Distributed encoding of action-mediated outcome drives consistent population dynamics during goal-directed reaching
nvidia/Gemma-4-26B-A4B-NVFP4 · academic_accessible/eval 94%/5 min read/Jul 12, 2026
Generated by a local model (nvidia/Gemma-4-26B-A4B-NVFP4) from a scientific paper, claim-checked against the full text. Provenance is open by design.
When animals reach for food, their brains do more than just coordinate muscles. They continuously track whether a reward is actually available. This "expectation" signal is spread across many different brain regions. It is distinct from the signals that control the physical movement itself. Understanding how these internal goals interact with outward actions is a fundamental challenge. It sits at the intersection of motor control and cognitive decision-making.
The blind spot in movement research
For decades, researchers have studied the neural correlates of movement. They have focused heavily on the motor cortex. Here, neurons encode kinematics (the geometry and speed of motion). While movement-related activity is distributed across widespread areas, most studies treat "internal" signals as isolated events. These signals, like the expectation of a reward, are usually studied before a movement starts or after a reward is consumed.
This creates a significant analytical gap. It is difficult to dissect purely cognitive signals from the massive wave of neural activity occurring during a physical reach. It remains unclear how much of the brain's "global" activity during movement is dedicated to mechanics versus mental anticipation. Without a way to separate these, we risk misinterpreting motor commands as cognitive intentions.
Disentangling intent from motion
To resolve this, the authors employed a multi-pronged approach. They isolated "intent" from "mechanics." They recorded the spiking activity of 22,036 neurons across multiple brain regions in head-fixed mice using high-density Neuropixels probes .
Figure 2: Heterogeneous modulation of neuron activity across regions during reaching. (a) Peri-spike timing histogram (PSTH) shows trialaveraged and range-normalized firing rate of all neurons (rows) from all recording sessions, aligned to time of cue and sorted by brain regions and product of modulation index and significance level. Shaded red interval indicates 500 ms window used to compute modulation index. (b) Schematic of sagittal brain atlas showing recorded region categories in colour. Abbreviation in parentheses indicate included subregions according to Allen Mouse Common Coordinate Framework. Black lines represent Neuropixels trajectories from different sessions approximating positions of active electrodes. (c) Volcano plot shows cue modulation index (log) and significance level (-log10) for each recorded neuron, color-coded by its brain region. Grey neurons are not significant after Bonferroni correction (p > 1.3 x 10 -06 ). Circled dots indicate example neurons from each region with spike rasters shown on the right (matching colours). (d) Spike rasters show spiking activity (dots) across all cued reach trials (rows) from individual neurons, aligned to time of cue (dashed vertical line). Vertical scale bar: 100 trials, horizontal scale bar: 1 s. (e) Scatter plot shows fraction of significantly positively modulated (index > 1) and negatively modulated (index < 1) neurons in each region (colour). Each transparent dot indicates an individual recording session, circled dots represent mean and standard error of mean across sessions per region. (f) Bar graph shows pooled number of neurons recorded in each region, subdivided into their activity change upon cue.
By observing mice performing a water-reaching task, they compared "cued" reaches (where a reward is guaranteed) with "spontaneous" reaches (where a reward might be absent).
The researchers used several mathematical layers to peel back these signals:
Principal Component Analysis (PCA): This technique extracts latent dynamics (stable, low-dimensional patterns of population activity). These patterns represent the collective state of the neural network. The authors found these patterns were remarkably consistent across different animals and sessions .
Figure 3: Consistent reach-related latent dynamics across sessions and regions are modulated by reward availability and reach distance. (a) Line plots show trial-averaged PC1 and PC2 dynamics during cued long reaches of each session (light blue lines) and animal-wise average (dark blue line) across animals (rows). PC dynamics were obtained from a PCA computed on trial-averaged activity using 20/80 cross-validation and scaled to the maximum peak of each session. The rightmost column shows state-space trajectories of individual sessions and of the animal-wise average. Three sessions with cross-validated R² < 0.8 were excluded. Shaded areas represent bootstrapped 95% confidence intervals (CI). (b) PSTH shows trial-averaged and z-scored firing rate of pooled neurons (rows) during cued reaches from all sessions, sorted by PC1 coefficient and brain region. (c) CCA was performed between each pair of sessions to quantify degree of similarity between sessions after alignment, lines indicate average canonical correlation across all session pairs for the top 10 CCA components of PCA dynamics calculated on trial-averaged activity of different movement types (colors) or shuffled control (dark gray). (d) Line plots show the mean fraction of variance explained (R²) by the top ten PCs across sessions for a PCA trained on cued long reaches. Thin lines indicate cross-condition projections, showing how well activity from other movement types (colors) is captured in this PCA subspace. Thick line indicates within-condition projections (cued long reaches), where R² reflects the variance explained for held-out test trials (20%) projected into a PCA subspace computed from training trials (80%, 20/80 cross-validation). (e) Trial-averaged PC1 traces of different reach types (color code as in panel c) were obtained by projecting activity into the PCA subspace trained on either rewarded cued long reach (left) or rewarded spontaneous long reach (right) and normalized within each session to the peak amplitude of PC1 in that training subspace. Thick lines represent across-session mean. As in panel d, within-type projections show 20/80 cross-validation. Bar plots show average of normalized PC1 peak amplitude across sessions, error bars indicate 95% CI. Asterisks show significant differences based on sign-rank test: **p < 0.01, ***p < 0.001. (f) Top: Movement position along the anterior-posterior (AP) axis for different movement types (colors), aligned to movement onset. Lines indicate the mean across sessions (lightness show reach distances). Bottom: Corresponding activity along PC1 trained on cued long reaches, normalized to the peak amplitude of long cued reaches within each session. Only sessions with at least five trials per reach type were included. Shaded areas represent 95% CI.
Generalized Linear Models (GLM): The authors used GLMs to quantify the unique contribution of each variable. They treated kinematics (like paw distance and speed) and outcome (reward presence) as separate predictors. This allowed them to mathematically "regress out" the movement to see what remained.
Unsupervised Clustering (Rastermap): The authors used Rastermap to group neurons based on shared temporal firing patterns. This allowed them to discover functional subpopulations without researcher bias.
Evidence for a distributed outcome signal
The results suggest the brain's "plan" is a dynamic, distributed process. The dominant latent subspace (the primary pattern of population activity) is strongly influenced by both movement and reward. Specifically, the paper finds that reward availability explains 26% of the variability in the first principal component (PC1) peak .
Figure 4: Independent contribution of outcome and kinematic encoding to population activity across regions. (a) Example continuous data of one session with PC1-3 expression over time (trained on cued reaches, gray lines) overlaid with GLM fits (red lines) and behavioural predictors below. Lines show continuous predictors, including position and speed. Vertical lines indicate event times, such as cue and reach onset and shaded bars show time periods of reward presence and consumption. These discrete events are modelled as kernels or step functions. (b) Bar graphs show mean R² (unique contribution based on leave-one-out approach, see Methods) per predictor for PC1-3 across sessions (n = 36 sessions, error bars indicate bootstrapped 95% CI). (c) Trial-averaged region-specific PC1 traces of different reach types (colors), obtained by projecting activity into the PCA subspace trained on within-region activity during rewarded cued long reaches. Lines represent across-session mean, shades represent 95% CI. (d) Heatmap shows session-mean of unique R² for each predictor (rows) in GLM fitted on PC1-3 activity calculated within individual regions (columns). The bottom row shows the outcome selectivity, defined as the ratio of explained variance by outcome predictors (reward presence and consumption) relative to kinematic predictors (movement onset, position, speed) for each region. (e) Sagittal schematic of recorded brain regions, color-coded by the across-session mean R² of reward presence on region-specific PC1 activity. (f) Scatter plots show relative amount of variance explained by kinematic vs. outcome encoding parameters on top PCA components within each region (colors). Dots represents mean across sessions, error bars show 95% CI, dashed line represents slope of 1 with equal contributions.
This means roughly one-quarter of the primary population signal is driven by the reward itself rather than just the movement.
Crucially, the authors demonstrate that this is not just a byproduct of moving differently when hungry. They performed a kinematic similarity analysis. They compared rewarded and unrewarded reaches that were nearly identical in speed and trajectory. Even when the physical movements were matched, the neural activity (PC1) remained significantly higher for rewarded reaches .
Figure 5: Kinematic similarity analysis shows persistent outcome encoding in matched reach trajectories. (a) Example session data showing top and side view of movement trajectories of long reaches during the 'approach period', defined as a period of continuous decrease in paw-droplet distance. Reaches were color coded by the presence (teal) or absence of reward (pink), irrespective of cue. Traces show trial-averaged data from this example session of reaches in both conditions for z-scored PC1-2 activity, speed and position across all axes. Vertical black lines show the approximate boundaries of the 'approach period' during which the averaged PC Δ firing rate was calculated (black area between curves). Note that the approach period was individually determined for each reach pair (see Methods). Shaded error indicates bootstrapped 95% CIs. (b) Overlay of all rewarded (teal) and unrewarded (pink) long reach trajectories during the approach period, shown in top and side view and aligned to start position (black dot). Data include sessions with at least five trials of each reach type (n = 27). Initial reach angles (thick lines) were calculated from start point to the point where the forepaw had travelled 10 mm along the path. (c) Within each session, rewarded-unrewarded reach pairs were binned by difference in kinematic features (x-axis of each panel) and the mean difference in PC1 activity (y-axis) was calculated for each bin. Bins close to zero on the x-axis represent reach pairs had nearly identical values in the respective kinematic feature. The black lines show the across-session difference in PC1 calculated from individual session-means, shaded error represents 95% CI at session level. Light gray lines show the across-session mean difference after shuffling reward labels. Histograms below each line plot show total number of reach pairs in each bin from all included sessions (n = 27). A distinct subpopulation of neurons encodes outcome-related information during movement.
The study also identifies a specific "driver" for this signal. The unsupervised clustering revealed a distinct subpopulation of neurons, labeled RP1, that specializes in encoding reward presence .
