---
title: "Information retrieval & search — papers"
description: "Information retrieval & search: 109 papers of the Elamur Research library, the most cited first."
canonical: "https://research.elamur.ai/free/en/topics/information-retrieval/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/topics/information-retrieval/)

[home](https://md.research.elamur.ai/free/en/) / [topics](https://md.research.elamur.ai/free/en/topics/) / Information retrieval & search

# 🔍 Information retrieval & search

109 papers · page 1 of 5 · the most cited first · [all topics](https://md.research.elamur.ai/free/en/topics/)

Finding the right documents: search engines, ranking, and retrieval for language models.

A survey · about 20 minutes to read · every paper named here has its own page in this library

**In this article**

1. [What it is](#a-what-it-is)
2. [History and milestones](#a-history-and-milestones)
3. [Main problems and tasks](#a-main-problems-and-tasks)
4. [Methods and schools](#a-methods-and-schools)
5. [Datasets, resources and benchmarks](#a-datasets-resources-and-benchmarks)
6. [Landmark papers in this library](#a-landmark-papers-in-this-library)
7. [People](#a-people)
8. [Institutions and places](#a-institutions-and-places)
9. [Open questions](#a-open-questions)
10. [Neighbouring topics](#a-neighbouring-topics)
11. [Key works in this library](#a-key-works-in-this-library)

Information retrieval is the study of how to find, among a very large collection of texts, the few items that answer a person's need. A user types a short query; the system must decide, usually in a fraction of a second, which documents, passages or entities deserve to be shown first. The field grew out of library science and early computing, became a mass technology with web search engines, and has been reshaped twice in recent years: first by neural ranking models built on pretrained transformers, and then by large language models that both use retrieval and are used inside it.

The papers here trace that arc. They begin with two classic works of 1998, one describing the architecture of an early web search engine and one recasting ranking as a problem of probabilistic language modelling, pass through learning to rank and neural matching, and arrive at the busy present: dense retrievers, sparse learned expansions, re-rankers built on large models, and retrieval-augmented generation (RAG), in which a language model consults an external store of text before it writes an answer. Most of the work collected on this page dates from the last few years, when retrieval became one of the main tools for making language models more factual, more current and easier to check.

## What it is

### Core notions

A retrieval system has three parts. A **collection** (a corpus of web pages, Wikipedia passages, scientific abstracts, legal cases, proof libraries) is turned into an **index**, a data structure that makes search fast. A **query** expresses a need, often vaguely and in few words. A **ranking function** scores each candidate item against the query, and the system returns a ranked list. What counts as a good list is defined by **relevance**, a judgement made by people about whether an item satisfies the need.

The classic index is the **inverted index**, which maps each term to the list of documents that contain it, with counts and positions. Scoring functions over such indexes reward documents that contain query terms often (term frequency), favour rare and informative terms over common ones (inverse document frequency), and correct for document length. BM25, the probabilistic weighting scheme developed by Stephen Robertson and colleagues in the Okapi system during the 1990s, is the best-known example and remains, decades later, the baseline against which almost every new method in this library is measured.

Modern systems are usually **multi-stage**. A cheap first-stage retriever pulls hundreds or thousands of candidates from millions; a more expensive re-ranker reorders that shortlist; and, increasingly, a reader or generator turns the top results into an answer.

### How it differs from its neighbours

Retrieval is distinct from [question answering](https://md.research.elamur.ai/free/en/topics/question-answering/), which aims at an answer rather than a list of sources, but the two have fused: open-domain QA systems retrieve first and read second, and many of the most cited papers here were written with QA benchmarks in mind. It is distinct from [information extraction](https://md.research.elamur.ai/free/en/topics/information-extraction/), which pulls structured facts out of text, and from [knowledge graphs and ontologies](https://md.research.elamur.ai/free/en/topics/knowledge-graphs/), which store such facts, though entity retrieval and graph-based RAG connect all three. It shares with [text generation](https://md.research.elamur.ai/free/en/topics/text-generation/) the task of presenting retrieved content. Finally, it is now inseparable from [language models](https://md.research.elamur.ai/free/en/topics/language-models/): language models are used to encode, expand and re-rank, and retrieval is used to ground and refresh language models.

## History and milestones

### Before the web

The scientific study of retrieval began in the 1950s and 1960s with work on mechanised library catalogues. The Cranfield experiments of the 1960s set the evaluation pattern that the field still follows: a fixed collection, a set of queries, and human relevance judgements, against which systems are compared by precision and recall. Gerard Salton's SMART system introduced the vector space model, Karen Spärck Jones's 1972 work gave rise to inverse document frequency, and probabilistic models culminating in BM25 followed. From 1992 the Text REtrieval Conference (TREC), run by the US National Institute of Standards and Technology, provided large shared collections and pooled judgements, and became the main proving ground for new ideas.

### The web and statistical ranking (late 1990s to 2000s)

Two papers of 1998 in this library mark the move into the modern era. Brin and Page described the Google prototype, which ranked pages using the link graph of the web (PageRank) and the anchor text of incoming links. In the same year, Ponte and Croft proposed ranking documents by the probability that a document's own language model would generate the query, which founded the language modelling approach to retrieval and a long line of work on smoothing.

The 2000s brought machine learning into ranking: the RankNet paper from Microsoft Research (2005) trained a neural network on document pairs and became the root of a family of learning-to-rank methods used in web search.

### Neural representations (2010s)

During the 2010s, neural networks began to learn representations of queries and documents directly from text and click logs. Microsoft's work on deep semantic matching, represented here by a multi-task network that shared one layer between query classification and web ranking, showed that learned vectors could carry useful meaning across tasks. The 2016 release of MS MARCO, built from real Bing queries, supplied the training data neural rankers had lacked.

### Pretrained transformers (2019–2021)

The arrival of BERT in late 2018 changed the field within months. In early 2019 Nogueira and Cho fine-tuned BERT as a cross-encoder re-ranker and took first place on the MS MARCO passage leaderboard by a wide margin. In 2019 and 2020 a series of papers made learned dense retrieval practical: ORQA trained a retriever end to end from question–answer pairs, and DPR showed that a simple dual encoder trained with in-batch negatives could beat BM25 on open-domain QA. ANCE, RocketQA, TAS-Balanced, Condenser and coCondenser then refined the training of dense retrievers, while SPLADE and ColBERTv2 developed learned sparse and late-interaction alternatives. BEIR (2021) tested all these families outside their training domain and found that none dominated.

### Retrieval meets large language models (2022 onwards)

From 2022, large generative models shifted the questions. Researchers used them to write hypothetical documents for search (HyDE), to expand queries (query2doc), to rank candidates directly (RankGPT), and to replace retrieval altogether by generating background text (GenRead). In the other direction, retrieval became a standard way to improve language models: REPLUG added retrieval to models available only through an API, IRCoT and ITER-RETGEN interleaved search with reasoning, and Search-R1 trained models with reinforcement learning to call a search engine on their own. Surveys, benchmarks and evaluation tools for RAG multiplied from 2023.

Landmark papers on this topic: the 12 most cited, in time The 12 most cited papers on this topic span 1998–2024; the dot grows with the citations.

