UCLA

This dissertation brings together four studies. The first two address practical representation, drawing on programming-language semantics and motor representation. The latter two are collaborative studies that develop and test methods for evaluating psychological depth in large language models. The first study examines Carlotta Pavese’s account of practical modes of presentation, in which operational semantic values serve as practical Fregean senses and algorithms as their referents. I show that the account does not establish how distinct operational semantic values can be more fine-grained than an algorithm while sharing it as their referent. I argue that algorithms cannot occupy the intermediate level of granularity between operational semantic values and the functions they compute. The second study develops a positive account of practical representation as a distinctive representational kind. I frame the central dispute as one between a mode of presentation view, in which a task is represented in a distinctively practical way, and a mode of use view, in which ordinary task representations are practically deployed in action. I examine how motor representational format bears on this dispute. I then take Pavese’s capacity-relative view of practical modes of presentation as the starting point for a positive account, and I introduce practical constancies to specify the capacities that play the relevant constitutive role. I propose that motor equivalences instantiate a distinctively practical form of representational constancy by preserving the same operationally articulated task across variations in execution. The third study introduces the Psychological Depth Scale, which is a reader-centered framework for assessing five dimensions of human- and LLM-generated fiction. The study first tests whether human readers can apply the PDS consistently, then examines whether those judgments can be approximated automatically. It finally uses the scale to compare human- and LLM-authored fiction, finding that GPT-4-generated stories score higher than highly rated human-written stories on empathy and narrative complexity and comparably on the remaining dimensions. The fourth study asks whether an LLM-based evaluator validated against human judgments on this original dataset continues to track human judgments when deployed on closely matched frontier-model outputs. Human readers show no uniform preference for reasoning-enhanced models, whereas the automated judge shows an almost uniform preference for reasoning-model outputs. Agreement with human ratings on one dataset therefore does not guarantee that an automated evaluator will track human judgments on a different distribution of outputs, especially when those judgments are subjective. Through philosophical and empirical inquiry, the dissertation seeks to address questions concerning meaning, computation, and reasoning.