This talk will be held online on Zoom, please register here to participate.

Abstract:

Recent work with large language models concludes that analogical reasoning, using what you know about one thing to infer knowledge about a new, somehow related instance, has emerged in these systems. My work shows something different: the newest (multimodal) large language models (LLMs) only appear to be “reasoning”, but when challenged they seem unable to generalize what they've learned to novel rules or unfamiliar domains. I will present a series of studies demonstrating how this is in stark contrast to how children learn to solve and generalize analogy solving.