English Español Français Deutsch
← Back to the list
Hallucination and factuality Foundational figures Directly upstream of ai100's methodology

Wen-tau Yih

Making the retrieval step inside retrieval-augmented systems good enough that the generation step has something to work with.

Long before RAG existed as a term, Yih was building the retrieval substrate it depends on. Dense Passage Retrieval (2020) is the paper that made dense-vector search competitive with traditional inverted-index methods for open-domain QA — virtually every modern RAG system descends from that result, including the ones that don't credit it. Reading Yih is the cleanest way to understand why the retrieval in retrieval-augmented generation is usually where systems silently break, and why FActScore (which he co-authored) bothers to score atomic facts against an external knowledge source rather than against the model's parametric memory.

Worth following when
you want to understand why a RAG-augmented model still hallucinates — and whether the failure is in retrieval or in generation.
Topics
dense passage retrieval; open-domain question answering as a discipline; what factual grounding does and doesn't fix.
Key works
Dense Passage Retrieval / DPR (2020, co-author); WikiQA (2015, co-author); FActScore (2023, co-author).