abstract = {This technical report presents the training methodology and evaluation results of the open-source multilingual E5 text embedding models, released in mid-2023. Three embedding models of different sizes (small / base / large) are provided, offering a balance between the inference efficiency and embedding quality. The training procedure adheres to the English E5 model recipe, involving contrastive pre-training on 1 billion multilingual text pairs, followed by fine-tuning on a combination of labeled datasets. Additionally, we introduce a new instruction-tuned embedding model, whose performance is on par with state-of-the-art, English-only models of similar sizes. Information regarding the model release can be found at https://github.com/microsoft/unilm/tree/master/e5 .},
language = {en},
urldate = {2026-06-03},
author = {Wang, Liang and Yang, Nan and Huang, Xiaolong and Yang, Linjun and Majumder, Rangan and Wei, Furu},
month = feb,
year = {2024},
}
@inproceedings{chen_axiomatic_2024,
address = {Washington DC USA},
title = {Axiomatic {Causal} {Interventions} for {Reverse} {Engineering} {Relevance} {Computation} in {Neural} {Retrieval} {Models}},
booktitle = {Proceedings of the 47th {International} {ACM} {SIGIR} {Conference} on {Research} and {Development} in {Information} {Retrieval}},
publisher = {ACM},
author = {Chen, Catherine and Merullo, Jack and Eickhoff, Carsten},
month = jul,
year = {2024},
pages = {1401--1410},
}
@misc{wang_text_2022,
title = {Text {Embeddings} by {Weakly}-{Supervised} {Contrastive} {Pre}-training},
url = {https://arxiv.org/abs/2212.03533v2},
abstract = {This paper presents E5, a family of state-of-the-art text embeddings that transfer well to a wide range of tasks. The model is trained in a contrastive manner with weak supervision signals from our curated large-scale text pair dataset (called CCPairs). E5 can be readily used as a general-purpose embedding model for any tasks requiring a single-vector representation of texts such as retrieval, clustering, and classification, achieving strong performance in both zero-shot and fine-tuned settings. We conduct extensive evaluations on 56 datasets from the BEIR and MTEB benchmarks. For zero-shot settings, E5 is the first model that outperforms the strong BM25 baseline on the BEIR retrieval benchmark without using any labeled data. When fine-tuned, E5 obtains the best results on the MTEB benchmark, beating existing embedding models with 40x more parameters.},
language = {en},
urldate = {2026-05-20},
author = {Wang, Liang and Yang, Nan and Huang, Xiaolong and Jiao, Binxing and Yang, Linjun and Jiang, Daxin and Majumder, Rangan and Wei, Furu},
month = dec,
year = {2022},
}
@inproceedings{parry_mechir_2025,
series = {Lecture {Notes} in {Computer} {Science}},
title = {{MechIR}: {A} {Mechanistic} {Interpretability} {Framework} for {Information} {Retrieval}},
author = {Parry, Andrew and Chen, Catherine and Eickhoff, Carsten and MacAvaney, Sean},
year = {2025},
pages = {89--95},
}
@misc{izacard_unsupervised_2022,
title = {Unsupervised {Dense} {Information} {Retrieval} with {Contrastive} {Learning}},
url = {http://arxiv.org/abs/2112.09118},
doi = {10.48550/arXiv.2112.09118},
abstract = {Recently, information retrieval has seen the emergence of dense retrievers, using neural networks, as an alternative to classical sparse methods based on term-frequency. These models have obtained state-of-the-art results on datasets and tasks where large training sets are available. However, they do not transfer well to new applications with no training data, and are outperformed by unsupervised term-frequency methods such as BM25. In this work, we explore the limits of contrastive learning as a way to train unsupervised dense retrievers and show that it leads to strong performance in various retrieval settings. On the BEIR benchmark our unsupervised model outperforms BM25 on 11 out of 15 datasets for the Recall@100. When used as pre-training before fine-tuning, either on a few thousands in-domain examples or on the large MS{\textasciitilde}MARCO dataset, our contrastive model leads to improvements on the BEIR benchmark. Finally, we evaluate our approach for multi-lingual retrieval, where training data is even scarcer than for English, and show that our approach leads to strong unsupervised performance. Our model also exhibits strong cross-lingual transfer when fine-tuned on supervised English data only and evaluated on low resources language such as Swahili. We show that our unsupervised models can perform cross-lingual retrieval between different scripts, such as retrieving English documents from Arabic queries, which would not be possible with term matching methods.},
urldate = {2026-03-16},
publisher = {arXiv},
author = {Izacard, Gautier and Caron, Mathilde and Hosseini, Lucas and Riedel, Sebastian and Bojanowski, Piotr and Joulin, Armand and Grave, Edouard},
