Fig. 6: Trajectory-specific tuning persists across rule changes.

a Top: Schematic of original, new, and restored rule conditions. Bottom: Corresponding learning performance color-coded for individual mice (n = 6 mice, purple: new rule, green: restored rule). Circles indicate average duration ± sem to criterion. C: coconut, V: vanilla. b Trajectory-specific tuning functions of all cells during original, new, and restored rules sorted by the original rule (n = 272, only left trajectories are shown). Since the last recording day in the original rule, 16–36 and 29–68 days passed until the mice had learned the new and restored rule, respectively. c Trajectory-specific correlation to the original rule decays over time but remains significant vs. shuffled (gray) during both new (t = 10.37, p = 10⁻⁴) and restored rules (t = 5.17, p = 0.003, time-correlation interaction effect: F = 13.68, p = 0.014, two-way repeated measures ANOVA followed by paired t-tests with Šidák correction). d Models trained on data obtained during the original rule allow the prediction of the animals’ linearized position during new (versus shuffled: t = −5.30, p = 0.003) and restored rules (t = −5.40, p = 0.003, two-way repeated measures ANOVA followed by paired t-tests with Šidák correction). e Models trained on the original rule allow the decoding of behavioral choice during new (versus shuffled: t = 7.78, p = 6*10⁻⁴) and restored conditions (t = 7.76, p = 6*10⁻⁴, two-way repeated measures ANOVA followed by paired t-tests with Šidák correction). For all comparisons, left and right trajectories were correlated separately and pooled subsequently. Boxes show median and upper/lower quartiles. Circles represent individual mice (n = 6 for all comparisons). Source data are provided as a Source Data file.