**The numbers**

| Paper | Citations |
| --- | --- |
| [1998 — The Anatomy of a Large-Scale Hypertextual Web Search Engine](https://md.research.elamur.ai/free/en/papers/brin-1998-anatomy-large-scale-hypertextual-web/) | 16,227 |
| [1998 — A Language Modeling Approach to Information Retrieval](https://md.research.elamur.ai/free/en/papers/ponte-1998-language-modeling-approach-information/) | 1,867 |
| [2005 — Learning to Rank using Gradient Descent](https://md.research.elamur.ai/free/en/papers/burges-2005-learning-rank-using-gradient/) | 2,896 |
| [2019 — Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) | 1,234 |
| [2019 — Passage Re-ranking with BERT](https://md.research.elamur.ai/free/en/papers/nogueira-2019-passage-re-ranking-bert/) | 667 |
| [2020 — Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) | 5,370 |
| [2020 — Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval](https://md.research.elamur.ai/free/en/papers/xiong-2020-approximate-nearest-neighbor-negative-contrastive/) | 681 |
| [2021 — Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) | 1,674 |
| [2021 — Unsupervised Dense Information Retrieval with Contrastive Learning](https://md.research.elamur.ai/free/en/papers/izacard-2021-unsupervised-dense-information-retrieval-contrastive/) | 824 |
| [2022 — Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions](https://md.research.elamur.ai/free/en/papers/trivedi-2022-interleaving-retrieval-chain-thought-reasoning/) | 711 |
| [2023 — Retrieval-Augmented Generation for Large Language Models: A Survey](https://md.research.elamur.ai/free/en/papers/gao-2023-retrieval-augmented-generation-large-language/) | 1,647 |
| [2024 — A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models](https://md.research.elamur.ai/free/en/papers/fan-2024-survey-rag-meeting-llms-towards/) | 1,006 |

Papers on this topic by year 109 dated papers from 1998 to 2026; the busiest year is 2023 with 24, and 88% of them appeared in 2018 or later.

**The numbers**

| Year | Papers |
| --- | --- |
| 2026 | 9 |
| 2025 | 6 |
| 2024 | 19 |
| 2023 | 24 |
| 2022 | 10 |
| 2021 | 14 |
| 2020 | 6 |
| 2019 | 6 |
| 2018 | 2 |
| 2017 | 2 |
| 2015 | 2 |
| 2013 | 1 |
| 2012 | 1 |
| 2005 | 3 |
| 2002 | 1 |
| 1999 | 1 |
| 1998 | 2 |

## Main problems and tasks

**Ad hoc document and passage ranking.** The central task: given a query, rank a fixed collection. Web search, TREC tracks and MS MARCO are its typical settings.

**Open-domain question answering.** A question is posed without the text that contains its answer. The system retrieves supporting passages, usually from Wikipedia, and a reader extracts or generates an answer. Much of the dense retrieval literature here was evaluated this way, and multi-hop variants require chaining several pieces of evidence that share few words with the question.

**Entity retrieval and linking.** Mentions in text must be mapped to entries in a knowledge base with millions of entities. The papers here replace hand-built alias tables with learned encoders that search over entity descriptions.

**Response selection.** Retrieval-based dialogue systems pick the reply that best fits a whole conversation.

**Cross-lingual and multilingual retrieval.** Queries in one language may need documents in another, either because the local collection is thin or because the topic is foreign. Multilingual sentence encoders, aligned word embeddings and benchmarks such as XOR QA and MIRACL address this, and several papers deal with morphologically rich languages such as Arabic and Hebrew, where stemming and lemmatisation matter for search.

**Retrieval-augmented generation.** A language model is given retrieved text as context and must use it faithfully. It must decide when and what to retrieve, resist irrelevant or false context, and cite sources.

**Evaluation.** Judging rankings and grounded answers fairly when labels are sparse and answers are long.

## Methods and schools

### Lexical and probabilistic retrieval

Lexical methods match query words to document words. They are fast, transparent, need no training data and handle rare names and codes precisely. Their weakness is **vocabulary mismatch**: a relevant document may use different words from the query. BM25 has proved remarkably durable: papers here find it competitive out of domain and use it to supply hard negatives for neural training.

### Learning to rank

Learning-to-rank methods combine many features of a query–document pair with a model trained on relevance labels or clicks. Pointwise methods predict a relevance score for each item; pairwise methods, such as RankNet, learn which of two items should rank higher; listwise methods optimise the ordering of whole lists.

### Neural re-ranking: cross-encoders

A cross-encoder reads the query and a candidate together through a transformer, so that every query token can attend to every document token. This gives the most accurate relevance estimates but cannot be precomputed, so it is used only on a shortlist. The BERT re-ranker of Nogueira and Cho made this the standard second stage; later work used T5 and then large decoder models.

### Dense retrieval: dual encoders

A dual encoder (bi-encoder) embeds queries and documents separately into the same vector space; relevance is a dot product or cosine, and documents can be encoded in advance and searched with approximate nearest-neighbour libraries such as FAISS or ScaNN. Training them is the central technical theme here:

- **Negatives.** Which wrong answers to contrast with the right one matters enormously. DPR used other questions' positives in the same batch plus one BM25 negative; ANCE argued that random in-batch negatives are too easy and refreshed an index during training to find hard ones; RocketQA shared negatives across devices and used a cross-encoder to remove false negatives, since many unlabelled passages are in fact relevant.
- **Distillation.** Re-rankers can teach retrievers. TAS-Balanced used two teacher models and topic-clustered batches so that a strong retriever could be trained on one consumer GPU; RocketQAv2 trained retriever and re-ranker jointly with a shared listwise objective.
- **Pretraining for retrieval.** Standard language models do not produce good single-vector text embeddings out of the box. ORQA used an inverse cloze task; Condenser changed the architecture during pretraining so that the summary vector gathers global meaning; coCondenser and Contriever added contrastive learning on spans of the same document, the latter with no labels at all.
- **Scale.** GTR showed that enlarging the encoder while keeping the embedding size fixed improves transfer to new domains; RepLLaMA and the Qwen3 and mGTE models built embedders from large decoder or long-context encoder backbones.

### Learned sparse retrieval

A middle path keeps the inverted index but learns the term weights. Doc2query appended predicted questions to each passage before indexing; SPLADE used BERT's masked-language-model head to assign weights to the whole vocabulary, so that documents and queries gain related terms, with a regulariser keeping the vectors sparse enough for fast search. These methods inherit the efficiency and exact matching of lexical systems while reducing vocabulary mismatch.

### Late interaction and multi-vector models

Between single-vector and full cross-attention sit models that keep several vectors per document. ColBERT keeps one vector per token and lets every query token pick its best-matching document token; ColBERTv2 compressed these vectors so the index shrinks to near single-vector size.

### Retrieval-augmented generation

RAG systems combine a retriever with a generator. Fusion-in-Decoder showed that a generative reader can combine evidence from a hundred passages, with accuracy still rising as more are added. Later work varied nearly every choice: iterative retrieval, query rewriting, routing by difficulty, filtering poor retrievals, compressing context, graph- or tree-shaped indexes, and training the retriever, the generator or both. Surveys from 2023 and 2024 group these into naive, advanced and modular pipelines.

