Hippocampal-entorhinal cognitive maps and cortical motor system represent action plans and their outcomes

The study investigates how the brain represents and compares multiple action-outcome associations to enable efficient goal-directed action selection. Traditional models posit modular controllers and look-up tables linking motor commands to outcomes, which may be inefficient given the vast repertoire of actions, context-dependence, and temporal variability. The authors propose that arbitrary action-outcome associations are organized as an abstract cognitive map in the hippocampal-entorhinal system, enabling relational representation across actions and outcomes. Two core questions guide the work: (1) whether signatures of cognitive mapping (e.g., grid-like hexadirectional modulation and distance sensitivity) arise when representing discrete, arbitrary action-outcome associations; and (2) whether hippocampal-entorhinal maps interact with cortical motor regions (e.g., SMA) that encode action plans during evaluation and selection. Combining immersive VR training to learn an action-outcome space with fMRI tasks probing comparisons along that space, the study aims to reveal the representational scaffolding that supports selection among alternative actions based on desired outcomes.

Prior research established spatial cognitive maps in hippocampal-entorhinal circuits (place cells, grid cells) and extended these map-like codes to non-spatial domains, including conceptual, semantic, social, sensory, and value spaces. Grid-like (six-fold) signals have been observed in human entorhinal cortex during abstract navigation and conceptual inference, while hippocampal activity often scales with distances within learned relational structures. Motor control theories (e.g., MOSAIC, multiple forward/inverse models) specify controllers for actions but provide limited accounts of how relations among action-outcome modules are organized neurally. Studies implicate SMA and premotor areas in action planning, imagery, sequence chunking, and linking actions to arbitrary consequences, and suggest interactions between hippocampal and cortical motor systems during learning and memory. The authors build on this literature to test whether action-outcome knowledge can be encoded as a cognitive map in HPC-EC and how it interfaces with motor system representations.

Participants: 52 recruited; 46 included in analyses (22 female; age 19–35, mean 26.6). Right-handed, normal or corrected vision. Ethics approved at Leipzig University (112/21-ek). Three-day protocol: two training days (immersive VR and computer tasks) and one fMRI day with two scanning sessions.

VR training: Participants manipulated two virtual joysticks to perform five discrete actions each (left, backward, up, right, forward). Each joystick action mapped (nonlinearly, randomized per participant) to one outcome dimension: Joystick 1 to probability of catching a ball; Joystick 2 to probability of the ball remaining visible across its trajectory. Outcome probabilities were 0.1, 0.3, 0.5, 0.7, 0.9 and were not explicitly disclosed. Six specific action combinations yielded uniquely colored balls (blue, green, purple, turquoise, red, yellow) serving as landmark outcomes. The conjunction defined a 5×5 two-dimensional abstract action-outcome space, with each action combination located at a coordinate. Guided Exploration exposed all combinations and color landmarks; a Goal-directed Action Task required selecting actions to achieve instructed outcomes relative to colored-ball references.

Computer tasks post-VR: Two Rating Tasks collected pairwise similarity judgments for (1) action combinations and (2) colored balls, relative to both outcome dimensions. Two Comparison Tasks presented pairs sequentially and queried outcome changes (increase/decrease/same; or truth statements integrating both dimensions). Behavioral sampling ensured coverage across quadrants and central space; pairs counterbalanced.

fMRI tasks: In session 1, Comparison Task 1 (action combinations), the space was sampled along 12 directions (approximate 30° increments: 0°, 28.35°, 61.65°, 90°, 118.35°, 151.65°, 180°, 208.35°, 241.65°, 270°, 298.35°, 331.65) from varying start positions, with controlled distances per direction (mean ≈2.8–3 units) and balanced trial structure (144 trials across 6 runs). In session 2, Comparison Task 2 (colored balls) presented all pairs (150 trials across 5 runs). Reflection periods encouraged mental comparison prior to responses.

MRI acquisition and preprocessing: 3T Siemens SkyraFit, multiband GE-EPI (TR 1.5 s, TE 22 ms, 2.5 mm isotropic, 63 slices), fieldmap correction (topup), slice-time correction, BBR co-registration to T1, ICA-AROMA denoising, smoothing (RSA: 2.5 mm FWHM; other analyses: 7.5 mm FWHM), nuisance regressors (motion, CompCor, globals, derivatives), high-pass filtering (0.01 Hz). Anatomical preprocessing via fMRIPrep; normalization to MNI spaces.

