The field of language models and cognitive science is rapidly evolving, with a focus on developing more accurate and interpretable models of human language and cognition. Recent research has explored the use of large language models to represent conceptual meaning and predict human behavior, with applications in areas such as natural language processing and human-computer interaction. Notably, studies have shown that language models can learn to represent physical concepts and predict outcomes in complex scenarios, and that multimodal approaches combining language models with other data sources can improve performance and interpretability.

Some noteworthy papers in this area include: Language models align with brain regions that represent concepts across modalities, which investigates the relationship between language models and brain activity. Deep Language Geometry: Constructing a Metric Space from LLM Weights, which introduces a novel framework for constructing a metric space of languages from large language model weights. VisionLaw: Inferring Interpretable Intrinsic Dynamics from Visual Observations via Bilevel Optimization, which proposes a bilevel optimization framework for inferring interpretable expressions of intrinsic dynamics from visual observations.