Figure 6: Region-specific prevalence of outcome-encoding subpopulation. (a) Example session PSTH showing trial-averaged and normalized firing rate of neurons sorted and clustered by Rastermap (horizontal lines, see Methods). Activity during different reach types are shown (columns). Neural activity was aligned to the end of reach, defined as the end of the 'approach' period. In rewarded reaches, this time point corresponds to the time when the droplet is brought to the mouth, separating the prior time period of reward presence (blue shaded area) from the subsequent period of reward consumption (purple). Traces on the right show overlaid trial-averaged activity of each cluster during different reach types (color). (b) Traces show the across-session mean activity of the top reward-presence cluster (RP1), defined in each session as the cluster with the highest GLM encoding of reward presence (see panel c and Methods). RP1 activity during different reach types (colors) are overlaid. Bar plots below show across-session mean of RP1 peak firing rate across long reach types. Shaded areas and error bars represent 95% CI. Asterisks show significant differences based on sign-rank test: ***p < 0.001. (c) Bar graphs show explained variance (R²) of each GLM predictor rastermap clusters 1-6 (adjacent bars) in the example session from panel a. The cluster with the highest reward presence R² (cluster 1) is labelled as 'Reward-Presence 1' (RP1, black bars) (d) Bar graphs show mean R² per predictor for RP1 (black) and other clusters that were sorted by the amount of variance explained by reward presence within each session (n = 36 sessions, error bars indicate bootstrapped 95% CI). (e) Bar plot shows across-session mean RP1 fraction relative to all neurons within each region (x-axis, color) and session. Horizontal gray line represents the average overall fraction of RP1 neurons (18%). Error bars indicate bootstrapped 95% CI. Statistical significance was determined using an LME with animal and session as nested random effect, *p < 0.05, ***p < 0.001. (f) Heatmap shows session-mean of unique R² for each predictor (rows) in GLM fitted on average RP1 activity within each region (columns). Outcome selectivity represents the ratio of explained variance between outcome to kinematic encoding. (g) Scatter plot of PC1 and PC2 coefficients of RP1 (blue) and other neurons (gray), each transparent dot represents a neuron. Vertical and horizontal line plots represent density distributions of each population.
The authors report that these RP1 neurons are not spread evenly. They are significantly enriched in frontal cortices and motor thalamic regions . These neurons contribute disproportionately to the global latent dynamics. They act as key nodes that broadcast outcome expectations across the brain.
Limitations of the reach-based model
While the findings provide a powerful framework, there are clear boundaries to the scope of this work. First, the study relies on head-fixed mice. Head-fixation is a standard requirement for high-density electrophysiology. However, it limits the ecological validity (the extent to which results apply to real-world settings). A mouse in a natural environment moves its head and body in ways that may alter how these signals are integrated.
Second, the findings are tied to a specific water-reaching task. The neural signatures of reward expectation might differ depending on the nature of the stimulus. It remains to be seen if this "actionable representation" follows the same mathematical structure during more complex tasks. Examples include navigating a maze or interacting with social cues.
Finally, the study characterizes the presence of the signal but does not investigate the mechanism of communication. We see that the frontal cortex acts as a hub. However, the paper does not explore the precise synaptic or circuit-level logic. We do not yet know exactly how these RP1 neurons shape the global manifold.
The verdict: A new blueprint for neural integration
The evidence suggests the brain does not process "what to do" and "how to do it" in isolation. Instead, outcome expectations are woven directly into motor-related population dynamics. This provides a compelling answer to why movement-related neural activity often contains more variance than simple kinematics can explain.
If you are building models of neural computation, the takeaway is clear. You cannot treat motor execution as a closed loop. The "intent" is a distributed, continuous signal. It is actively updated by the outcome of the action. The study implies that the brain uses a shared, low-dimensional manifold to multiplex motor commands and cognitive goals. This shifts the focus from individual "command neurons" to the coordination of functional subpopulations like the RP1 cluster.
Figures from the paper
Figure 1: Performance on a goal directed reaching task leads to cued and spontaneous movements with different reward-related outcomes. (a) Schematic of head-fixed recording setup and water reaching task. Cue and droplet are presented after 3-6 s holding period, red shaded area illustrates reaction time until reach onset (bar release) and reach duration until spout touch. (b) Front and side video allowed projection of automatically estimated joint position into 3D space. Black trajectories represent movement of left digit II from the side view and top view during one session. White dot and line represent rest bar. Orange dot shows spout position, blue dot indicates location of reward consumption. (c) Side view of movement trajectories of left digit II from example session at 1:1 scale. Holding bar is located at bottom left, droplet position is at top right. Colors indicate different movement types, lightness categorizes maximum reach distance within 500 ms. (d) Example session data showing top and side view of movement trajectories of long reaches during the 'approach period'. Reaches were color coded by the presence (teal) or absence of reward (pink), irrespective of cue. (e) Example session trial-averaged traces of reaches in both conditions for kinematic parameters. (f) Boxplots show distribution of session-averages (gray lines) in kinematic features between long rewarded (teal) and long unrewarded reaches (pink). ***p < 0.001, calculated by LME.