### The main debates

- **Sparse versus dense.** Dense retrievers win in domain with plenty of training data, but zero-shot evaluations show BM25 holding its ground on many collections, and simple entity-centric questions expose dense models' weakness with rare names. Hybrids that combine both kinds of score are a common practical answer.
- **Retrieve or generate.** Some work argues that a large model's own parameters can stand in for an index; most of the evidence here suggests retrieval still adds accuracy, freshness and traceability, especially for rare entities and recent events.
- **Does the generator use what it retrieves?** Studies of long-form QA and of RAG robustness show that models may ignore retrieved text, be misled by it, or cite it unfaithfully, which makes grounding a research problem of its own.

## Datasets, resources and benchmarks

The field rests on shared test collections. The TREC ad hoc collections of the 1990s served the early language modelling work; the TREC Deep Learning tracks of 2019 and 2020 provide the judged web queries on which many neural rankers here report results. **MS MARCO**, built from real search queries, became the main training set for neural retrievers, though its licence restricts commercial use, which motivated zero-shot approaches. **Natural Questions**, **TriviaQA**, **WebQuestions** and **SQuAD** in open-domain form are the standard tests for QA retrievers, and **HotpotQA**, **2WikiMultihopQA** and **MuSiQue** for multi-hop retrieval. **TREC-CAR** (complex answer retrieval over Wikipedia) appears in early BERT ranking work.

Several benchmarks were introduced in papers in this library. [SearchQA](https://md.research.elamur.ai/free/en/papers/dunn-2017-searchqa-new-q-dataset-augmented/) paired Jeopardy! clues with real search snippets to test reading from noisy retrieved context. [BEIR](https://md.research.elamur.ai/free/en/papers/thakur-2021-beir-heterogenous-benchmark-zero/) gathered heterogeneous English collections for zero-shot testing and is now the default yardstick for generalisation. [MIRACL](https://md.research.elamur.ai/free/en/papers/zhang-2023-miracl-multilingual-retrieval-dataset-covering/) extends judged retrieval to typologically diverse languages, and [XOR QA](https://md.research.elamur.ai/free/en/papers/asai-2021-xor-qa-cross-lingual-open/) poses questions that must be answered from another language's Wikipedia. ColBERTv2 introduced LoTTE, built from StackExchange communities, to test long-tail topics. For retrieval-augmented generation, [RGB](https://md.research.elamur.ai/free/en/papers/chen-2023-benchmarking-large-language-models-retrieval/), [MIRAGE with the MedRAG toolkit](https://md.research.elamur.ai/free/en/papers/xiong-2024-benchmarking-retrieval-augmented-generation-medicine/), [MultiHop-RAG](https://md.research.elamur.ai/free/en/papers/tang-2024-multihop-rag-benchmarking-retrieval-augmented/) and [RAGTruth](https://md.research.elamur.ai/free/en/papers/niu-2024-ragtruth-hallucination-corpus-developing-trustworthy/) probe robustness, domain use, multi-step questions and hallucination, while [Ragas](https://md.research.elamur.ai/free/en/papers/es-2023-ragas-automated-evaluation-retrieval-augmented/) scores pipelines without reference answers. Smaller sets include [Gretino](https://md.research.elamur.ai/free/en/papers/toyin-2026-gretino-greek-latin-dataset-benchmark/) for Greek and Latin and [ArabicaQA](https://md.research.elamur.ai/free/en/papers/abdallah-2024-arabicaqa-comprehensive-dataset-arabic-question/) for Arabic.

## Landmark papers in this library

Citations of the papers on this topic, by when they were published Papers published in 2020–2024 gather 51% of the 56,624 citations of the papers on this topic.

**The numbers**

| Published | Papers and their citations |
| --- | --- |
| 2025–2029 | 15 papers, 657 citations |
| 2020–2024 | 73 papers, 28,984 citations |
| 2015–2019 | 12 papers, 5,313 citations |
| 2010–2014 | 2 papers, 214 citations |
| 2005–2009 | 3 papers, 3,324 citations |
| 2000–2004 | 1 papers, 3 citations |
| 1995–1999 | 3 papers, 18,129 citations |

### The anatomy of a web search engine (1998)

[The Anatomy of a Large-Scale Hypertextual Web Search Engine](https://md.research.elamur.ai/free/en/papers/brin-1998-anatomy-large-scale-hypertextual-web/), by [Sergey Brin](https://md.research.elamur.ai/free/en/authors/sergey-brin/) and [Lawrence Page](https://md.research.elamur.ai/free/en/authors/lawrence-page/), is by far the most cited work on this page. It presented the Stanford prototype of Google, described PageRank as the stationary probability of a random surfer moving along links, credited anchor text to the page a link points to, and set out the whole crawl–index–query pipeline at a time when such designs were rarely published.

### The language modelling approach (1998)

[A Language Modeling Approach to Information Retrieval](https://md.research.elamur.ai/free/en/papers/ponte-1998-language-modeling-approach-information/) by Ponte and Croft ranked each document by how likely its own estimated language model was to produce the query, with collection-wide statistics standing in for unseen words. On TREC topics it outperformed a tuned tf.idf system without hand-set weighting heuristics. The query-likelihood framework, and the smoothing methods developed after it, became a pillar of statistical IR.

### RankNet (2005)

[Learning to Rank using Gradient Descent](https://md.research.elamur.ai/free/en/papers/burges-2005-learning-rank-using-gradient/) introduced RankNet: a scorer trained on document pairs, with a logistic link between score differences and the probability that one item should precede another. Tested on commercial search data, it beat ordinal regression approaches. Its pairwise loss underlies many later rankers.

### ORQA (2019)

[Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) (ORQA) learned retrieval and reading jointly from question–answer strings alone, treating the chosen evidence block as a hidden variable. To start the retriever, it pretrained on an inverse cloze task, asking the model to find the passage a sentence was taken from. It set the stage for DPR.

### BERT re-ranking (2019)

[Passage Re-ranking with BERT](https://md.research.elamur.ai/free/en/papers/nogueira-2019-passage-re-ranking-bert/) fed query and passage as a sentence pair to BERT and used its classification output as a relevance probability. The jump over earlier neural rankers on MS MARCO and TREC-CAR was large, and the BERT cross-encoder became the standard re-ranking baseline. Its companion, [Document Expansion by Query Prediction](https://md.research.elamur.ai/free/en/papers/nogueira-2019-document-expansion-query-prediction/), moved neural computation to indexing time by appending generated questions to each passage.

### Dense Passage Retrieval (2020)

[Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) (DPR) is the most cited neural paper here. Two BERT encoders, trained only on existing question–passage pairs with in-batch and BM25 negatives, beat BM25 by a wide margin on top-20 retrieval across several QA datasets, without the extra pretraining ORQA needed. It became the default first stage for open-domain QA and early RAG systems.

### ANCE (2020)

[ANCE](https://md.research.elamur.ai/free/en/papers/xiong-2020-approximate-nearest-neighbor-negative-contrastive/) gave a convergence argument for why easy in-batch negatives slow training and replaced them with hard negatives retrieved from the whole corpus by the model being trained, using an index rebuilt asynchronously. It brought single-vector retrieval close to the accuracy of a BERT re-ranking cascade at a fraction of the cost.

