Give your AI
Research Superpowers

Three papers anchor the area: RAG [1] introduced the retrieve-then-generate recipe, REALM [2] pretrained retrieval end to end, and Fusion-in-Decoder [3] scaled it across many passages. More recent work makes retrieval adaptive: Self-RAG [4] lets the model decide when to retrieve and critique what it generates. Here is RAG's setup, in its own words:

Self-RAG [1] already trains a model to decide when to retrieve and to critique its own output with reflection tokens, so “know when to retrieve” on its own won't read as novel. The open gap: that decision is brittle under domain shift, and almost nobody evaluates it cost-aware. A top-venue contribution needs a measurable delta over Self-RAG, not a re-framing.

Reviewers will expect three baselines: BM25, a dense retriever like DPR [1], and a no-retrieval control. Report exact match and F1 on Natural Questions [2] and TriviaQA. Ablate retrieval depth k and the reranker. In this area the most common desk-reject is dropping the no-retrieval control, so keep it in.

Drafted, with each citation checked against the source so nothing is misattributed: