Fig. 7: Model performance.

a Difference in 90th percentile response time (disengaged - engaged) across mice for models with (purple) and without GLM-T (blue), indicating a higher separation of engaged and disengaged states in terms of response times for the model with GLM-T. This was for n = 37 mice (biological replicates; independent IBL animals). Sessions/trials are replicates aggregated within mouse. Points show the per-mouse difference in the 90th percentile response time. Error bars denote 95% bootstrap percentile confidence intervals (2.5–97.5th) computed from 5000 resamples of trials within each mouse. b Average state probabilities across trials for all mice, showing an initial warm-up effect where mice start in disengaged states (orange) before transitioning to engaged states (green). Additionally, this suggests a link between disengagement and satiety, as greater past rewards predict transitions to disengaged states, leading to an increased occupancy of disengagement states toward the end of the session when mice have accumulated more rewards throughout the experiment. c Relationship between filtered reward and the probability of being in the disengaged state, showing a positive correlation between them (light blue points are data points and blue line is the trend line). d Test log-likelihood (LL) comparison between models for synthetic data sampled from a model without GLM-T (left) and a model with GLM-T (right). When data has trial-dependent transitions, the GLM-T model performs better, improving by 0.009 bits per trial. Conversely, when transitions are independent of trials, the non-GLM-T model performs almost similar. The relative improvement, shows that it provides a meaningful advantage for data with transitions. Models were run on synthetic data produced for n = 37 mice by reconstructing each mouse’s best-fitting model and simulating session-by-session using the true stimulus sequences. Data are presented as mean test LL; error bars denote  ± standard deviation across 5-fold cross-validation. e Scatter plot of inferred states with and without GLM-T, with a reference y = x line (red dashed). Most points lie above the red line, indicating that GLM-T states are more probable than non-GLM-T states. Source data are provided as a Source Data file. Diseng., Disengaged.