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

Weijia Shi

Making retrieval-augmented generation work when the language model itself is a closed box you can't fine-tune.

The dominant assumption in retrieval-augmented systems is that you train the retriever and the generator together, or at least fine-tune one to fit the other. REPLUG (2023, lead author) showed that this isn't necessary — the retriever can be trained against a closed-source language model treated as a frozen black-box scoring function, and the resulting system improves the LLM's outputs without ever touching its weights. The methodological consequence is that retrieval-augmentation research stopped being something that only applied to open-weight models, which is the only condition under which it's relevant for evaluating systems like the major commercial AI engines.

Worth following when
you want to understand how retrieval improves closed-source LLMs that don't expose their weights or training signals.
Topics
black-box retrieval-augmented generation; retriever training without LM access; the architectural assumptions buried in earlier RAG research.
Key works
REPLUG: Retrieval-Augmented Black-Box Language Models (2023, lead author); In-Context RALM (2023, co-author).