Building Compositional Tasks with Shared Neural Subspaces
How the brain flexibly performs multiple tasks by compositionally combining task-relevant neural representations in shared subspaces of neural activity.
Nature · Volume 650 · 2026Experience the three compositional categorization tasks (S1, C1, C2) used in the monkey experiments. Categorize stimuli by shape or color and respond on the correct axis.
Read the complete text with all figures, including main results on shared sensory and motor subspaces, cross-task generalization, and task belief encoding.
Explore the theoretical framework with interactive simulations. Adjust parameters to see how shared subspaces, gain modulation, and task belief affect neural representations.
Examine key results with simulated data reproducing the paper's main findings: cross-task decoding, sensory-motor transformation, and dynamic subspace engagement.
Explore proposed future experiments, analyses, and extensions discussed in the paper, plus additional ideas for computational and experimental work.
External links to the original paper, data, and code.
Key Findings
Core discoveries from neural recordings in five brain regions during three compositional tasks
Shared Sensory Subspaces
The same subspaces of LPFC neural activity represent colour category information across the C1 and C2 tasks. Classifiers trained on one task generalize to the other, indicating a shared representational geometry.
Shared Motor Subspaces
Motor response representations in the S1 and C1 tasks occupy the same neural subspace. This shared motor code was found across all recorded brain regions (LPFC, FEF, PAR, aIT, STR).
Sequential Transformation
Tasks engage shared subspaces sequentially: sensory information in the shared colour subspace is transformed into the appropriate motor subspace in a task-specific manner.
Task Belief Modulation
Monkeys maintain an internal belief about the current task. This belief dynamically scales the engagement of shared sensory subspaces — amplifying relevant and suppressing irrelevant representations.
Gain Modulation / Compression
Task-relevant stimulus dimensions are amplified while irrelevant dimensions are compressed. The Compression Index (CPI) correlates with the strength of task belief in LPFC.
Compositionality
The C1 task is composed from components shared with C2 (colour categorization) and S1 (axis 1 response). This compositional structure enables flexible behavior and potential for rapid learning.
Task Design Overview
Three compositionally related categorization tasks sharing sensory and motor components