### Fusion-in-Decoder (2021)

[Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) introduced Fusion-in-Decoder: a T5 encoder processes each retrieved passage separately with the question, and the decoder attends to all of them at once. Since each passage is encoded alone, the reader scales to a hundred of them, and accuracy kept rising with more.

### SPLADE (2021)

[SPLADE](https://md.research.elamur.ai/free/en/papers/formal-2021-splade-sparse-lexical-expansion/) learned sparse term weights over BERT's vocabulary, adding expansion terms and dropping uninformative ones, with a regulariser that keeps posting lists short. It approached dense models in accuracy while remaining servable from an ordinary inverted index.

### Contriever (2021)

[Contriever](https://md.research.elamur.ai/free/en/papers/izacard-2021-unsupervised-dense-information-retrieval-contrastive/) trained a dense retriever with no relevance labels, using randomly cropped spans of the same text as positive pairs and a large queue of negatives. It beat BM25 on recall for most BEIR datasets, was a stronger starting point for fine-tuning than plain BERT, and in a multilingual version retrieved across scripts.

### BEIR (2021)

[BEIR](https://md.research.elamur.ai/free/en/papers/thakur-2021-beir-heterogenous-benchmark-zero/) assembled diverse English retrieval tasks, from fact checking and argument retrieval to biomedical and financial search, and evaluated lexical, sparse, dense, late-interaction and re-ranking systems without in-domain training. It showed that in-domain strength predicts transfer poorly.

### ColBERTv2 (2022)

[ColBERTv2](https://md.research.elamur.ai/free/en/papers/santhanam-2022-colbertv2-effective-efficient-retrieval-lightweight/) combined token-level late interaction with distillation and hard-negative training, and compressed each token vector into a centroid plus a quantised residual. It led on most out-of-domain tests while keeping indexes close in size to single-vector systems.

### HyDE (2022)

[Precise Zero-Shot Dense Retrieval without Relevance Labels](https://md.research.elamur.ai/free/en/papers/gao-2022-precise-zero-shot-dense-retrieval/) (HyDE) asked an instruction-following model to write a plausible answer passage for the query and then searched for real documents near that passage's embedding. The invented text may be wrong in detail, but it sits in the right region of the space. With no training at all, it often matched retrievers fine-tuned on MS MARCO.

### Large models as re-rankers (2023)

[Is ChatGPT Good at Search? Investigating Large Language Models as Re-Ranking Agents](https://md.research.elamur.ai/free/en/papers/sun-2023-chatgpt-good-search-investigating-large/) prompted large chat models to output a permutation of numbered candidates, using a sliding window for long lists. GPT-4 outperformed strong supervised re-rankers; its orderings were distilled into a far smaller student.

### Surveys of RAG (2023)

[Retrieval-Augmented Generation for Large Language Models: A Survey](https://md.research.elamur.ai/free/en/papers/gao-2023-retrieval-augmented-generation-large-language/) organised the fast-growing RAG literature into naive, advanced and modular paradigms and catalogued tasks, datasets and metrics. Together with [A Survey on RAG Meeting LLMs](https://md.research.elamur.ai/free/en/papers/fan-2024-survey-rag-meeting-llms-towards/) and [Large Language Models for Information Retrieval: A Survey](https://md.research.elamur.ai/free/en/papers/zhu-2023-large-language-models-information-retrieval/), it is among the most used maps of the area.

### IRCoT (2022)

[Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions](https://md.research.elamur.ai/free/en/papers/trivedi-2022-interleaving-retrieval-chain-thought-reasoning/) (IRCoT) alternated between writing one reasoning step and using that step as the next search query. It raised recall and accuracy on multi-hop benchmarks, even with small models. Related iterative methods include [ITER-RETGEN](https://md.research.elamur.ai/free/en/papers/shao-2023-enhancing-retrieval-augmented-large-language/), [Demonstrate-Search-Predict](https://md.research.elamur.ai/free/en/papers/khattab-2022-demonstrate-search-predict-composing-retrieval/) and [Search-R1](https://md.research.elamur.ai/free/en/papers/jin-2025-search-r1-training-llms-reason/).

## People

Authors with the most papers on this topic Jimmy Lin leads with 5 papers; the 12 authors shown wrote 40 of the papers on this topic between them, counting each paper once per author.

**The numbers**

| Author | Papers here · their citations |
| --- | --- |
| [Jimmy Lin](https://md.research.elamur.ai/free/en/authors/jimmy-lin/) | 5 papers · 1,689 citations |
| [Wen-tau Yih](https://md.research.elamur.ai/free/en/authors/wen-tau-yih/) | 4 papers · 6,180 citations |
| [Luke Zettlemoyer](https://md.research.elamur.ai/free/en/authors/luke-zettlemoyer/) | 4 papers · 1,442 citations |
| [Edouard Grave](https://md.research.elamur.ai/free/en/authors/edouard-grave/) | 3 papers · 2,975 citations |
| [Gautier Izacard](https://md.research.elamur.ai/free/en/authors/gautier-izacard/) | 3 papers · 2,629 citations |
| [Ming-Wei Chang](https://md.research.elamur.ai/free/en/authors/ming-wei-chang/) | 3 papers · 1,677 citations |
| [Sebastian Riedel](https://md.research.elamur.ai/free/en/authors/sebastian-riedel/) | 3 papers · 1,573 citations |
| [Kyunghyun Cho](https://md.research.elamur.ai/free/en/authors/kyunghyun-cho/) | 3 papers · 1,527 citations |
| [Jamie Callan](https://md.research.elamur.ai/free/en/authors/jamie-callan/) | 3 papers · 1,134 citations |
| [Luyu Gao](https://md.research.elamur.ai/free/en/authors/luyu-gao/) | 3 papers · 1,134 citations |
| [Yi Luan](https://md.research.elamur.ai/free/en/authors/yi-luan/) | 3 papers · 943 citations |
| [Mike Lewis](https://md.research.elamur.ai/free/en/authors/mike-lewis/) | 3 papers · 824 citations |

[Jimmy Lin](https://md.research.elamur.ai/free/en/authors/jimmy-lin/) of the [University of Waterloo](https://md.research.elamur.ai/free/en/universities/university-of-waterloo/) is the most prolific author on this topic. His work here spans document expansion with predicted queries, efficient distillation for dense retrievers with balanced topic-aware sampling, zero-shot retrieval with hypothetical documents, and fine-tuning LLaMA-2 as both retriever and re-ranker.

[Wen-tau Yih](https://md.research.elamur.ai/free/en/authors/wen-tau-yih/), now at the [University of Washington](https://md.research.elamur.ai/free/en/universities/university-of-washington/), co-authored DPR and later work on black-box retrieval augmentation (REPLUG), passage re-ranking through zero-shot question generation, and instruction-aware retrieval.

