Multimodal temporal modeling of behavioral reorganization across threat learning and extinction in rats

After threat disrupts a learned activity, does engagement return as defensive responding later declines? Tri-reference-point theory describes decisions under risk relative to a minimum requirement, the status quo and a goal; we extended this logic to defense and engagement. We followed freezing, 22-kHz and 50-kHz calls, and food-reinforced lever responses across conditioning, extinction, delayed retesting and context changes in 29 male rats (45,300 five-second windows). A hidden Markov model described recurring response combinations, or configurations, and their persistence. For every animal, it predicted behavioral sequences excluded from fitting better than a matched model that ignored temporal order. Among models with two to eight configurations, eight predicted best; four provided a coarser description. This coarser model separated relatively brief press-active patterns from more persistent vocal-active and low-output patterns. Low-output patterns contained few calls and lever responses, but could involve either little or substantial freezing. In the repeated-session comparison, the estimated share of time in press-active patterns increased on average in the full extinction group (n=9) and decreased in cue-free context controls (n=8), from different first-session levels. These descriptive trajectories included renewed instrumental engagement alongside continued defensive immobility. We separately tested whether configuration summaries added predictive value beyond the measured behaviors for later responding or whole-recording fecal-pellet counts, finding no consistent benefit. Tracking how behavioral combinations change and persist provides a computational framework for evaluating functional adaptation beyond defensive decline.