Nov 21, 2023 GS.09
Recent work with OpenAI’s GPT language models concludes that analogical reasoning, using what you know about one thing to infer knowledge about a new, somehow related instance, is emerging in these systems. My work shows something different: the newest GPT models are autoregression at its best, excelling at next-word prediction, but they can’t generalize what they’ve learned to novel domains, and they are not “reasoning”. I will present two studies on analogical reasoning and demonstrate through a series of experiments comparing GPTs performance to children that analogical reasoning has not emerged in these systems. I will go on to discuss why I do not think analogical reasoning or the ability to generalize knowledge will emerge with the current language models’ architecture and training goals.