[Luke Zettlemoyer](https://md.research.elamur.ai/free/en/authors/luke-zettlemoyer/) (University of Washington and [Meta](https://md.research.elamur.ai/free/en/companies/meta/)) appears on BLINK, the dense entity linker, on REPLUG, and on [RA-DIT](https://md.research.elamur.ai/free/en/papers/lin-2023-ra-dit-retrieval-augmented-dual/), which tunes both the language model and the retriever to work together. [Mike Lewis](https://md.research.elamur.ai/free/en/authors/mike-lewis/) and [Weijia Shi](https://md.research.elamur.ai/free/en/authors/weijia-shi/) share much of this line, Shi also co-authoring [RECOMP](https://md.research.elamur.ai/free/en/papers/xu-2023-recomp-improving-retrieval-augmented-lms/), which compresses retrieved documents before generation.

[Gautier Izacard](https://md.research.elamur.ai/free/en/authors/gautier-izacard/) and [Edouard Grave](https://md.research.elamur.ai/free/en/authors/edouard-grave/) wrote Fusion-in-Decoder and, with [Sebastian Riedel](https://md.research.elamur.ai/free/en/authors/sebastian-riedel/), [Armand Joulin](https://md.research.elamur.ai/free/en/authors/armand-joulin/) and [Piotr Bojanowski](https://md.research.elamur.ai/free/en/authors/piotr-bojanowski/), Contriever. Riedel, now at [Google DeepMind](https://md.research.elamur.ai/free/en/companies/google-deepmind/) and [University College London](https://md.research.elamur.ai/free/en/universities/university-college-london/), also co-authored BLINK.

[Kenton Lee](https://md.research.elamur.ai/free/en/authors/kenton-lee/), [Ming-Wei Chang](https://md.research.elamur.ai/free/en/authors/ming-wei-chang/) and [Kristina Toutanova](https://md.research.elamur.ai/free/en/authors/kristina-toutanova/) wrote ORQA; Chang and [Yi Luan](https://md.research.elamur.ai/free/en/authors/yi-luan/) continued with large dual encoders (GTR) and [Promptagator](https://md.research.elamur.ai/free/en/papers/dai-2022-promptagator-few-shot-dense-retrieval/), which builds task-specific retrievers from a handful of examples, and Luan and Toutanova analysed sparse, dense and multi-vector representations.

[Kyunghyun Cho](https://md.research.elamur.ai/free/en/authors/kyunghyun-cho/) ([New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/)) and [Rodrigo Nogueira](https://md.research.elamur.ai/free/en/authors/rodrigo-nogueira/) (now at UNICAMP in Brazil) produced the BERT re-ranker and Doc2query; Cho also co-created SearchQA.

[Jamie Callan](https://md.research.elamur.ai/free/en/authors/jamie-callan/) and [Luyu Gao](https://md.research.elamur.ai/free/en/authors/luyu-gao/) of [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) developed Condenser and coCondenser, which make dense retrievers trainable on modest hardware, and HyDE with Jimmy Lin.

[Danqi Chen](https://md.research.elamur.ai/free/en/authors/danqi-chen/), [Sewon Min](https://md.research.elamur.ai/free/en/authors/sewon-min/), [Patrick Lewis](https://md.research.elamur.ai/free/en/authors/patrick-lewis/) (now at [Cohere](https://md.research.elamur.ai/free/en/companies/cohere/)) and [Ledell Wu](https://md.research.elamur.ai/free/en/authors/ledell-wu/) are co-authors of DPR; Chen also co-wrote the study showing that [simple entity-centric questions challenge dense retrievers](https://md.research.elamur.ai/free/en/papers/sciavolino-2021-simple-entity-centric-questions-challenge/), and Wu was first author of BLINK.

[Akari Asai](https://md.research.elamur.ai/free/en/authors/akari-asai/) and [Hannaneh Hajishirzi](https://md.research.elamur.ai/free/en/authors/hannaneh-hajishirzi/) (University of Washington and the [Allen Institute for Artificial Intelligence](https://md.research.elamur.ai/free/en/labs/allen-institute-for-artificial-intelligence/)) worked on graph-based retrieval of multi-hop reasoning paths over Wikipedia, cross-lingual open-retrieval QA, and retrievers that follow task instructions.

[Mohit Iyyer](https://md.research.elamur.ai/free/en/authors/mohit-iyyer/), with [Kalpesh Krishna](https://md.research.elamur.ai/free/en/authors/kalpesh-krishna/), questioned whether long-form QA systems actually use what they retrieve, and proposed retrieval over stored generations as a defence against paraphrased machine text.

[Nandan Thakur](https://md.research.elamur.ai/free/en/authors/nandan-thakur/) (University of Waterloo) led BEIR with [Nils Reimers](https://md.research.elamur.ai/free/en/authors/nils-reimers/) and [Iryna Gurevych](https://md.research.elamur.ai/free/en/authors/iryna-gurevych/), and worked on domain adaptation by generative pseudo-labelling and on the multilingual MIRACL collection.

[Haifeng Wang](https://md.research.elamur.ai/free/en/authors/haifeng-wang/), [Hua Wu](https://md.research.elamur.ai/free/en/authors/hua-wu/), [Jing Liu](https://md.research.elamur.ai/free/en/authors/jing-liu/) and [Ruiyang Ren](https://md.research.elamur.ai/free/en/authors/ruiyang-ren/) of the Baidu group, with [Wayne Xin Zhao](https://md.research.elamur.ai/free/en/authors/wayne-xin-zhao/) and [Ji-Rong Wen](https://md.research.elamur.ai/free/en/authors/ji-rong-wen/) of Renmin University, built RocketQA and RocketQAv2.

Earlier work is represented by [Jianfeng Gao](https://md.research.elamur.ai/free/en/authors/jianfeng-gao/), [Xiaodong He](https://md.research.elamur.ai/free/en/authors/xiaodong-he/) and [Li Deng](https://md.research.elamur.ai/free/en/authors/li-deng/) (Microsoft's deep semantic matching). [Omar Khattab](https://md.research.elamur.ai/free/en/authors/omar-khattab/), [Keshav Santhanam](https://md.research.elamur.ai/free/en/authors/keshav-santhanam/), [Christopher Potts](https://md.research.elamur.ai/free/en/authors/christopher-potts/) and [Matei Zaharia](https://md.research.elamur.ai/free/en/authors/matei-zaharia/) at Stanford developed ColBERTv2 and the Demonstrate-Search-Predict framework.

## Institutions and places

Where the authors on this topic work now Carnegie Mellon University is the current home of authors of 16 papers on this topic; counted are the authors with a page here, by their latest institution.

**The numbers**

| Institution | Papers |
| --- | --- |
| [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) | 16 papers |
| [Google (United States)](https://md.research.elamur.ai/free/en/companies/google/) | 16 papers |
| [Google DeepMind (United Kingdom)](https://md.research.elamur.ai/free/en/companies/google-deepmind/) | 9 papers |
| [Microsoft (United States)](https://md.research.elamur.ai/free/en/companies/microsoft/) | 9 papers |
| [University of Washington](https://md.research.elamur.ai/free/en/universities/university-of-washington/) | 9 papers |
| [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/) | 8 papers |
| [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/) | 8 papers |
| [University of Waterloo](https://md.research.elamur.ai/free/en/universities/university-of-waterloo/) | 8 papers |
| [New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/) | 7 papers |
| [Tsinghua University](https://md.research.elamur.ai/free/en/universities/tsinghua-university/) | 7 papers |

The work is concentrated in a few universities and industrial laboratories, led by the United States, then China. [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/), long a centre of IR research, is the current home of the most authors here, and [Google](https://md.research.elamur.ai/free/en/companies/google/) is equally prominent. [Microsoft](https://md.research.elamur.ai/free/en/companies/microsoft/) contributed RankNet, deep semantic matching, ANCE and query2doc; [Meta](https://md.research.elamur.ai/free/en/companies/meta/) DPR, BLINK, Contriever and REPLUG.

