Uncovering the semantics of teaching in experiential learning with Large Language Models

Language is the cornerstone of teaching, but we still lack principled ways to identify what makesexplanations effective in terms of their composition. In a two-stage behavioural paradigm,“Teacher” participants learned two-armed bandit tasks through trial and error and wrote free-textlessons. “Pupil” participants received these lessons before completing the same tasks. Wefound that receiving a well-constructed lesson improved experiential learning. To understandwhat made those lessons “Good”, we uncovered their latent semantic dimensions through aninferential use of large language models. These dimensions (“Memorization”, “PatternRecognition”, “Option Ranking”, “Randomness”) predicted expert judges’ ranking of lessons andbehavioural outcomes. Further, artificially-generated Good and Bad lessons based on thesedimensions replicated the observed effect of good explanations on experiential performance.Overall, results show that the semantic structure of instructional language measurably shapeshow people learn from experience, and that those structures can be discovered, interpreted,and used to refine our understanding of the cognitive mechanisms underlying teaching andlearning.