Neural subspace reorganization reflects value-based decision-making Authors: Huidi Li, Nikolaos Chrysanthidis, Scott L Brincat, Jonas Rose, Earl K Miller Published: 2026 ABSTRACT SummaryValue-based decision-making requires maintaining and comparing multiple options to guide behavior. While the neural circuits supporting these decisions are increasingly understood, how decisions reshape option representations remains unclear. We recorded from lateral prefrontal cortex (LPFC) while non-human primates held two sequentially presented option-value pairs in working memory and selected the higher value option. We found that, once a decision could be made, representations of the chosen and unchosen options rotated into orthogonal neural subspaces, while the chosen representation expanded. Before the decision, representations were organized by presentation order. After the decision, the chosen option, regardless of presentation order, was transformed into a common subspace aligned with the prescribed action. These findings suggest that value-based decisions reorganize LPFC population activity to selectively represent the chosen option, enabling efficient downstream readout for action selection.Keywords: value-based decision making, working memory, neural population analysis, prefrontal cortex, subspace coding Introduction To make an economic choice, we must compare the values of goods available and choose the goods with greater value. The process of making this choice can be broken down into multiple steps including representing available options, assigning values to them, comparing those values, and selecting the option with greater value.1,2,3,4,5 Previous research has studied how values are encoded and compared to make a choice.6,7,8,9 But it remains less well understood how multiple options are simultaneously represented in working memory and how these representations are transformed and reorganized by value-based choices. Lateral prefrontal cortex (LPFC) is a likely candidate site for linking options, values, and actions, and for selecting actions based on their associated values.10,11,12,13 Multiple potential options often need to be maintained in working memory during value assignment. The LPFC is also implicated in encoding both choice values and action plans. It receives direct input from the orbitofrontal cortex,14,15 which is involved in processing reward value and other economic factors.1,2,16 It also has dense projections to motor systems.17,18,19 Neural activity in the LPFC has been shown to reflect the transformation of value into action.11 Many recent studies have shown that population spiking activity in LPFC,20,21,22,23,24,25,26,27,28,29 as well as other frontal areas,30,31,32,33 exhibits structured correlations across neurons. As a result, neural activity tends to be restricted to low-dimensional “subspaces” within the high-dimensional space of all possible population activity patterns.34,35,36 Structuring activity within such subspaces is thought to flexibly organize information processing20,21,22,23,24,25,31,37,38,39,40 and communication.41,42 Motivated by these observations, we recorded from LPFC while non-human primates (NHPs) performed a value-based selection task. To segregate option representations before and after decisions, we used a novel task paradigm that temporally dissociated presentation of options from their value assignment. We examined the representational geometry of the options before and after their relative values were assigned and a decision could be made. We found that, before the decision, the two potential options were held in near-orthogonal subspaces. Once the decision was made, they were transformed to different subspaces that were now orthogonalized by whether their associated actions were chosen vs. unchosen, and the representation of the chosen action was expanded. Chosen actions were encoded within the same subspace, regardless of the initial presentation order of their associated option, facilitating their readout to generate behavior. These results show that organized changes in representational geometry of options support value-based decision-making. Results Behavior and analysis In each trial (Figure 1A), NHPs were presented with a series of two spatial targets (T1 and T2). Each was followed by a reward cue (R1 and R2, respectively) whose color signaled the reward value available if the preceding target was chosen. After the two targets and their values were presented, there was a “post-decision delay”. The NHPs had all the information needed to decide which target to choose. But they had to wait for a “go” cue (the reappearance of the two targets) before choosing one by saccading to it. Both NHPs most often chose the higher-value target over the lower value target (97.77% for NHP T; 78.56% for NHP I). Both NHPs reliably chose the higher value target across all pairs of first and second reward values (Figure 1B; all p < 0.001, one-sample t test). NHP T’s choices were more binary, consistent with their overall higher performance (Figure 1C, top), whereas NHP I’s choices showed a more graded dependence on the relative reward difference (Figure 1C, bottom). Nonetheless, all major effects were robust across both NHPs. Our analyses focus on trials in which NHPs chose the higher value target. Experimental paradigm and behavioral results (A) Time course of a task trial, proceeding from left to right. Two spatial targets (T1 and T2) were shown, paired with abstract cues (R1 and R2) signaling the reward value available if the preceding target was chosen. The NHPs’ task was to choose the target associated with the higher reward. (B) Probability of choosing the high-reward target for NHP T (top) and NHP I (bottom) across all combinations of first and second reward values. Values along the diagonal—where the same reward was offered for both targets—were not tested. (C) Probability (mean ± SD across sessions) of choosing the first presented target at different relative reward levels (first-target reward–second-target reward). Both NHPs reliably chose the higher-value target across all pairs of first and second reward values. Not all trials required an explicit comparison between the reward value of the targets. If the reward value of the first target (R1) was the highest possible value, the NHPs could immediately decide to choose it. Alternatively, if R1 was the lowest possible value, the NHPs could proactively decide to choose the second target when it appeared. We instead focused neurophysiological analysis on the remainder of trials. On those, the NHPs had to wait until the second reward value cue (R2) before they could decide which target had the higher value. Thus, we focused on the time periods just before and after R2. After the second target (T2) but before its value was known (R2), NHPs needed to hold both targets in memory (pre-decision delay). Then after R2, they could compare the two targets’ reward values and make a decision (post-decision delay). Our data analysis was based on the spiking activity of neurons in LPFC for two NHPs (NHP T: 316 neurons; NHP I: 422 neurons). LPFC encoded both reward and target location information We examined how reward and target information evolved during the decision process. We trained a linear discriminant analysis (LDA) classifier to decode each reward cue’s value and each target’s location from spike rates across the entire neural population. We found an increase of reward information (decoding accuracy) after each reward cue was presented and the reward information was maintained until NHPs made responses (Figure 2A). There was also information about other value-related variables, including chosen value and relative value (Figure S1). Next, we compared location information for chosen vs. unchosen targets separately for trials when the first vs. second target was chosen. Before the decision, both target locations could be decoded with similar accuracy. But after the second reward cue (R2), when the decision could be made, decoder accuracy for the location of the chosen target increased (Figures 2B and 2C). Information reflecting the first target location increased when it was chosen (Figure 2B), and there was a corresponding increase for the second target when it was chosen (Figure 2C). This resulted in an interaction between target order and choice (p < 0.001, bootstrap test on difference-of-differences; see STAR Methods). We also observed a small subject-specific modulation of this effect that paralleled differences in their behavior (Figure S2). NHP I was slightly biased toward choosing the first target and decoding accuracy difference between chosen and unchosen targets was also slightly greater when the first target was chosen, while choice behavior and neural information were both slightly biased toward the second target for NHP T. LPFC encoded information about both reward cue value and target location (A) Decoding accuracy of first (R1; cyan) and second (R2; purple) reward cue value (mean ± SD of pseudopopulation bootstraps). (B and C) Decoding accuracy for the first (T1; green) and second (T2; orange) target location in trials where the first target was chosen (R1>R2; B) and where the second target was chosen (R1R2 and R1R2 and p = 0.398 in R1R2 and p = 1 in R1R2) vs. trials when the second target was chosen (R1R2] vs. choose T2 [R1