Among universities, the [University of Washington](https://md.research.elamur.ai/free/en/universities/university-of-washington/), [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/), [New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/), the [Courant Institute](https://md.research.elamur.ai/free/en/universities/courant-institute-of-mathematical-sciences/) and the [University of Massachusetts Amherst](https://md.research.elamur.ai/free/en/universities/university-of-massachusetts-amherst/), home of the Center for Intelligent Information Retrieval where the language modelling approach originated, stand out in the United States. The [University of Waterloo](https://md.research.elamur.ai/free/en/universities/university-of-waterloo/) is the main Canadian centre, joined by [Cohere](https://md.research.elamur.ai/free/en/companies/cohere/). In China, the [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/), the [University of Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/universities/university-of-chinese-academy-of-sciences/), [Tsinghua University](https://md.research.elamur.ai/free/en/universities/tsinghua-university/), [Peking University](https://md.research.elamur.ai/free/en/universities/peking-university/) and [Zhejiang University](https://md.research.elamur.ai/free/en/universities/zhejiang-university/) are prominent. In Europe, [University College London](https://md.research.elamur.ai/free/en/universities/university-college-london/) and the French [Centre National de la Recherche Scientifique](https://md.research.elamur.ai/free/en/labs/centre-national-de-la-recherche-scientifique/) appear, the latter through the Paris-based researchers behind Contriever and Fusion-in-Decoder.

Countries of the authors on this topic United States comes first: authors working there now wrote 62 papers on this topic.

**The numbers**

| Country | Papers |
| --- | --- |
| United States | 62 papers |
| China | 42 papers |
| United Kingdom | 21 papers |
| Canada | 11 papers |
| France | 8 papers |
| Germany | 7 papers |
| Hong Kong | 7 papers |
| Israel | 6 papers |
| United Arab Emirates | 5 papers |
| Bulgaria | 4 papers |

The papers appear mostly at the main computational linguistics venues, EMNLP, NAACL, ACL and its Findings, and the Transactions of the ACL, alongside the SIGIR conference for core IR work, the NeurIPS datasets track for BEIR, and many influential preprints.

Where the papers on this topic were published Proceedings of the Fifteenth Language Resources and Evaluation Conference carries the most, 4 of 109.

**The numbers**

| Venue | Papers |
| --- | --- |
| Proceedings of the Fifteenth Language Resources and Evaluation Conference | 4 papers |
| Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing | 3 papers |
| Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies | 3 papers |
| Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing | 3 papers |
| Transactions of the Association for Computational Linguistics | 3 papers |
| Computational Linguistics | 2 papers |
| Findings of the Association for Computational Linguistics: ACL 2024 | 2 papers |
| Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) | 2 papers |
| Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies | 2 papers |
| Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing | 2 papers |

## Open questions

**Generalisation.** Neural retrievers trained on one large collection still transfer unevenly. Zero-shot methods (Contriever, HyDE), synthetic training data ([GPL](https://md.research.elamur.ai/free/en/papers/wang-2022-gpl-generative-pseudo-labeling-unsupervised/), Promptagator, [question generation for re-ranking](https://md.research.elamur.ai/free/en/papers/sachan-2022-improving-passage-retrieval-zero-shot/)) and scale (GTR) all help, but no method is reliably best across domains.

**Rare entities and exact matching.** Dense models struggle with names and specific facts that lexical methods find easily. Hybrid retrieval is the common workaround; a single model that handles both well remains an open target.

**Faithful use of retrieved text.** Generators may ignore, misuse or miscite retrieved passages. An audit of commercial generative search engines ([Evaluating Verifiability in Generative Search Engines](https://md.research.elamur.ai/free/en/papers/liu-2023-evaluating-verifiability-generative-search-engines/)) found many cited claims unsupported. Methods such as [CRAG](https://md.research.elamur.ai/free/en/papers/yan-2024-corrective-retrieval-augmented-generation/), [Chain-of-Note](https://md.research.elamur.ai/free/en/papers/yu-2024-chain-note-enhancing-robustness-retrieval/) and [robustness training against irrelevant context](https://md.research.elamur.ai/free/en/papers/yoran-2023-making-retrieval-augmented-language-models/) are partial answers.

**When and how much to retrieve.** Adaptive strategies route queries by difficulty, but classifiers for this remain weak. As for long-context models, [Retrieval meets Long Context Large Language Models](https://md.research.elamur.ai/free/en/papers/xu-2023-retrieval-meets-long-context-large/) finds that the two are complementary.

**Evaluation.** Sparse labels penalise systems that find unjudged relevant documents, ROUGE fails for long answers, and model-based judges raise their own validity questions.

**Freshness and time.** Knowledge changes; [FreshLLMs](https://md.research.elamur.ai/free/en/papers/vu-2023-freshllms-refreshing-large-language-models/) shows that search engine results can keep model answers current, but how to maintain and evaluate time-sensitive systems is unsettled.

**Languages beyond English.** Most benchmarks and training data are English. Morphologically rich and low-resource languages need suitable tokenisation, stemming and training data, as papers on [Arabic lemmatisation for retrieval](https://md.research.elamur.ai/free/en/papers/langstream-2012-el-shishtawy-accurate-arabic-root-based-lemmatizer/), [Arabic stemming](https://md.research.elamur.ai/free/en/papers/bessou-2019-accuracy-enhanced-stemming-algorithm-arabic/), [Hebrew search](https://md.research.elamur.ai/free/en/papers/aclbulk-2002-ornan-morphological-syntactic-semantic-search-engine/) and [dense retrieval for Amharic](https://md.research.elamur.ai/free/en/papers/yeshambel-2025-dense-retrieval-low-resource-languages/) illustrate.

**Cost.** Large re-rankers and agentic loops with many search calls are expensive. Distillation, compression and efficient inference such as [DuoAttention](https://md.research.elamur.ai/free/en/papers/xiao-2024-duoattention-efficient-long-context-llm/) aim to cut that cost.

## Neighbouring topics

Neighbouring topics: papers shared Language models & pretraining shares the most: 30 of the 109 papers here (28%) stand on both.

