Real-time geometric complexity of language is associated with global functioning across psychosis and major depression

Language disturbances are prominent across psychiatric disorders, yet their underlying organization and relationship to clinical functioning remain poorly understood. We conceptualized spontaneous speech as a temporally evolving sequence of contextual representations through semantic space, in which successive concepts define transitions between large-language-model embeddings. We characterized the local dynamics of these representations using Lyapunov-inspired measures of sensitivity to perturbation and anisotropy—the extent to which nearby neighboring contextual representations diverge differently depending on direction in representational space. We applied this approach to picture descriptions from 129 German-speaking participants with major depressive disorder (n = 43), schizophrenia spectrum disorders (n = 42), or no psychiatric diagnosis (n = 44). Individual differences in local geometric dynamics predicted multiple clinical dimensions, with the strongest association observed for global functioning, for which the geometric measures explained 51% of the variance. Clinically informative geometric structure was already detectable within the first three sentences of speech, where distinct dynamical profiles were associated with differences in negative and positive symptoms, anxiety, depression, and global functioning. These associations suggest that the dynamics of local representational divergence and contraction during speech reflect broad variation in cognitive and social functioning rather than diagnosis alone. Together, these findings suggest that psychiatric variation is reflected not simply in what people say, but in the dynamics by which semantic representations evolve during ongoing language production. A dynamical-geometric analysis of language may provide a compact, transdiagnostic marker of clinically meaningful variation in psychopathology.