Home / Curated / Algorithms / UPR (Unsupervised Passage Re-ranking with Zero-Shot Question Generation)

UPR (Unsupervised Passage Re-ranking with Zero-Shot Question Generation)

Algorithms AIMech:UPRUnsupervisedPassageRerankingWithQuestionGeneration DRAFT

Aliases: Unsupervised Passage Re-ranker, Unsupervised Passage Re-ranking, UPR Zero-Shot Question Generation Re-Ranker

An unsupervised zero-shot passage re-ranking algorithm for information retrieval and open-domain question answering that re-orders top-K candidate passages Z = {z_1, ..., z_K} from any first-stage retriever by applying Bayes' rule under a uniform passage prior and using an off-the-shelf pre-trained language model Theta with an appended natural-language instruction prompt to compute the length-normalized conditional log-likelihood of generating the input question tokens given each passage, log p(q | z_i) = (1 / |q|) sum_{t=1}^{|q|} log p(q_t | q_{<t}, z_i; Theta).

Identity Notes & Concept Scope Introduced by Sachan, Lewis, Joshi, Aghajanyan, Yih, Pineau, and Zettlemoyer (EMNLP 2022, pp. 3781-3797, arXiv:2204.07496). Distinguished from classical unigram query-likelihood language models (Ponte & Croft, 1998; Zhai & Lafferty, 2001), from supervised relevance-token cross-encoders such as monoT5 (Nogueira et al., 2020), and from synthetic query-generation dataset augmentation methods such as Doc2Query, InPars, or Promptagator that train a downstream retriever rather than re-scoring test queries zero-shot via p(q | z_i).

Claims & Grounded Evidence

3 Claims

Mechanism #upr-bayesian-question-likelihood-mechanism

Given top-K retrieved passages Z = {z_1, ..., z_K} for a query q, UPR decomposes log p(z_i | q) = log p(q | z_i) + log p(z_i) + c via Bayes' rule, assumes a uniform passage prior log p(z_i), appends a zero-shot instruction prompt (such as 'Please write a question based on this passage') to each passage z_i, and ranks passages by the average token log-likelihood log p(q | z_i) = (1 / |q|) sum_t log p(q_t | q_{<t}, z_i; Theta), forcing token-level cross-attention to explain every token in q.

supports Source: Improving Passage Retrieval with Zero-Shot Question Generation · Locator: Section 2.2 (Unsupervised Passage Re-ranking (UPR)) & Figure 2, pp. 2-3
supports Source: Improving Passage Retrieval with Zero-Shot Question Generation · Locator: upr_question_log_likelihood & main

Scope #upr-question-vs-passage-generation-asymmetry

Re-ranking top-1000 retrieved passages with UPR (using T0-3B) improves unsupervised retrievers (MSS, BM25, Contriever) by 6% to 18% absolute and supervised retrievers (DPR, MSS-DPR) by up to 12% in top-20 accuracy across SQuAD-Open, TriviaQA, Natural Questions, and WebQuestions, whereas the reverse conditional score log p(z_i | q; Theta) = (1 / |z_i|) sum_t log p(z_{i,t} | z_{i,<t}, q; Theta) degrades accuracy below the first-stage retriever.

supports Source: Improving Passage Retrieval with Zero-Shot Question Generation · Locator: Section 4.1 (Table 2) & Section 4.2.1 (Importance of Question Generation, Figure 3), pp. 4-6
supports Source: Improving Passage Retrieval with Zero-Shot Question Generation · Locator: passage_generation_log_likelihood & main

Limitation #upr-cross-attention-latency-and-claim-query-limitation

UPR incurs inference latency that scales linearly with the number of re-ranked candidate passages K and proportionally to the product of question length, passage length, and PLM layer count (reaching 11.6 seconds per question on an A100 GPU for K = 1000 with T0-3B), is upper-bounded by the recall of the first-stage top-K retrieval pool, and can degrade nDCG@10 on fact-verification datasets (such as FEVER and Climate-FEVER in BEIR) where queries are declarative claims rather than questions.

supports Source: Improving Passage Retrieval with Zero-Shot Question Generation · Locator: Section 4.2.3 (Figure 4), p. 6; Limitations, pp. 9-10; Appendix A.4 (Table 10), pp. 14, 18

⚠️ Curated Limitations & Negative Results

High inference latency when re-ranking large candidate pools (e.g., K = 1000) due to full token-level cross-attention across a multi-billion-parameter language model.
Recall is strictly upper-bounded by the first-stage retriever's top-K candidate pool, and question-generation instructions underperform on declarative claim-verification queries such as FEVER and Climate-FEVER.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	Uses a pre-trained neural sequence-to-sequence or autoregressive transformer language model for zero-shot conditional question likelihood re-ranking.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r744-primary-pdf`	Improving Passage Retrieval with Zero-Shot Question Generation Devendra Singh Sachan, Mike Lewis, Mandar Joshi, Armen Aghajanyan, Wen-tau Yih, Joelle Pineau, Luke Zettlemoyer	primary paper full_text	ID: https://dblp.org/search/publ/api?q=Improving+Passage+Retrieval+with+Zero-Shot+Question+Generation+Sachan&h=5&format=xml sha256:3e8d200937c8…
`r744-arxiv-abs`	Improving Passage Retrieval with Zero-Shot Question Generation (arXiv abstract) Devendra Singh Sachan, Mike Lewis, Mandar Joshi, Armen Aghajanyan, Wen-tau Yih, Joelle Pineau, Luke Zettlemoyer	primary paper abstract_only	ID: https://arxiv.org/abs/2204.07496 sha256:640fa10570e3…
`r744-metadata-cr`	Improving Passage Retrieval with Zero-Shot Question Generation (Crossref metadata) Devendra Singh Sachan, Mike Lewis, Mandar Joshi, Armen Aghajanyan, Wen-tau Yih, Joelle Pineau, Luke Zettlemoyer	primary paper abstract_only	ID: https://dblp.org/search/publ/api?q=Improving+Passage+Retrieval+with+Zero-Shot+Question+Generation+Sachan&h=5&format=xml sha256:7160130d27bf…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T13:47:00Z

Updated: 2026-10-19T13:47:00Z

Source File: data/curated/UPRUnsupervisedPassageRerankingWithQuestionGeneration.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