English Español Français Deutsch
← Back to the list
Information retrieval Methodologists Directly upstream of ai100's methodology

Julian McAuley

Whether language models trained on the open web are already doing recommendation — and what that implies for products that compete with traditional recommenders.

The standard recommender system was a specialized model trained on a particular catalog and user-behavior history. McAuley's 2023 work on "Large Language Models as Zero-Shot Conversational Recommenders" demonstrated that an off-the-shelf LLM, with no recommender-system training at all, performs competitively on conversational-recommendation benchmarks — partly because the training data contained a billion implicit recommendations from forums, reviews, and listicles. For ai100, this is the precise mechanism behind "the model mentions brand X and not brand Y in answer to an open-ended question" — it's conversational recommendation in everything but the label.

Worth following when
you want to understand the recommendation behavior buried inside open-ended LLM responses, and why it isn't always called by that name.
Topics
LLMs as zero-shot conversational recommenders; recommendation behavior in open-ended LLM outputs; the training-data conditions that made out-of-the-box LLM recommendation work.
Key works
"Large Language Models as Zero-Shot Conversational Recommenders" (2023, senior author); Amazon product-review datasets (mid-2010s onward); foundational publications connecting NLP and recommendation.