Neural dynamics become significantly faster and more nonlinear at the peak of information processing
The temporal scale at which neural populations preserve stimulus information is often treated as fixed, even though population activity changes rapidly during sensory processing. Whether the information-optimal scale changes within a single stimulus-evoked episode, and what determines its evolution, remains unclear. We developed Integration-Split, a decoding-based framework that estimates the time-varying temporal resolution of population spike trains. Across four visual paradigms and throughout the mouse visual hierarchy, the optimal scale contracted after stimulus onset and subsequently re-expanded. This pattern persisted across stimulus temporal frequencies and was largely absent in a hippocampal control. The optimal scale varied inversely with stimulus decodability, reaching its finest resolution when stimulus identity was most accurately decoded. Beyond this association, independently estimating the signal-noise autocorrelation balance across candidate window widths identified the window that maximized the autocorrelation-based signal-to-noise ratio, and this optimal width closely followed the Integration-Split trajectory. Consistent with these findings, stronger nonlinear predictive structure coincided with finer optimal temporal resolution, consistent with the idea that shorter integration windows are needed to preserve informative nonlinear population structure that would otherwise be attenuated by temporal averaging. To test the functional relevance of this organization, we trained a Transformer solely to decode stimulus identity, without providing the optimal temporal-scale pattern or Integration-Split outputs. Its learned attention exhibited a similar contraction-and-recovery profile, and disrupting heads with stronger correspondence produced larger decoding deficits. Together, these findings reveal a dynamically changing, information-optimal temporal scale in population coding, shaped by the evolving balance of signal and noise autocorrelation and by nonlinear population structure, and independently recapitulated in the learned attention of a Transformer trained without temporal-scale supervision.