Gaze-centered gating, reactivation, and reevaluation of economic value in orbitofrontal cortex

The study investigates how gaze dynamics interact with valuation and choice in economic decision-making. Prior work shows that people and animals sequentially fixate options and tend to look longer at ultimately chosen items, suggesting a link between attention and value computation. However, because many paradigms keep stimuli visible, encoding and evaluation are conflated, leaving unclear whether multiple offers are encoded in parallel or sequentially and how gaze influences these processes. The authors pose the hypothesis that economic choice relies on a default sequential evaluation mechanism, and that gaze—even directed to empty locations where offers were previously shown—facilitates memory retrieval and reevaluation of option values. To isolate evaluation from visual encoding, offers were presented sequentially and separated by blank-screen delays while gaze remained unconstrained, enabling tests of look-at-something vs. look-at-nothing effects on behavior and orbitofrontal cortex (OFC) activity.

Evidence from neuroeconomics indicates OFC and vmPFC neurons encode economic value and can be modulated by gaze and attention. Competing frameworks propose parallel value computation by partially segregated populations versus attention-dependent, alternating encoding. Some studies show OFC encodes values of fixated cues and that attention/gaze shifts modulate value signals; others report simultaneous encoding under central fixation. Behavioral literature shows sequential fixations guide choice and that gaze bias both reflects and influences preferences. The look-at-nothing phenomenon suggests gaze to previously relevant but currently empty locations aids memory retrieval. These mixed findings motivate testing whether OFC encodes values sequentially, depends on gaze, and whether looking at empty locations reinstates value signals for reevaluation.

Participants: Two adult male rhesus macaques performed a two-alternative risky choice task while freely moving their eyes. Data comprised 5,971 valid trials (2,463 subject 1; 3,328 subject 2) and recordings from 248 OFC neurons (areas 11 and 13; 163 subject 1; 85 subject 2). Eye position was sampled at 1 kHz. Task: Two vertical bar offers were shown sequentially on opposite screen sides for 400 ms each (offer 1/offer 2), separated by blank delays (600 ms; delay 1/delay 2). After delay 2, a central fixation was reacquired, then both offers reappeared to cue saccadic choice and a 200 ms choice-hold. Reward magnitude (m) was indicated by color (small/gray 125 µL; medium/blue 165 µL; large/green 240 µL). Probability (p) was indicated by bar height (red top showed 1−p). Offers included safe small options (p=1) and gambles (medium/large with p∼Uniform[0,1]). Expected value EV=mp; risk σ²=mp(1−p). Subjective values SV were session-wise estimated via logistic regression of choice including EV, m, and σ². Behavioral analyses: Trials were mirrored so offer 1 mapped to left, offer 2 to right. Gaze variables included fraction of time looking right f_R=t_R/(t_R+t_L) within epochs, and precision-restricted looking within offer locations. Logistic regressions predicted right choice using EV_L, EV_R, m_L, m_R, σ²_L, σ²_R, order of presentation, and f_R per epoch; additional models assessed choice as a function of (t_R−t_L)/(t_R+t_L) in each epoch. Neural recordings and analyses: Single units from OFC (areas 11/13) were recorded with 32-contact probes. Spike counts were computed in sliding 200 ms windows every 10 ms. Linear regressions tested encoding of EV_L or EV_R per neuron/bin (OLS; significance via F-test; chance via bin-wise trial-order permutations; cluster-based run-length analysis for temporal contiguity). Gaze conditioning split bins into LookL (average gaze<0) vs LookR (>0). Additional control analyses balanced trial counts across gaze conditions, excluded central gaze, and time-locked encoding to first midline-crossing gaze shifts during delays. Subjective value and choice encoding: A full model n=β0+β1 SV_L+β2 SV_R+β_ch·ch (ch=+1 right, −1 left) assessed SV and choice signals across epochs. To avoid circularity, trials were split into two disjoint subsets: one to fit behavioral SV weights and the other to regress neural activity, symmetrically swapped. Reevaluation tests via residuals/choice probability: During delay 2, residuals z=n−(β0+β1 SV_L+β2 SV_R) were computed. For each cell, preferred vs non-preferred choices were defined by the sign of tuning (β1 or β2) relative to chosen side. Choice Probability (CP) quantified how residual fluctuations predicted preferred vs non-preferred choice, computed separately for LookL and LookR trials; ROC-based analyses corroborated CP results.

