The medial and lateral orbitofrontal cortex jointly represent the cognitive map of task space

Cognitive maps are internal representations of relationships among entities, enabling flexible behavior by integrating observable stimuli with relevant memories for predicting outcomes in new situations. The human orbitofrontal cortex (OFC) has been implicated in constructing cognitive maps that guide decision-making, integrating value and expected rewards, influencing sensory processing, and combining multisensory inputs with prior information to represent task space. The OFC is functionally heterogeneous, encompassing medial (mOFC) and lateral (lOFC) subregions that may specialize in different aspects of cognitive map representation. The cognitive map of task space comprises task states and their relational structure. Prior work suggests mOFC represents the current state when it includes unobservable components, while lOFC has been linked to simulating outcomes and may also create and modify maps by encoding relationships among entities and integrating current sensory inputs with prior information. The present study asks how mOFC and lOFC differentially contribute and interact to form a cognitive map of task space in humans. The authors hypothesize that mOFC represents individual state components, especially hidden information, defining the current location in task space, whereas lOFC encodes relational information and abstract rules (e.g., whether age/category switched) governing structure knowledge across states. They further propose that mOFC may provide component information via functional connectivity, enabling lOFC to integrate and construct relational representations.

The paper situates its work within research showing the OFC’s role in flexible behavior, value coding, and cognitive maps of task space across species. It reviews evidence that mOFC represents current, partially observable task states, and lOFC participates in both using and constructing cognitive maps, encoding relationships among entities and integrating prior with current information. Related findings highlight OFC’s direct projections to sensory cortices, influencing sensorimotor associations and multisensory integration, and prefrontal cortex (including dlPFC) involvement in abstract rule encoding and relational coding. The authors reference studies indicating that hippocampus and OFC can jointly represent task structure during memory-guided decisions and cognitive maps for abstract spaces, while also noting mixed findings on hippocampal engagement depending on task demands. Meta-analytic connectivity modeling suggests differential functional connectivity of mOFC versus lOFC with dlPFC, consistent with the proposed division of labor in state versus rule representation.