**The numbers**

| Topic | Papers shared |
| --- | --- |
| [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) | 30 shared · 28% |
| [Question answering](https://md.research.elamur.ai/free/en/topics/question-answering/) | 17 shared · 16% |
| [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) | 9 shared · 8% |
| [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) | 6 shared · 6% |
| [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) | 4 shared · 4% |
| [Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) | 4 shared · 4% |
| [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) | 3 shared · 3% |
| [Morphology](https://md.research.elamur.ai/free/en/topics/morphology/) | 3 shared · 3% |
| [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) | 3 shared · 3% |
| [Efficient models & compression](https://md.research.elamur.ai/free/en/topics/efficient-models/) | 2 shared · 2% |

The closest neighbour is [language models and pretraining](https://md.research.elamur.ai/free/en/topics/language-models/): pretrained encoders power neural retrievers and re-rankers, and retrieval augments large language models. [Question answering](https://md.research.elamur.ai/free/en/topics/question-answering/) shares the open-domain QA literature, from ORQA and DPR to multi-hop retrieval. [Evaluation and benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) covers BEIR, RGB and RAG evaluation; [reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) covers iterative retrieval with chain-of-thought.

Multilingual work links to [multilingual and cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) and [low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/), and work on Arabic and Hebrew search connects to [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/), [Hebrew and Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/) and [morphology](https://md.research.elamur.ai/free/en/topics/morphology/). Benchmarks for classical texts touch [Ancient Greek](https://md.research.elamur.ai/free/en/topics/ancient-greek/) and [Latin](https://md.research.elamur.ai/free/en/topics/latin/). Retrieval-based reply selection and grounded conversation tie the topic to [dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/); retrieval used for adapting translation models and for idiom translation, to [machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) and [idioms and multiword expressions](https://md.research.elamur.ai/free/en/topics/multiword-expressions/). Entity retrieval links to [named entity recognition](https://md.research.elamur.ai/free/en/topics/named-entity-recognition/) and [coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/), which an early paper here used for question answering. Further neighbours include [summarization](https://md.research.elamur.ai/free/en/topics/summarization/), [efficient models](https://md.research.elamur.ai/free/en/topics/efficient-models/) and [legal AI](https://md.research.elamur.ai/free/en/topics/legal-ai/).

The shelves this topic stands on Natural language processing: 46 of the 109 papers here.

**The numbers**

| Shelf | Papers |
| --- | --- |
| [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) | 46 papers |
| [Other](https://md.research.elamur.ai/free/en/shelves/other/) | 28 papers |
| [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) | 25 papers |
| [Computing & GPUs](https://md.research.elamur.ai/free/en/shelves/computing/) | 13 papers |
| [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/ai/) | 9 papers |
| [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/) | 4 papers |
| [Law](https://md.research.elamur.ai/free/en/shelves/law/) | 1 paper |
| [Life sciences & medicine](https://md.research.elamur.ai/free/en/shelves/life-sciences/) | 1 paper |
| [Mathematics & statistics](https://md.research.elamur.ai/free/en/shelves/mathematics/) | 1 paper |

Within the site's shelves, the topic stands mainly on [natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/), with substantial parts on [linguistics and languages](https://md.research.elamur.ai/free/en/shelves/linguistics/), [computing and GPUs](https://md.research.elamur.ai/free/en/shelves/computing/), [AI and machine learning](https://md.research.elamur.ai/free/en/shelves/ai/) and [machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/), and single papers on [law](https://md.research.elamur.ai/free/en/shelves/law/), [life sciences and medicine](https://md.research.elamur.ai/free/en/shelves/life-sciences/) and [mathematics and statistics](https://md.research.elamur.ai/free/en/shelves/mathematics/).

## Key works in this library

- [The Anatomy of a Large-Scale Hypertextual Web Search Engine](https://md.research.elamur.ai/free/en/papers/brin-1998-anatomy-large-scale-hypertextual-web/) — PageRank, anchor text and the architecture of early Google.
- [A Language Modeling Approach to Information Retrieval](https://md.research.elamur.ai/free/en/papers/ponte-1998-language-modeling-approach-information/) — ranking by query likelihood under a document language model.
- [Learning to Rank using Gradient Descent](https://md.research.elamur.ai/free/en/papers/burges-2005-learning-rank-using-gradient/) — RankNet and the pairwise probabilistic ranking loss.
- [Passage Re-ranking with BERT](https://md.research.elamur.ai/free/en/papers/nogueira-2019-passage-re-ranking-bert/) — the BERT cross-encoder as a re-ranker.
- [Document Expansion by Query Prediction](https://md.research.elamur.ai/free/en/papers/nogueira-2019-document-expansion-query-prediction/) — enriching passages with generated queries before indexing.
- [Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) — ORQA and the inverse cloze task.
- [Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) — DPR, the dual-encoder baseline.
- [Scalable Zero-shot Entity Linking with Dense Entity Retrieval](https://md.research.elamur.ai/free/en/papers/wu-2020-scalable-zero-shot-entity-linking/) — BLINK: bi-encoder retrieval plus cross-encoder re-ranking of entities.
- [Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval](https://md.research.elamur.ai/free/en/papers/xiong-2020-approximate-nearest-neighbor-negative-contrastive/) — ANCE and corpus-wide hard negatives.
- [RocketQA](https://md.research.elamur.ai/free/en/papers/qu-2021-rocketqa-optimized-training-approach-dense/) — cross-batch and denoised negatives for dense retrieval.
- [Efficiently Teaching an Effective Dense Retriever with Balanced Topic Aware Sampling](https://md.research.elamur.ai/free/en/papers/hofstatter-2021-efficiently-teaching-effective-dense-retriever/) — strong dense retrieval on a single GPU.
- [Condenser](https://md.research.elamur.ai/free/en/papers/gao-2021-condenser-pre-training-architecture-dense/) and [coCondenser](https://md.research.elamur.ai/free/en/papers/gao-2022-unsupervised-corpus-aware-language-model/) — pretraining that readies the summary vector for retrieval.
- [SPLADE](https://md.research.elamur.ai/free/en/papers/formal-2021-splade-sparse-lexical-expansion/) — learned sparse expansion for first-stage ranking.
- [Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) — Fusion-in-Decoder.
- [Unsupervised Dense Information Retrieval with Contrastive Learning](https://md.research.elamur.ai/free/en/papers/izacard-2021-unsupervised-dense-information-retrieval-contrastive/) — Contriever, label-free dense retrieval.
- [BEIR](https://md.research.elamur.ai/free/en/papers/thakur-2021-beir-heterogenous-benchmark-zero/) — the zero-shot retrieval benchmark.
- [ColBERTv2](https://md.research.elamur.ai/free/en/papers/santhanam-2022-colbertv2-effective-efficient-retrieval-lightweight/) — compressed late interaction.
- [Precise Zero-Shot Dense Retrieval without Relevance Labels](https://md.research.elamur.ai/free/en/papers/gao-2022-precise-zero-shot-dense-retrieval/) — HyDE, searching with a hypothetical document.
- [Is ChatGPT Good at Search?](https://md.research.elamur.ai/free/en/papers/sun-2023-chatgpt-good-search-investigating-large/) — permutation-based re-ranking by large models.
- [Fine-Tuning LLaMA for Multi-Stage Text Retrieval](https://md.research.elamur.ai/free/en/papers/ma-2023-fine-tuning-llama-multi-stage/) — RepLLaMA and RankLLaMA.
- [REPLUG](https://md.research.elamur.ai/free/en/papers/shi-2023-replug-retrieval-augmented-black-box/) — retrieval for language models reachable only through an API.
- [Search-R1](https://md.research.elamur.ai/free/en/papers/jin-2025-search-r1-training-llms-reason/) — reinforcement learning to make models search on their own.