Behavior:

• Monkeys chose the higher-EV offer most often (subject 1: 72.19%; subject 2: 75.72%); choice probability increased with EV_R−EV_L (p<0.001).
• Gaze tracked offers when visible and exhibited look-at-nothing during delays: during delay 2, animals split gaze between former offer locations.
• Choice was biased by gaze beyond EV, m, and σ². Logistic models showed that the fraction of time looking to a side within each epoch predicted choosing that side, including delays with no stimuli. In a model including EV, m, σ², order, and f_R per epoch, f_R during delay 2 had a significant positive effect on choosing right (p=0.029, FDR-corrected). Choice probability rose monotonically with (t_R−t_L)/(t_R+t_L) across epochs. Neural (OFC):
• Across all trials, encoding alternated with stimulus order: EV_L encoding rose after offer 1 and into delay 1, while EV_R encoding rose after offer 2 and into delay 2; pre-offer baseline showed no encoding.
• Gaze-gated encoding: During offer epochs, neurons encoded the value of the currently fixated (ipsilateral) offer. Critically, during blank delays, encoding persisted only when gaze remained on the side of the last viewed offer (maintenance), and was reduced when gaze shifted away.
• Reactivation: During delay 2, looking at the empty location of the first offer (LookL) reactivated EV_L encoding even if the second offer had been viewed in between; looking right maintained EV_R encoding. Reactivation persisted after controlling for potential confounds (trial balancing, excluding central gaze, varying window sizes, permutation tests) and occurred specifically after gaze shifts from right to left during delay 2.
• Overlapping populations: Weights encoding an offer during its presentation correlated with weights during reactivation at its empty location in delay 2, indicating overlapping neural populations for encoding and reactivation. Populations encoding offer 1 and offer 2 during their respective presentations also showed correlated weights, consistent with asynchronous, sequential encoding.
• Choice-related activity and reevaluation: A significant fraction of cells encoded choice beyond SV in late epochs. During delay 2, residual fluctuations (after regressing SV_L and SV_R) predicted choice in a gaze-dependent manner. Choice Probability during LookL (reactivation of left) had median CP=0.56 for SV_L-tuned residuals (83% of cells CP>0.5; p≈1.4×10⁻⁴¹) and median CP=0.44 for SV_R-tuned residuals (85% CP<0.5; p≈1.1×10⁻⁴⁸). These effects indicate that gaze-triggered reactivation supports reevaluation that biases upcoming choice.

The findings directly address whether value encoding in OFC is parallel or sequential and how gaze modulates these computations. OFC primarily encodes the value of the currently inspected option and, when gaze returns to an empty location where an option was shown, reinstates that option’s value representation. This reactivation is not merely a passive memory trace of the most recent stimulus; it can revive the earlier offer’s value after an intervening distractor and, crucially, its moment-to-moment fluctuations predict choice even after controlling for subjective value. Thus, look-at-nothing behavior serves as a gaze-centered memory indexing mechanism that enables reevaluation of previously viewed offers. These results support a sequential default mode of economic choice in which options are evaluated one at a time, with gaze acting as a control signal to gate maintenance, reactivation, and comparison processes. The data reconcile prior discrepancies by highlighting task demands and gaze state: under free viewing with complex, feature-integration offers, OFC coding is strongly gaze-dependent and sequential, whereas central fixation or overtrained stimuli may permit more simultaneous encoding. The work suggests coordination with other regions (e.g., ACC, hippocampus) that could maintain broader offer maps or support memory-based deliberation, while OFC focuses on the currently sampled or reinstated option.

This work demonstrates that gaze both biases choice and gates OFC value signals: fixation on an offer elicits value encoding; looking away weakens it; and looking back to the empty location reinstates value coding of that offer. Reactivation is functionally meaningful, correlating with upcoming choice, and involves overlapping neurons that encode during stimulus viewing and during reactivation. Behaviorally, look-at-nothing time predicts choice beyond EV, magnitude, and risk. Collectively, the results reveal a sequential, gaze-centered mechanism for evaluation and reevaluation during economic choice. Future research should causally manipulate gaze and OFC activity to test directionality, probe interactions with ACC and hippocampus in maintaining multi-option maps, examine whether similar mechanisms operate for abstract/non-spatial options, and assess generalization across tasks, species, and overtraining regimes.

Causality remains unproven: the study is correlational and cannot definitively establish that gaze-triggered reactivation causes reevaluation or choices. A motor preparation component cannot be fully ruled out. The sample comprises two male macaques and recordings were limited to OFC areas 11/13, which may limit generalizability across sexes, species, and brain regions. The paradigm relies on spatially presented visual offers; it is unclear how findings translate to non-spatial or purely abstract choices. Gaze analyses, while extensive, required binning and mirroring conventions and sometimes unbalanced trial counts, though multiple controls mitigate these concerns.