Evaluating TRIBE v2 Beyond Its Original Benchmarks: From Functional Localizers to Naturalistic Emotion

Brain foundation models that predict neural responses from multimodal stimuli offer a potential route toward in-silico neuroscience. Recent models such as TRIBE v2 have reproduced several well-established functional contrasts, but generalization to independent, dynamic paradigms and distributed affective phenomena such as emotion remains unclear. Here, we evaluate TRIBE v2 beyond its original benchmarks using two complementary settings: a dynamic audiovisual multifunction localizer spanning perceptual and higher-order cognitive contrasts, and naturalistic affective movie stimuli with continuous emotion annotations. We compare TRIBE-predicted and measured fMRI response maps under a consistent evaluation framework. TRIBE v2 reproduced several perceptual contrasts as well as Multiple Demand and Theory of Mind responses, but showed far weaker correspondence for controlled auditory-language contrasts. In naturalistic movie viewing, model-brain correspondence varied widely across emotion categories. Together, these results show that TRIBE v2 generalizes to several independent perceptual and higher-order contrasts, but that model-brain correspondence is highly domain-specific, with selective failures in auditory language and some affective conditions.