month = aug,
year = {2022},
note = {arXiv:2112.09118},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language, Computer Science - Information Retrieval},
}
@article{izacard_atlas_2023,
title = {Atlas: few-shot learning with retrieval augmented language models},
volume = {24},
issn = {1532-4435},
abstract = {Large language models have shown impressive few-shot results on a wide range of tasks. However, when knowledge is key for such results, as is the case for tasks such as question answering and fact checking, massive parameter counts to store knowledge seem to be needed. Retrieval-augmented models are known to excel at knowledge intensive tasks without the need for as many parameters, but it is unclear whether they work in few-shot settings. In this work we present Atlas, a carefully designed and pre-trained retrieval-augmented language model able to learn knowledge intensive tasks with very few training examples. We perform evaluations on a wide range of tasks, including MMLU, KILT and Natural Questions, and study the impact of the content of the document index, showing that it can easily be updated. Notably, Atlas reaches over 42\% accuracy on Natural Questions using only 64 examples, outperforming a 540B parameter model by 3\% despite having 50x fewer parameters.},
number = {1},
journal = {J. Mach. Learn. Res.},
author = {Izacard, Gautier and Lewis, Patrick and Lomeli, Maria and Hosseini, Lucas and Petroni, Fabio and Schick, Timo and Dwivedi-Yu, Jane and Joulin, Armand and Riedel, Sebastian and Grave, Edouard},
month = jan,
year = {2023},
keywords = {information retrieval, language models, retrieval augmented language models},
}
@misc{xu_search---chain_2024,
title = {Search-in-the-{Chain}: {Interactively} {Enhancing} {Large} {Language} {Models} with {Search} for {Knowledge}-intensive {Tasks}},
shorttitle = {Search-in-the-{Chain}},
url = {http://arxiv.org/abs/2304.14732},
doi = {10.48550/arXiv.2304.14732},
abstract = {Making the content generated by Large Language Model (LLM), accurate, credible and traceable is crucial, especially in complex knowledge-intensive tasks that require multi-step reasoning and each step needs knowledge to solve. Retrieval-augmented generation is good potential to solve this problem. However, where and how to introduce Information Retrieval (IR) to LLM is a big challenge. Previous work has the problems that wrong knowledge retrieved by IR misleads the LLM and interaction between IR and LLM breaks the reasoning chain of LLM. This paper proposes a novel framework named {\textbackslash}textbf\{Search-in-the-Chain\} (SearChain) for the interaction between LLM and IR to solve the challenges. First, LLM generates the reasoning chain named Chain-of-Query (CoQ) where each node consists of an IR-oriented query-answer pair. Second, IR verifies the answer of each node of CoQ. It corrects the answer that is not consistent with the retrieved information when IR gives high confidence, which improves the credibility. Third, LLM can indicate its missing knowledge in CoQ and rely on IR to provide this knowledge to LLM. These operations improve the accuracy in terms of reasoning and knowledge. Finally, SearChain generates the reasoning process and marks references to supporting documents for each reasoning step, which improves traceability. Interaction with IR in SearChain forms a novel reasoning path based on a tree, which enables LLM to dynamically modify the direction of reasoning. Experiments show that SearChain outperforms state-of-the-art baselines on complex knowledge-intensive tasks including multi-hop Q{\textbackslash}\&A, slot filling, fact checking, and long-form Q{\textbackslash}\&A.},
urldate = {2026-03-16},
publisher = {arXiv},
author = {Xu, Shicheng and Pang, Liang and Shen, Huawei and Cheng, Xueqi and Chua, Tat-Seng},
month = feb,
year = {2024},
note = {arXiv:2304.14732},
keywords = {Computer Science - Computation and Language},
}
@misc{zhu_static_2025,
title = {From {Static} to {Dynamic}: {A} {Streaming} {RAG} {Approach} to {Real}-time {Knowledge} {Base}},
shorttitle = {From {Static} to {Dynamic}},
url = {http://arxiv.org/abs/2508.05662},
doi = {10.48550/arXiv.2508.05662},