ROIs: Entorhinal cortex (EC) and hippocampus (HPC) from Julich atlas (75% prob threshold; EC intersected with average brain masks). Premotor cortex (PMC) and supplementary motor area (SMA) from Harvard-Oxford (PMC 75%, SMA 50%). Lingual gyrus (LG; 50%) as control. Whole-brain searchlight and FDR corrections applied for confirmatory analyses.

Analyses:

• RSA grid-like test (Comparison Task 1): GLM modeled reflection-phase onsets for 12 directions; EC pattern similarity across directions computed via Mahalanobis distance to form neural RDM; Spearman correlation with model RDM expecting higher similarity for directions aligned modulo 60° (six-fold). Control periodicities (2–5 fold) and start/end-position control RDMs tested; cross-validated RSA via Crossnobis; EC laterality assessed; control ROIs tested; searchlight performed.
• Distance-based BOLD adaptation (Comparison Task 1): Parametric GLM regressor aligned to second stimulus modeling Euclidean distance between action combinations; ROI tests in HPC and EC; dimension-specific controls; subsampling to balance short/long trials; whole-brain confirmation.
• Distance-based adaptation for landmark outcomes (Comparison Task 2): Parametric regressor aligned to second colored ball modeling Euclidean distance between landmark positions; 12 extra regressors captured first/second ball colors to control for stimulus-specific activity; ROI and whole-brain tests; laterality assessed.
• Action similarity-based adaptation (Comparison Task 1): Trials categorized by shared actions between pairs (none, one shared, reversed both); parametric regressor (0,1,2) aligned to second stimulus; two-sided tests in SMA and PMC; controls in HPC, EC, LG; GLM including both distance and shared-action regressors verified independence; subsampling to equalize trial counts.
• gPPI connectivity (Comparison Task 1): Seed ROIs in left HPC and SMA; psychological conditions defined by long vs short distances; model included seed time series, condition regressors, and interaction (PPI) terms; extracted parameter estimates for connectivity between HPC and SMA; tested group contrast long vs short; control LG connectivity; whole-brain maps (uncorrected) for visualization.

Statistics: Non-parametric permutation-based t-tests with 10,000 sign-flips; one- or two-sided per a priori hypotheses; Bonferroni corrections across masks/hemispheres; whole-brain FDR voxel-level p<0.01.

Behavioral:

• Participants learned action-outcome relations and used them for selection; high accuracy in scanner Comparison Tasks (Task 1: M=93.7%, SD=7.99%; Task 2: M=92.9%, SD=8.80%).
• Rating Tasks: Subjective similarity correlated with Euclidean distances in the action-outcome space; Action combinations (Task 1): M Spearman’s rho=0.451 (SD=0.288), t(45)=10.63, p<0.001; increased on Day 2 to M=0.604 (SD=0.282), t(45)=14.52, p<0.001; across-day effect t(45)=-4.28, p<0.001. Colored balls (Task 2): M rho=0.389 (SD=0.270), t(45)=9.79, p<0.001; Day 2 M=0.571 (SD=0.261), t(45)=14.84, p<0.001; across-day effect t(45)=-4.82, p<0.001. MDS+Procrustes reconstructions fit original positions better than chance (Action combinations Day1 data distance M=0.368 vs critical 0.526; Day2 data M=0.257 vs critical 0.528. Colored balls Day1 data M=0.371 vs critical 0.382; Day2 data M=0.247 vs critical 0.372).

Neuroimaging:

• EC grid-like hexadirectional modulation (Comparison Task 1): Bilateral EC showed significant six-fold periodicity of pattern similarity, t(45)=2.32, p=0.011, Cohen’s d=0.34; left EC significant t(45)=2.54, p=0.011; right EC not significant. Control periodicities (2–5 fold) non-significant (all p>0.563). Control models for start/end positions non-significant. Searchlight confirmed EC effect and revealed medial prefrontal cortex cluster.
• Hippocampal distance coding (Comparison Task 1): HPC showed BOLD adaptation scaling with shorter distances between action combinations, t(45)=2.52, p=0.015, d=0.37; more apparent in left HPC t(45)=2.73, p=0.009; whole-brain FDR map peak MNI 30, −23, −20, t=4.87. EC trend did not reach significance, t(45)=1.89, p=0.055.
• Landmark distance coding (Comparison Task 2): HPC adaptation for colored ball pairs closer in space, t(45)=2.76, p=0.005, d=0.40; EC also significant, t(45)=2.57, p=0.013, d=0.38. Whole-brain HPC confirmation peak MNI 27, −8, −18, t=5.155.
• SMA action representation (Comparison Task 1): Trials with shared actions elicited increased SMA activity (repetition enhancement), t(45)=−3.07, p=0.005, d=−0.45; PMC non-significant t(45)=−1.78, p=0.078. Whole-brain FDR confirmation peak MNI −6, 4, 67, t=−4.673. Effect persisted when controlling for 2D distance; control ROIs (HPC, EC, LG) showed no motor repetition effects.
• HPC–SMA connectivity (gPPI, Comparison Task 1): Functional connectivity increased for long vs short distances, averaged across seeds, t(45)=2.42, p=0.021, d=0.35; replicated per seed; control LG showed no task-related connectivity changes. Whole-brain maps indicated mutual HPC–SMA coupling and involvement of posterior parietal cortex.

Overall, results indicate an abstract 2D cognitive map of action-outcome relations in HPC-EC (grid-like EC modulation; hippocampal distance sensitivity), parallel encoding of individual actions in SMA, and distance-modulated HPC–SMA functional coupling during evaluation of action plans.

The findings address the core hypotheses by demonstrating that arbitrary, discrete action-outcome associations are organized as a relational cognitive map within the hippocampal-entorhinal system. EC exhibited hexadirectional modulation characteristic of grid-like codes when participants mentally compared action plans, indicating a generalized structural representation of the abstract action-outcome space. Hippocampal activity scaled with distances between positions in this space, both for action combinations and for landmark outcomes (colored balls), showing modality-agnostic retrieval of the relational map.

In contrast, SMA encoded individual action plans and showed repetition enhancement when pairs shared actions, consistent with representations of motor primitives and action sequences rather than relational spaces. The observed increase in HPC–SMA connectivity for trials spanning longer distances suggests a complementary interaction: SMA may forward-simulate specific actions while HPC-EC situates these action-outcome links within an abstract map to support comparison and selection. Additional clusters (mPFC, IOFC) imply broader network participation in constructing and leveraging cognitive maps for behavior.

These results challenge classical separations of declarative vs procedural memory by showing that flexible action selection relies on explicit relational knowledge in HPC-EC interacting with motor planning systems. Map-like dimensional representations can enable efficient, potentially parallel evaluation of multiple action alternatives across outcome dimensions, providing a neurally plausible mechanism for goal-directed choice beyond the spatial domain.

This work demonstrates that the hippocampal-entorhinal system constructs and uses a cognitive map of arbitrary action-outcome associations, evidenced by EC grid-like hexadirectional modulation and hippocampal distance-sensitive adaptation during comparisons. SMA represents individual actions and exhibits repetition enhancement for overlapping action plans, and HPC–SMA connectivity scales with the relational distance between action options, indicating coordinated evaluation across memory and motor systems. Behaviorally, participants formed a 2D relational structure enabling accurate comparisons and efficient action selection.

Future directions include clarifying the directionality of information flow between HPC-EC and SMA using high-temporal-resolution methods (e.g., intracranial recordings), examining how probabilities are represented (e.g., log-odds transforms), probing the roles of mPFC/OFC in map construction and inference, and extending to naturalistic settings with richer outcome dimensions. Investigations into how map-like structures evolve with experience, value, and context, and how they integrate with motor sequence control, could further illuminate mechanisms of goal-directed behavior.

• Directionality of HPC–SMA interactions cannot be inferred from gPPI; causal flow remains unknown.
• No a priori power analysis; sample size based on related domains; generalizability should be interpreted cautiously.
• SMA representations could not be fully disentangled from outcome features due to tight action–outcome correlations; action vs outcome coding in SMA may overlap.
• Design constraints led to slight imbalances in distance sampling for cardinal directions; subsampling controls mitigate but do not eliminate potential biases.
• Gender effects were not analyzed; right-hand-only control used to standardize actions.
• Outcome probabilities were not explicitly revealed; learning relied on experiential inference, which may vary across individuals.
• Whole-brain connectivity maps reported uncorrected for multiple comparisons for visualization; ROI-based statistics used for inference.
• Data privacy restrictions limit access to raw data; preprocessed/processed data available upon request/OSF.