## All 109 papers on this topic

**1** [2](https://md.research.elamur.ai/free/en/topics/information-retrieval/page/2/) … [5](https://md.research.elamur.ai/free/en/topics/information-retrieval/page/5/) [next →](https://md.research.elamur.ai/free/en/topics/information-retrieval/page/2/)

- [The Anatomy of a Large-Scale Hypertextual Web Search Engine](https://md.research.elamur.ai/free/en/papers/brin-1998-anatomy-large-scale-hypertextual-web/) — Sergey Brin, Lawrence Page · 1998 · 16 227 citations
- [Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) — Vladimir Karpukhin, Barlas Oguz, Sewon Min, et al. · 2020 · 5 370 citations
- [Learning to Rank using Gradient Descent](https://md.research.elamur.ai/free/en/papers/burges-2005-learning-rank-using-gradient/) — Chris Burges, Tal Shaked, Erin Renshaw, et al. · 2005 · 2 896 citations
- [A Language Modeling Approach to Information Retrieval](https://md.research.elamur.ai/free/en/papers/ponte-1998-language-modeling-approach-information/) — Jay M. Ponte, W. Bruce Croft · 1998 · 1 867 citations
- [Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) — Gautier Izacard, Edouard Grave · 2021 · 1 674 citations
- [Retrieval-Augmented Generation for Large Language Models: A Survey](https://md.research.elamur.ai/free/en/papers/gao-2023-retrieval-augmented-generation-large-language/) — Yunfan Gao, Yun Xiong, Xinyu Gao, et al. · 2023 · 1 647 citations
- [Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) — Kenton Lee, Ming-Wei Chang, Kristina Toutanova · 2019 · 1 234 citations
- [A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models](https://md.research.elamur.ai/free/en/papers/fan-2024-survey-rag-meeting-llms-towards/) — Wenqi Fan; Yujuan Ding; Liangbo Ning; Shijie Wang; Hengyun Li; Dawei Yin; Tat-Seng Chua;, Qing Li · 2024 · 1 006 citations
- [Unsupervised Dense Information Retrieval with Contrastive Learning](https://md.research.elamur.ai/free/en/papers/izacard-2021-unsupervised-dense-information-retrieval-contrastive/) — Gautier Izacard, Mathilde Caron, Lucas Hosseini, et al. · 2021 · 824 citations
- [Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions](https://md.research.elamur.ai/free/en/papers/trivedi-2022-interleaving-retrieval-chain-thought-reasoning/) — Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, et al. · 2022 · 711 citations
- [Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval](https://md.research.elamur.ai/free/en/papers/xiong-2020-approximate-nearest-neighbor-negative-contrastive/) — Lee Xiong, Chenyan Xiong, Ye Li, et al. · 2020 · 681 citations
- [Passage Re-ranking with BERT](https://md.research.elamur.ai/free/en/papers/nogueira-2019-passage-re-ranking-bert/) — Rodrigo Nogueira, Kyunghyun Cho · 2019 · 667 citations
- [Benchmarking Large Language Models in Retrieval-Augmented Generation](https://md.research.elamur.ai/free/en/papers/chen-2023-benchmarking-large-language-models-retrieval/) — Jiawei Chen, Hongyu Lin, Xianpei Han, et al. · 2023 · 624 citations
- [Sequential Matching Network: A New Architecture for Multi-turn Response Selection in Retrieval-Based Chatbots](https://md.research.elamur.ai/free/en/papers/wu-2017-sequential-matching-network-new-architecture/) — Yu Wu, Wei Wu, Chen Xing, et al. · 2017 · 624 citations
- [Scalable Zero-shot Entity Linking with Dense Entity Retrieval](https://md.research.elamur.ai/free/en/papers/wu-2020-scalable-zero-shot-entity-linking/) — Ledell Wu, Fabio Petroni, Martin Josifoski, et al. · 2020 · 618 citations
- [RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/qu-2021-rocketqa-optimized-training-approach-dense/) — Yingqi Qu, Yuchen Ding, Jing Liu, et al. · 2021 · 601 citations
- [Multilingual Universal Sentence Encoder for Semantic Retrieval](https://md.research.elamur.ai/free/en/papers/yang-2020-multilingual-universal-sentence-encoder-semantic/) — Yinfei Yang, Daniel Cer, Amin Ahmad, et al. · 2020 · 571 citations
- [Precise Zero-Shot Dense Retrieval without Relevance Labels](https://md.research.elamur.ai/free/en/papers/gao-2022-precise-zero-shot-dense-retrieval/) — Luyu Gao, Xueguang Ma, Jimmy Lin, et al. · 2022 · 545 citations
- [SearchQA: A New Q&A Dataset Augmented with Context from a Search Engine](https://md.research.elamur.ai/free/en/papers/dunn-2017-searchqa-new-q-dataset-augmented/) — Matthew Dunn, Levent Sagun, Mike Higgins, et al. · 2017 · 543 citations
- [SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking](https://md.research.elamur.ai/free/en/papers/formal-2021-splade-sparse-lexical-expansion/) — Thibault Formal, Benjamin Piwowarski, Stéphane Clinchant · 2021 · 510 citations
- [Sparse, Dense, and Attentional Representations for Text Retrieval](https://md.research.elamur.ai/free/en/papers/luan-2021-sparse-dense-attentional-representations-text/) — Yi Luan, Jacob Eisenstein, Kristina Toutanova, et al. · 2021 · 500 citations
- [Is ChatGPT Good at Search? Investigating Large Language Models as Re-Ranking Agents](https://md.research.elamur.ai/free/en/papers/sun-2023-chatgpt-good-search-investigating-large/) — Weiwei Sun, Lingyong Yan, Xinyu Ma, et al. · 2023 · 493 citations
- [REPLUG: Retrieval-Augmented Black-Box Language Models](https://md.research.elamur.ai/free/en/papers/shi-2023-replug-retrieval-augmented-black-box/) — Weijia Shi, Sewon Min, Michihiro Yasunaga, et al. · 2023 · 482 citations
- [ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction](https://md.research.elamur.ai/free/en/papers/santhanam-2022-colbertv2-effective-efficient-retrieval-lightweight/) — Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, et al. · 2022 · 477 citations
- [Loss in Translation: Learning Bilingual Word Mapping with a Retrieval Criterion](https://md.research.elamur.ai/free/en/papers/joulin-2018-loss-translation-learning-bilingual-word/) — Armand Joulin, Piotr Bojanowski, Tomas Mikolov, et al. · 2018 · 477 citations

**1** [2](https://md.research.elamur.ai/free/en/topics/information-retrieval/page/2/) … [5](https://md.research.elamur.ai/free/en/topics/information-retrieval/page/5/) [next →](https://md.research.elamur.ai/free/en/topics/information-retrieval/page/2/)

- [← previous topic Information extraction](https://md.research.elamur.ai/free/en/topics/information-extraction/)
- [next topic → Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/)
- [↑ back to topics](https://md.research.elamur.ai/free/en/topics/)