https://doi.org/10.1016/j.neuron.2018.10.002
- Blodgett, H. C. (1929). The effect of the introduction of reward upon the maze performance of rats. University of California Publications in Psychology, 4(8), 113–134. https://openlibrary.org/works/OL38572948W/The_effect_of_the_introduction_of_reward_upon_the_maze_performance_of_rats
- Constantinescu, A. O., O'Reilly, J. X., & Behrens, T. E. J. (2016). Organizing conceptual knowledge in humans with a gridlike code. Science, 352(6292), 1464–1468. https://doi.org/10.1126/science.aaf0941
- Courellis, H. S., Minxha, J., Cardenas, A. R., Kimmel, D. L., Reed, C. M., Valiante, T. A., Salzman, C. D., Mamelak, A. N., Fusi, S., & Rutishauser, U. (2024). Abstract representations emerge in human hippocampal neurons during inference. Nature, 632, 841–849. https://doi.org/10.1038/s41586-024-07799-x
- Doeller, C. F., Barry, C., & Burgess, N. (2010). Evidence for grid cells in a human memory network. Nature, 463, 657–661. https://doi.org/10.1038/nature08704
- Epstein, R. A., Patai, E. Z., Julian, J. B., & Spiers, H. J. (2017). The cognitive map in humans: Spatial navigation and beyond. Nature Neuroscience, 20, 1504–1513. https://doi.org/10.1038/nn.4656
- Garvert, M. M., Dolan, R. J., & Behrens, T. E. J. (2017). A map of abstract relational knowledge in the human hippocampal–entorhinal cortex. eLife, 6, e17086. https://doi.org/10.7554/eLife.17086
- Hafting, T., Fyhn, M., Molden, S., Moser, M.-B., & Moser, E. I. (2005). Microstructure of a spatial map in the entorhinal cortex. Nature, 436, 801–806. https://doi.org/10.1038/nature03721
- Jensen, R. (2006). Behaviorism, latent learning, and cognitive maps: Needed revisions in introductory psychology textbooks. The Behavior Analyst, 29(2), 187–209. https://doi.org/10.1007/BF03392130
- Momennejad, I., Russek, E. M., Cheong, J. H., Botvinick, M. M., Daw, N. D., & Gershman, S. J. (2017). The successor representation in human reinforcement learning. Nature Human Behaviour, 1, 680–692. https://doi.org/10.1038/s41562-017-0180-8
- Morris, R. G. M., Garrud, P., Rawlins, J. N. P., & O'Keefe, J. (1982). Place navigation impaired in rats with hippocampal lesions. Nature, 297, 681–683. https://doi.org/10.1038/297681a0
- O'Keefe, J., & Dostrovsky, J. (1971). The hippocampus as a spatial map: Preliminary evidence from unit activity in the freely-moving rat. Brain Research, 34(1), 171–175. https://doi.org/10.1016/0006-8993(71)90358-1
- Packard, M. G., & McGaugh, J. L. (1996). Inactivation of hippocampus or caudate nucleus with lidocaine differentially affects expression of place and response learning. Neurobiology of Learning and Memory, 65(1), 65–72. https://doi.org/10.1006/nlme.1996.0007
- Qiu, Y., Li, H., Liao, J., Chen, K., Wu, X., Liu, B., & Huang, R. (2024). Forming cognitive maps for abstract spaces: The roles of the human hippocampus and orbitofrontal cortex. Communications Biology, 7, Article 517. https://doi.org/10.1038/s42003-024-06214-5
- Schapiro, A. C., Rogers, T. T., Cordova, N. I., Turk-Browne, N. B., & Botvinick, M. M. (2013). Neural representations of events arise from temporal community structure. Nature Neuroscience, 16, 486–492. https://doi.org/10.1038/nn.3331
- Schuck, N. W., Cai, M. B., Wilson, R. C., & Niv, Y. (2016). Human orbitofrontal cortex represents a cognitive map of state space. Neuron, 91(6), 1402–1412. https://doi.org/10.1016/j.neuron.2016.08.019
- Stachenfeld, K. L., Botvinick, M. M., & Gershman, S. J. (2017). The hippocampus as a predictive map. Nature Neuroscience, 20, 1643–1653. https://doi.org/10.1038/nn.4650
- Sun, W., Winnubst, J., Natrajan, M., Lai, C., Kajikawa, K., Bast, A., Michaelos, M., Gattoni, R., Stringer, C., Flickinger, D., Fitzgerald, J. E., & Spruston, N. (2025). Learning produces an orthogonalized state machine in the hippocampus. Nature, 640, 165–175. https://doi.org/10.1038/s41586-024-08548-w
- Tacikowski, P., Kalender, G., Ciliberti, D., & Fried, I. (2024). Human hippocampal and entorhinal neurons encode the temporal structure of experience. Nature, 635, 160–167. https://doi.org/10.1038/s41586-024-07973-1
- Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological Review, 55(4), 189–208. https://doi.org/10.1037/h0061626
- Tolman, E. C., & Honzik, C. H. (1930). Introduction and removal of reward and maze performance in rats. University of California Publications in Psychology, 4(17), 257–275. https://iucat.iu.edu/iub/6635413
- Tolman, E. C., Ritchie, B. F., & Kalish, D. (1946a). Studies in spatial learning: Orientation and the short-cut. Journal of Experimental Psychology, 36, 13–24. https://doi.org/10.1037/h0053944
- Tolman, E. C., Ritchie, B. F., & Kalish, D. (1946b). Studies in spatial learning: Place learning versus response learning. Journal of Experimental Psychology, 36, 221–229. https://doi.org/10.1037/h0060262
- Wang, F., Schoenbaum, G., & Kahnt, T. (2020). Interactions between human orbitofrontal cortex and hippocampus support model-based inference. PLOS Biology, 18(1), e3000578. https://doi.org/10.1371/journal.pbio.3000578
- Whittington, J. C. R., McCaffary, D., Bakermans, J. J. W., & Behrens, T. E. J. (2022). How to build a cognitive map. Nature Neuroscience, 25, 1257–1272. https://doi.org/10.1038/s41593-022-01153-y
- Wikenheiser, A. M., & Schoenbaum, G. (2016). Over the river, through the woods: Cognitive maps in the hippocampus and orbitofrontal cortex. Nature Reviews Neuroscience, 17, 513–523. https://doi.org/10.1038/nrn.2016.56
- Wilson, R. C., Takahashi, Y. K., Schoenbaum, G., & Niv, Y. (2014). Orbitofrontal cortex as a cognitive map of task space. Neuron, 81(2), 267–279. https://doi.org/10.1016/j.neuron.2013.11.005