Hippocampal spatio-predictive cognitive maps adaptively guide reward generalization

Humans must generalize from past experience to novel situations. Cognitive maps in the hippocampal formation have been proposed as a substrate for such generalization. Stimuli can be embedded simultaneously in multiple relational structures, such as policy-independent spatial (Euclidean) relationships and policy-dependent predictive (transition) relationships derived from experience. The core question is how the brain selects and adapts the most beneficial relational map to guide reward-based decisions and inference when multiple maps exist. The orbitofrontal cortex (OFC) is implicated in representing task states and may arbitrate between maps, but its interaction with hippocampal map representations during adaptive generalization is not well understood. This study tests whether spatial and predictive maps are represented and used for value generalization, and how their use is adaptively tuned by outcome evidence.

Prior work shows hippocampal cognitive maps encode spatial distances and support flexible behavior beyond stimulus-response learning. Recent studies extend this to nonspatial domains, including conceptual, temporal, and associative structures, often formed incidentally and automatically. Predictive or successor representations can capture policy-dependent transition relations, while spatial maps capture Euclidean distances. OFC is known to represent latent task states and may guide selection of relevant representations. However, how OFC signals relate to updates in hippocampal maps for guiding generalization had been unclear. This work builds on literature on hippocampal spatial and nonspatial mapping, statistical learning of temporal structure, and OFC state representations to test parallel spatial and predictive mapping and their adaptive selection.

Participants: 48 healthy adults (mean age 26.8 ± 3.8 years, 20–34 years; 27 male) completed a 3-day study (approved by the University of Leipzig ethics committee). Day 1: participants freely explored a circular virtual arena (radius 15 m) containing 12 monster stimuli at fixed but individually randomized locations, with landmarks outside the walls. Stimuli appeared when approached (<3 m). After exploration, an object location memory task required navigating to cued stimulus locations; feedback reflected distance error. Blocks alternated until average replacement error <3 virtual meters (10% of diameter), minimum five and maximum ten blocks. Day 2: picture-viewing fMRI session (3 blocks; each monster shown on red/blue backgrounds in random order, occasional two-alternative distance judgments, no feedback). Object location memory without feedback before/after scanning. Day 3: in-scanner choice task (100 trials) with two contexts (signaled by background color) mapping spatial locations to reward magnitudes; contexts alternated in 10-trial blocks. Two inference stimuli per context were never presented during choice. After choice, picture-viewing fMRI (distance or value similarity judgments with incentives), then object location memory without feedback, and post-tests (value ratings by context, liking, similarity arena, spatial arena placement). Modeling: Gaussian Process (GP) regression modeled value generalization using kernels capturing (1) spatial similarity (RBF over Euclidean distances), (2) predictive similarity (diffusion kernel from participant-specific transition probabilities estimated via successor representation learned from day-1 visitation histories), and (3) compositional spatio-predictive kernel (additive combination). A mean-tracker (no generalization) served as baseline. Mixed-effects logistic regression related model-predicted value differences to choices; model frequencies computed via leave-one-trial-out cross-validated log-likelihood. Participant-specific spatial and predictive effects and dynamic trial-by-trial weights were estimated; logistic functions were fitted to spatial weight trajectories to obtain slopes. fMRI: Event-related GLMs in SPM12. Picture-viewing GLM included separate stimulus onsets with parametric modulators for spatial and predictive distances to the preceding stimulus (z-scored, non-orthogonalized) to measure cross-stimulus adaptation/enhancement scaling with relational distance. Choice-task GLMs included: (i) value modulators, (ii) spatial weight update (trial-to-trial change), (iii) reward prediction error (RPE) based on compositional map, and (iv) relative map accuracy (difference in unsigned prediction errors: spatial vs predictive). Small-volume corrections used anatomically defined masks for hippocampal formation, OFC, and caudate. ROI-based correlations linked neural map strength to behavioral weights and inference performance. Mediation analyses tested: (1) hippocampal spatial map → spatial weight → inference error, and (2) OFC relative map accuracy → hippocampal spatial weight update → change in hippocampal spatial map.

