GAR Generation-Augmented Retrieval for Open-Domain Question Answering
Algorithms AIMech:GARGenerationAugmentedRetrievalForOpenDomainQA DRAFT
Aliases: GAR, Generation-Augmented Retrieval, GAR and GAR+ Multi-Context Query Expansion for Open-Domain QA
A query-expansion and retrieval algorithm for open-domain question answering that trains sequence-to-sequence pre-trained language models (such as BART) via standard seq2seq supervision to generate three complementary in-domain query contexts from an input question q—(1) the default target answer, (2) the passage sentence containing the answer, and (3) the title of a passage containing the answer—appends each generated context to q to form separate generation-augmented queries, retrieves top candidate passages for each augmented query using sparse lexical retrieval (BM25) and fuses the retrieved rankings (or fuses them with DPR dense retrieval rankings as GAR^+), and aggregates candidate answer spans across retrieved passages in the extractive reader via Passage-level Span Voting by summing normalized span probabilities p(S_i[j]) = softmax(D)[i] * softmax(S_i)[j] over all spans sharing the same surface string.
Identity Notes & Concept Scope
Distinct from RM3 pseudo-relevance feedback, which expands queries using terms from initially retrieved passages rather than parametric knowledge generated by a seq2seq PLM; distinct from HyDE (AIMech:HyDEHypotheticalDocumentEmbeddings), which encodes zero-shot LLM-generated hypothetical documents into dense vectors with an unsupervised contrastive encoder; and distinct from RAG and FiD, which use a generator as the downstream reader over DPR-retrieved passages rather than generating multi-target query expansions prior to retrieval.
Claims & Grounded Evidence
4 Claims
Mechanism
#gar-multi-target-query-context-generation-and-fusion
GAR trains seq2seq generators using freely available in-domain supervision to generate three distinct query contexts—(1) Answer, (2) Sentence containing the answer, and (3) Title of a positive passage—appends each to the question to form generation-augmented queries, and fuses their separate BM25 retrieval lists (Sections 3.2-3.3), substantially increasing query-passage ROUGE-1/2/L overlap on Natural Questions from 6.00/2.36/5.01 to 13.21/6.99/10.27 (sentence) and mitigating hallucinated answers via complementary contexts (Tables 2-3, Figure 1).
Mechanism
#gar-passage-level-span-voting
For extractive reading over retrieved passages D = [d_1, ..., d_k] with passage relevance logits D and top-N span logits S_i per passage, GAR computes normalized span probabilities p(S_i[j]) = softmax(D)[i] * softmax(S_i)[j] and sums them across all candidate spans sharing the same surface string across all retrieved passages (Passage-level Span Voting, Section 4.1).
Application
#gar-openqa-benchmark-applications
On Natural Questions (NQ) and TriviaQA, GAR with BM25 outperforms DPR at Top-500/1000 on NQ and at all k >= 5 on TriviaQA while training and indexing at a fraction of DPR's compute cost (Tables 4 and 8), and fusing GAR with DPR (GAR^+) reaches 88.9% Top-100 retrieval accuracy on NQ and 43.8 extractive EM / 45.3 generative EM (Tables 4, 6, and 7).
Limitation
#gar-small-k-and-multi-generator-overhead-limitation
Because GAR relies on sparse BM25 matching over queries expanded with potentially noisy PLM-generated contexts, standalone GAR underperforms dense retrieval (DPR) on Natural Questions at very small retrieval cutoffs (Top-5: 60.9% vs. 68.3% in Table 4, which also limits generative reader EM when only 10 passages are read in Table 6) and requires training and running three separate BART-large generators for the three context types.
⚠️ Curated Limitations & Negative Results
- Standalone GAR with BM25 trails dense retrieval (DPR) at very small retrieval cutoffs on Natural Questions (e.g., Top-5 accuracy 60.9% vs. 68.3% in Table 4), limiting generative readers that only consume a few top passages unless fused with DPR (GAR^+).
- Requires training and executing three separate seq2seq generators (for answer, sentence, and title targets) and running three separate BM25 retrievals per query before rank fusion.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
related_match | GAR uses neural sequence-to-sequence Transformer language models (BART) to generate query expansion contexts and a neural BERT extractive or BART generative reader. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
r769-primary-pdf
|
Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, Weizhu Chen
|
primary paper
full_text
|
ID: https://arxiv.org/bibtex/2009.08553
sha256:6f5c902858c8…
|
r769-arxiv-abs
|
Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, Weizhu Chen
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/2009.08553
sha256:5c7a5b29954a…
|
r769-metadata-cr
|
Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, Weizhu Chen
|
primary paper
abstract_only
|
ID: https://arxiv.org/bibtex/2009.08553
sha256:68b1ab497bda…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T14:12:00Z
Updated: 2026-10-19T14:12:00Z
Source File:
data/curated/GARGenerationAugmentedRetrievalForOpenDomainQA.yaml
· Schema:
curated_concept.yaml
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories