Cognitive inspirations for Large Language Models

Cognitive science provides many frameworks for studying how humans learn and reason. As Large Language Models (LLMs) become larger and more complex, ideas from cognitive science become increasingly relevant. In this talk, I draw on two lines of work that connect cognitive principles to practical insights about LLM training and inference. The first line takes inspiration from a well-known phenomenon in human cognition: rephrasing a question can help elicit a better answer, especially if that rephrasing is tuned to the agent answering the question. We find that the same holds for LLMs. Across six text classification benchmarks, augmenting data with paraphrases at both train and test time consistently improves LLM performance beyond what parameter-efficient fine-tuning alone can achieve. A related intuition comes from the cumulative nature of many learning scenarios: complex problems often require learning in stages, with later competencies building on earlier ones. Using a controlled benchmark, we find that a transformer does exactly this: it acquires simple skills before progressing on to more complex ones. We also find an interesting asymmetry: learning to combine simple rules is much easier for the model than learning to take complex examples apart. Together, these results suggest that cognitive intuitions offer not just metaphors, but actionable hypotheses for understanding and improving language models.