Is This Just Fantasy? Language Model Representations Reflect Human Judgments of Event Plausibility

Researchers have identified 'modal difference vectors' in language models that can distinguish between possible, impossible, and nonsensical statements, revealing better modal categorization abilities than previously thought. The study shows these vectors emerge consistently as models become more capable and can even predict human judgment patterns about event plausibility.

• →Language models possess more reliable modal categorization abilities than recent studies suggested, accessible through linear representations called modal difference vectors.
• →Modal difference vectors emerge in a predictable order as models become more competent through training steps, layers, and parameter scaling.
• →These vectors can model fine-grained human categorization behavior and correlate with human ratings of interpretable features.
• →The research uses mechanistic interpretability techniques to provide new insights into how both AI and humans distinguish between modal categories.
• →Findings challenge previous assessments that questioned language models' ability to categorize sentences by modality.