English Español Français Deutsch
← Back to the list
Retrieval-augmented generation On-mission junior researchers Directly upstream of ai100's methodology

Akari Asai

How a language model decides that its own internal knowledge is insufficient and that it should reach for an external source instead.

Most retrieval-augmented systems are wired rigidly — always query the external store, then answer. Asai's Self-RAG (2023) showed that the model can be trained to make the decision itself: whether retrieval is needed here, what kind, and whether to trust what comes back. For ai100 this is a directly upstream line of work — modern AI engines in web-search mode already operate this way, and the decision to mention or omit a brand often happens at the retrieval layer rather than in the final generation.

Worth following when
you need to understand why a model in web-search mode includes one source and ignores another on the same query.
Topics
active retrieval in LLMs; adaptive RAG; the robustness of retrieval decisions to adversarial or noisy contexts.
Key works
Self-RAG (2023, lead author); OpenScholar (2024, lead author); CRAG critic-corrective RAG (2024).