• Behavior and inference:
  • Participants learned spatial layout (replacement error criterion achieved; stable across days 2–3). Choices depended on value differences (context 1: t(47)=10.0, P<0.001; context 2: t(47)=12.1, P<0.001).
  • Inference stimuli (never chosen) were rated correctly (RM-ANOVA high vs low value: F(1,46)=21.4, P<0.001). Map reproduction error correlated with inference rating error (r=0.37, P=0.01).
• Computational modeling:
  • Spatio-predictive GP (additive spatial + predictive similarities) best explained choices (model frequency=0.681, s.d.=0.065, P_exc>0.999). Spatial: 0.23; Predictive: 0.08; Mean-tracker: 0.005.
  • Model reproduced inference effects (F(1,47)=2602.3, P<0.0001) and prediction error correlated with participants’ inference error (r=0.85, P<0.001).
  • End-of-study inference ratings also best fit by spatio-predictive model (exceedance probability ~0.988). For experienced stimuli, mean-tracker fit best, indicating episodic recall where possible.
  • Spatial and predictive effects on choice were both >0 and negatively correlated (r=-0.45, P=0.001). Greater relative spatial weight predicted better inference (r=-0.43, P=0.003).
• Hippocampal maps:
  • Cross-stimulus enhancement scaling with spatial distance observed in right hippocampal formation after the choice task (session 3): peak t(47)=3.86, P_SVC=0.045, [24, -28, -16]; left hippocampal trend t=3.63, P=0.08.
  • ROI parameter estimates correlated with behavioral weights: spatial (r=0.37, P=0.01) and predictive (r=0.40, P=0.005). Spatial map strength correlated with lower inference error (r=-0.44, P=0.002), predictive map did not (r=0.06, P=0.68).
  • Second-level covariates: stronger behavioral spatial effects associated with greater hippocampal spatial map (peak t(47)=4.45, P_SVC=0.009, [22, -28, -18]); stronger predictive effects associated with hippocampal clusters (peaks t=4.19–3.91, P_SVC=0.02–0.04).
  • Inference error covariate: smaller errors linked to stronger hippocampal spatial map (peaks t=5.08 and 4.95, P_SVC=0.002) and weaker predictive map (peak t=4.53, P_SVC=0.007).
  • Mediation: hippocampal spatial map → spatial weights → inference error (a=0.3±0.1, P=0.01; b=-3.4±0.9, P=0.004; c′=-1.1±0.4, P=0.02; indirect ab=-0.9±0.4, P≈0.0003–0.0004).
• Adaptive selection and updating:
  • Over trials, spatial weights increased and predictive decreased; logistic slopes of spatial weight trajectories predicted better choice and inference performance (inference: r=-0.44, P=0.002).
  • Change in hippocampal spatial map from day 2→3 correlated with slope (r=0.38, P=0.008); predictive map decreased across participants (t(47)=-2.1, P=0.04). Changes in spatial vs predictive map were negatively correlated (r=-0.62, P<0.001).
  • Hippocampal spatial weight update signal at feedback (left hippocampus: t(47)=4.14, P_SVC=0.02, [-18, -32, -18]). Individuals with stronger hippocampal weight update showed larger hippocampal spatial map increase (peak t(47)=4.21, P_SVC=0.018, [18, -14, -25]).
• OFC and RPE signals:
  • RPE tracking covaried with hippocampal weight update in OFC (t(47)=4.75, P_SVC=0.02, [-4, 32, -20]), striatum (right t=3.57, P_SVC=0.029; left t=3.63, P_SVC=0.051), and hippocampus (right t=5.06, P_SVC=0.002; left t=4.00, P_SVC=0.04).
  • Relative map accuracy (spatial vs predictive unsigned PE difference) tracked by medial OFC (P_FWE cluster=0.03, [14, 46, -13]) and scaled with hippocampal updating.
  • Mediation: OFC relative map accuracy → hippocampal spatial weight update → increase in hippocampal spatial map (a=0.7±0.3, P=0.02; b=14.1±4.6, P=0.001; indirect ab=10.2±5.7, P=0.01; direct c≈2.9±6.6, P=0.57).

The study shows that the hippocampus concurrently represents spatial and predictive relations among stimuli learned during navigation. The relative strength of each hippocampal map predicts how individuals generalize values during decision-making, linking representational geometry to behavior. Despite only spatial location determining rewards, participants initially relied more on predictive relations but adaptively shifted toward spatial generalization as accumulating evidence favored the spatial map. This shift was mirrored neurally by increased hippocampal spatial mapping and reduced predictive mapping. The OFC encoded signals indexing both reward prediction error and the relative accuracy of spatial versus predictive maps, and these OFC signals were associated with hippocampal weight updates and subsequent strengthening of the spatial map representation. Together, these results support a mechanism whereby OFC evaluates which relational structure best explains outcomes and biases hippocampal representations toward the task-relevant map, enabling flexible inference and improved generalization. The findings integrate theories of hippocampal cognitive maps with OFC state representations and arbitration in adaptive decision-making.

This work demonstrates that humans use both spatial and predictive cognitive maps for value generalization, with hippocampal representations of these maps predicting and mediating behavior. As task demands reveal which map is more relevant, OFC encodes the relative accuracy of competing maps, drives hippocampal weight updates, and facilitates adaptive strengthening of the appropriate hippocampal map. Contributions include: (1) behavioral evidence for compositional generalization over spatial and predictive relations, (2) neural evidence for parallel hippocampal maps linked to generalization and inference, and (3) an OFC–hippocampus mechanism for adaptive selection and updating of cognitive maps. Future work could more fully dissociate whether two distinct maps or a single experience-distorted map underlies behavior, manipulate reward–space correlations to separate consolidation from reward-driven updating, and test causal roles of OFC–hippocampal interactions (e.g., via perturbation) in adaptive map selection.

• Correlation between reward distributions and spatial distances limits disentangling effects of reward feedback versus consolidation in driving representational changes.
• Modeling cannot definitively distinguish two distinct maps from a single experience-dependent combined map that distorts spatial information.
• Repetition enhancement vs suppression mechanisms in fMRI adaptation are not fully understood, complicating interpretation of directionality.
• Interindividual variability in exploration policies and map reliance introduces variability; although modeled, residual confounds may remain.
• Some effects (e.g., predictive-map voxels, left hippocampal spatial effects) were weaker or trend-level at strict corrections, and small-volume corrections focused on a priori ROIs.