Proceedings of the Annual Meeting of the Cognitive Science Society

Large language models (LLMs) excel at personality prediction but are often impractical for deployment due to high computational demands. This work introduces cognitive distillation with Chain-of-Thought calibration, a novel framework for transferring structured reasoning from large LLMs to smaller, efficient models. Inspired by cognitive architectures like ACT-R, our method aligns intermediate inference steps using exemplars from single or multiple LLMs. A 1.5B-parameter model distilled through this process surpasses Qwen1.5-110B in predictive accuracy, achieving a 28% improvement in Pearson correlation while using just 1.36% of its parameters. Ablation studies reveal that moderate sampling diversity and multi-model ensembles enhance cognitive skill transfer and construct validity. These findings demonstrate that high-level reasoning can be efficiently and faithfully transferred, enabling psychometrically robust personality assessment in resource-constrained settings. This approach bridges AI and cognitive science, offering a scalable path toward plausible, cognitively grounded language models.