AuthorsPatrick Lewis, Ethan Perez, Aleksandra Piktus et al.
TL;DR
RAG combines a DPR retriever with a BART generator to marginalize over retrieved Wikipedia passages and reach 44.5 EM on Natural Questions, beating DPR by 3.0 points.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Large pre trained language models store factual knowledge in parameters but still have limited ability to access and precisely manipulate knowledge, especially on knowledge intensive tasks.
On open domain QA benchmarks like Natural Questions and TriviaQA, purely parametric seq2seq systems trail specialized retrieve and extract architectures, leading to lower accuracy and no provenance or easy knowledge updates.
HOW IT WORKS
RAG combines a Retriever pη(z|x) based on DPR, a Generator pθ(yi|x,z,y1:i−1) using BART, and two formulations RAG-Sequence and RAG-Token that treat retrieved documents as latent variables.
You can think of RAG like a CPU using fast parametric memory as registers and an external non parametric memory index as disk, fetching relevant Wikipedia chunks on demand instead of memorizing everything.
By marginalizing over top K retrieved passages per sequence or per token, RAG accesses more evidence than a fixed context window and generates answers even when no single passage contains the exact span.
DIAGRAM
This diagram shows how RAG Token retrieves top K passages once and marginalizes over them at each decoding step to compute token probabilities.
DIAGRAM
This diagram shows how RAG is evaluated on Natural Questions, TriviaQA, WebQuestions, CuratedTrec, MS MARCO, Jeopardy question generation, and FEVER.
PROCESS
Retriever DPR
RAG uses the Retriever pη(z|x) based on DPR to encode the query with BERTq and retrieve top K Wikipedia passages via Maximum Inner Product Search.
Generator BART
RAG feeds the input x concatenated with each retrieved passage z into the Generator pθ(yi|x,z,y1:i−1) implemented with BART large to compute token probabilities.
RAG Sequence Model
RAG Sequence treats a single retrieved document as a latent variable for the whole sequence and marginalizes pη(z|x)pθ(y|x,z) over the top K passages.
RAG Token Model
RAG Token allows a different latent document per token, marginalizing over pη(z|x)pθ(yi|x,z,y1:i−1) at each step and decoding with a standard beam search.
KEY CONTRIBUTIONS
Retrieval Augmented Generation models
RAG introduces RAG Sequence and RAG Token, combining a DPR based Retriever pη(z|x) with a BART Generator pθ(yi|x,z,y1:i−1) and training them end to end over latent documents.
State of the art open domain QA
RAG achieves 44.5 EM on Natural Questions and 68.0 EM on TriviaQA Wiki, surpassing DPR and T5 11B plus SSM without specialized salient span masking pre training.
Knowledge intensive generation improvements
RAG improves MS MARCO Bleu 1 by 2.6 points over BART and yields more factual and specific Jeopardy questions, with human evaluators preferring RAG in 42.7 percent of factuality comparisons.
RESULTS
Natural Questions EM
44.5
+3.0 over DPR
TriviaQA Wiki EM
68.0
+10.1 over DPR
WebQuestions EM
45.2
+4.1 over DPR
MS MARCO Bleu 1
44.2
+2.6 over BART
These metrics come from open domain Natural Questions, TriviaQA Wiki, WebQuestions, and open MS MARCO NLG, showing that RAG consistently improves Exact Match and Bleu over DPR and BART on knowledge intensive QA and generation.
BENCHMARK
These metrics come from open domain Natural Questions, TriviaQA Wiki, WebQuestions, and open MS MARCO NLG, showing that RAG consistently improves Exact Match and Bleu over DPR and BART on knowledge intensive QA and generation.
BENCHMARK
Exact Match on Natural Questions open domain QA comparing RAG to T5 and DPR.
KEY INSIGHT
RAG achieves 11.8 percent accuracy on Natural Questions even when the correct answer is not present in any retrieved document, where extractive systems score zero.
This is surprising because retrieval augmented systems are usually assumed to rely entirely on retrieved spans, but RAG shows parametric memory can still generate correct answers without explicit textual support.
WHY IT MATTERS
RAG enables a single architecture to handle open domain QA, abstractive QA, Jeopardy question generation, and FEVER fact verification using the same retriever and generator.
Builders can now plug different document indices into RAG to update world knowledge without retraining, and jointly fine tune retrieval and generation for new knowledge intensive applications.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Yijie Zhong, Yunfan Gao, Haofen Wang
· 2026
HingeMem combines Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment dialogues into element-indexed hyperedges and plan query-specific retrieval. On LOCOMO, HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge score, surpassing the best baseline Zep (56.9 F1) by 7.0 F1 without using category-specific QA formats.
Answers use this explainer on Memory Papers.
Checking…