Participants: 34 right-handed healthy adults were recruited; 7 were excluded (3 quit, 2 poor performance >3 SD error rate, 2 excessive signal dropout in ventral PFC). Final N=27 (17F/10M), mean age 21.6 ± 3.2 years. IRB approval obtained; written informed consent provided. Stimuli: 42 images (FACES database and Haemmerer & Ebner), comprising young/old faces and houses. Each trial presented an overlapped face+house stimulus where each category could be young or old; stimuli could be age-congruent or incongruent between face and house. Task design: Multi-step sequential task requiring integration of current sensory inputs and memory from the previous trial. Initial run trial provided a cue indicating which category to judge (face or house). Participants continued judging the same category until the age of that category changed, at which point the judgment category switched. Task rules therefore depended on whether an age switch or a category switch occurred. Four binary components defined each task state: current age (observable), current category (hidden), previous age (hidden), previous category (hidden), yielding 2^4=16 states with equal transition probabilities (P=0.5). Trials were classified as Enter (category switch), Internal (no switch), and Exit (age switch within same category, signaling that the next trial will be a category switch). Behavioral training ensured proficiency on single-category age judgment and on the sequential switching rules before scanning. Task procedure: Responses via bimanual 2-button box; options randomized left/right. Each trial: fixation+options 0.70–4.70 s (avg 1.20 s), overlapped stimulus 0.55–8.30 s (avg 3.30 s), total avg trial 4.50 s. Response deadline 2.75 s; incorrect or missed trials repeated with feedback. If current trial age did not change, next trial age changed with 50% probability. Scanning protocol: 3T Siemens Prisma-fit, 64-channel coil. Task fMRI: 4 runs × 97 trials (388 trials total), GE-EPI SMS: TR=1500 ms, TE=31 ms, FOV=211×211 mm, matrix 88×88, voxel 2.4 mm isotropic, 60 slices, slice tilt 30° to improve OFC signal. Field-map with double-echo GRE: TR=620 ms, TE1/TE2=4.92/7.38 ms, voxel 2.4 mm^3, 60 slices. Structural T1 MPRAGE: TR=1800 ms, TE=2.07 ms, voxel 0.8 mm isotropic, matrix 320×320, 208 sagittal slices. Each participant also had resting-state fMRI. Preprocessing: fMRIPrep 20.2.7 (Nipype 1.6.1): motion correction (6 params), slice-timing correction, distortion correction using field-map, realignment, co-registration, tissue segmentation, normalization to MNI 2 mm space. ROI definitions: Bilateral mOFC, lOFC, hippocampus (HP), fusiform face area/parahippocampal place area (FFA/PPA) from Desikan–Killiany atlas with FreeSurfer; surface ROIs converted to volumetric and transformed to native functional space using ANTs. dlPFC (BAs 9/46) defined with WFU PickAtlas. Multivoxel pattern analysis (MVPA): Trial-wise beta maps from voxel-wise GLM (AFNI). Excluded first trial of each run, error trials, and first trial after errors. Events modeled as boxcars with duration equal to RT; 16 state regressors convolved with canonical HRF; 6 motion parameters per run. Averaged trial-wise beta maps per state within each run to obtain 16 beta maps/run; total 64 beta maps per participant. Within each ROI, beta maps Z-scored and smoothed (FWHM=4 mm). Classification with scikit-learn linear SVM (C=0.001), leave-one-run-out CV. Multiclass SVM (one-vs-rest) for 16-state classification; six binary SVMs for components: task-relevant (current age, previous age, current category, previous category) and task-irrelevant (age, category from two trials ago). Representational similarity analysis (RSA): For each ROI, computed Pearson correlations between state patterns across runs to avoid temporal autocorrelation; averaged across runs to form 16×16 correlation matrices. Hierarchical clustering (SciPy, agglomerative, average linkage, Euclidean distance) applied to correlation matrices. Four model RDMs constructed: Enter_non-Enter (category switch), Exit_non-Exit (age switch), Control1 (category switch only within same age), Control2 (age switch within same and different categories). Multiple regression RSA compared neural RDM (1−correlation) to model RDMs, using lower triangles as predictors with an intercept; Bonferroni-corrected one-sample t-tests against zero on beta coefficients. Whole-brain searchlight RSA (BrainIAK, 5 mm radius) with Z-scored beta maps and cross-run correlations; TFCE with 10,000 permutations and FWE correction. Connectivity-based searchlight: Seed sphere (8 mm radius) centered in right mOFC (MNI: 3, 44, −14). Computed whole-brain beta-series correlations per trial, averaged within runs for 16 states, then used as inputs to a LOO-CV linear SVM (C=0.001) searchlight (6 mm radius) to classify the hidden component "previous age"; significance assessed via one-sample t-test against 50% with FDR correction (p<0.001). Statistics and behavioral analyses: Linear mixed-effects models (lme4) tested effects of time course and trial type on error rate (avg 4.7%) and RT (avg 1137 ms). Error rates: Enter 7.1%, Internal 3.5%, Exit 3.7%; RTs: Enter 1202 ms, Internal 1074 ms, Exit 1143 ms. One-sample t-tests and permutation tests (10,000 iterations) assessed decoding above chance; bootstrapped tests (10,000 iterations) related state similarity to error rates.

• Both mOFC and lOFC multivoxel patterns decoded the 16 task states above chance (chance=6.25%; p<0.001 for each ROI; Table S1). • Component-level decoding: mOFC significantly encoded all hidden components (previous age, previous category, current category) above chance (50%) (p<0.05), but did not represent the observable current age. lOFC encoded current and previous category (p<0.05) but not previous age (p>0.05), and also did not represent current age. • Task-irrelevant components (age/category from two trials ago) were not decodable in mOFC or lOFC (p>0.05), indicating specificity to task-relevant hidden information. • Control ROIs: HP encoded current and previous category (p<0.05). dlPFC and FFA/PPA significantly decoded current age, current category, and previous category, and showed significant decoding of task-irrelevant components, unlike mOFC/lOFC (Supplementary Fig. S2). • Behavioral relevance: Greater similarity (higher cross-run correlation) between neural patterns for state pairs predicted lower error rates for mental category switches (r=−0.76, t26=−2.87, p=0.028, CI [−0.98, −0.17]) and mental age switches (r=−0.88, t26=−4.59, p=0.003, CI [−0.98, −0.53]) in mOFC. No such relationships in lOFC, HP, dlPFC, or FFA/PPA. • Hierarchical clustering: lOFC and dlPFC grouped states by Exit vs non-Exit (age-switch rule; silhouette 0.44 both), whereas mOFC and HP grouped states by Enter vs non-Enter (category-switch rule; silhouette 0.39 and 0.43, respectively), showing distinct rule-based organization. • ROI RSA: lOFC and dlPFC significantly represented both task rules (Enter_non-Enter and Exit_non-Exit) after Bonferroni correction (lOFC: t26=4.41 and 3.22; dlPFC: t26=6.14 and 3.41). Control2 (age switch across categories) also significant in lOFC (t26=3.44) and dlPFC (t26=3.26), suggesting participants recognized a mental age switch even when not required between Exit→Enter. mOFC represented category-switch rule (Enter_non-Enter: t26=4.50) but not age-switch rule (Exit_non-Exit: t26=1.19). • Whole-brain searchlight RSA confirmed lOFC and dlPFC significance after FWE correction; anterior insula also reached significance (Fig. 5C; Table S2). • Connectivity-based searchlight: Right mOFC exhibited state-related functional connectivity supporting classification of the hidden component "previous age" with left lOFC (MNI −38, 21, −11; t26=4.41, p<0.001), extending into left insula and inferior PFC; right putamen (31, −7, −9; t26=5.20, p<0.001), right dlPFC (21, 25, 49; t26=5.88, p<0.001), and left supramarginal gyrus (−65, −53, 33; t26=5.21, p<0.001) (Fig. 6; Table S3). This suggests lOFC/dlPFC and other regions may access hidden state information via mOFC connectivity to track age switches.

Findings support a dissociation and interaction between mOFC and lOFC in constructing cognitive maps of task space. The mOFC selectively represents the current location in task space by encoding hidden task-relevant components that cannot be directly observed, and the fidelity of this state representation predicts behavioral accuracy during mental switches. In contrast, lOFC, together with dlPFC and anterior insula, represents abstract relational rules governing transitions between states (age-switch and category-switch structure), consistent with broader prefrontal contributions to rule-guided behavior and relational coding. Despite lOFC not directly encoding all hidden components (notably previous age), connectivity analyses indicate that lOFC, dlPFC, and other nodes (putamen, supramarginal gyrus, insula) may obtain unobservable information through functional coupling with mOFC to support rule tracking. These results integrate with views of lOFC as a cognitive cartographer that organizes relational information and of mOFC as a hub for representing partially observable task states. The work extends frameworks of cognitive maps in non-spatial, abstract task spaces, clarifying complementary roles within OFC and their network-level interactions.

The study demonstrates that medial and lateral orbitofrontal cortex play complementary, dissociable roles in representing cognitive maps of task space. The mOFC encodes the hidden components defining current task states, enabling accurate tracking of partially observable positions. The lOFC, along with dlPFC and insula, encodes abstract rules that structure relationships across states (age/category switches), forming the relational architecture of the map. Functional connectivity indicates exchange of hidden state information from mOFC to lOFC and other areas to support rule tracking. These findings advance understanding of how the brain constructs and utilizes abstract cognitive maps in non-physical, task-relevant spaces to guide flexible decision-making. Future work should probe how these representations support model-based inference, generalization across tasks, and interactions with sensory cortices.

The study did not examine how OFC subregions collaborate to utilize cognitive maps for model-based inference and generalization across tasks, nor the dynamics of inter-regional information flow regulating sensory cortex integration. Methods such as informational connectivity and dynamic causal modeling could elucidate information transfer and local circuit contributions within OFC and allied regions. The hippocampus (HP) did not show representation of all hidden components or relational structure in this task, though prior work implicates HP in abstract and physical maps. Task demands here may rely more on immediate representations and working memory, engaging prefrontal or sensory-processing regions more than HP, or may not have sufficiently engaged HP under these parameters, warranting further investigation.