abstract = {Dynamic streams from news feeds, social media, sensor networks, and financial markets challenge static RAG frameworks. Full-scale indices incur high memory costs; periodic rebuilds introduce latency that undermines data freshness; naive sampling sacrifices semantic coverage. We present Streaming RAG, a unified pipeline that combines multi-vector cosine screening, mini-batch clustering, and a counter-based heavy-hitter filter to maintain a compact prototype set. We further prove an approximation bound {\textbackslash}\$E{\textbackslash}[R(K{\textbackslash}\_t)] {\textbackslash}ge R{\textasciicircum}{\textbackslash}* - L Δ{\textbackslash}\$ linking retrieval quality to clustering variance. An incremental index upsert mechanism refreshes prototypes without interrupting queries. Experiments on eight real-time streams show statistically significant gains in Recall{\textbackslash}@10 (up to 3 points, p {\textless} 0.01), end-to-end latency below 15 ms, and throughput above 900 documents per second under a 150 MB budget. Hyperparameter sensitivity analysis over cluster count, admission probability, relevance threshold, and counter capacity validates default settings. In open-domain question answering with GPT-3.5 Turbo, we record 3.2-point gain in Exact Match and 2.8-point gain in F1 on SQuAD; abstractive summarization yields ROUGE-L improvements. Streaming RAG establishes a new Pareto frontier for retrieval augmentation.},
abstract = {Continual learning refers to the capability of a machine learning model to learn and adapt to new information, without compromising its performance on previously learned tasks. Although several studies have investigated continual learning methods for neural information retrieval (NIR) tasks, a well-defined task definition is still lacking, and it is unclear how typical learning strategies perform in this context. To address this challenge, a systematic task definition of continual NIR is presented, along with a multiple-topic dataset that simulates continuous information retrieval. A comprehensive continual neural information retrieval framework consisting of typical retrieval models and continual learning strategies is then proposed. Empirical evaluations illustrate that the proposed framework can successfully prevent catastrophic forgetting in neural information retrieval and enhance performance on previously learned tasks. The results also indicate that embedding-based retrieval models experience a decline in their continual learning performance as the topic shift distance and dataset volume of new tasks increase. In contrast, pretraining-based models do not show any such correlation. Adopting suitable learning strategies can mitigate the effects of topic shift and data augmentation in continual neural information retrieval.},
urldate = {2026-03-16},
journal = {Information Sciences},
author = {Hou, Jingrui and Cosma, Georgina and Finke, Axel},
month = jan,
year = {2025},
keywords = {Catastrophic forgetting, Continual learning, Data augmentation, Neural information retrieval, Topic shift},
pages = {121368},
}
@misc{ni_towards_2025,
title = {Towards {Trustworthy} {Retrieval} {Augmented} {Generation} for {Large} {Language} {Models}: {A} {Survey}},
abstract = {Retrieval-Augmented Generation (RAG) is an advanced technique designed to address the challenges of Artificial Intelligence-Generated Content (AIGC). By integrating context retrieval into content generation, RAG provides reliable and up-to-date external knowledge, reduces hallucinations, and ensures relevant context across a wide range of tasks. However, despite RAG's success and potential, recent studies have shown that the RAG paradigm also introduces new risks, including robustness issues, privacy concerns, adversarial attacks, and accountability issues. Addressing these risks is critical for future applications of RAG systems, as they directly impact their trustworthiness. Although various methods have been developed to improve the trustworthiness of RAG methods, there is a lack of a unified perspective and framework for research in this topic. Thus, in this paper, we aim to address this gap by providing a comprehensive roadmap for developing trustworthy RAG systems. We place our discussion around five key perspectives: reliability, privacy, safety, fairness, explainability, and accountability. For each perspective, we present a general framework and taxonomy, offering a structured approach to understanding the current challenges, evaluating existing solutions, and identifying promising future research directions. To encourage broader adoption and innovation, we also highlight the downstream applications where trustworthy RAG systems have a significant impact.},
urldate = {2026-02-03},
publisher = {arXiv},
author = {Ni, Bo and Liu, Zheyuan and Wang, Leyao and Lei, Yongjia and Zhao, Yuying and Cheng, Xueqi and Zeng, Qingkai and Dong, Luna and Xia, Yinglong and Kenthapadi, Krishnaram and Rossi, Ryan and Dernoncourt, Franck and Tanjim, Md Mehrab and Ahmed, Nesreen and Liu, Xiaorui and Fan, Wenqi and Blasch, Erik and Wang, Yu and Jiang, Meng and Derr, Tyler},
title = {A {Survey} on {RAG} {Meeting} {LLMs}: {Towards} {Retrieval}-{Augmented} {Large} {Language} {Models}},
shorttitle = {A {Survey} on {RAG} {Meeting} {LLMs}},
url = {http://arxiv.org/abs/2405.06211},
doi = {10.48550/arXiv.2405.06211},
abstract = {As one of the most advanced techniques in AI, Retrieval-Augmented Generation (RAG) can offer reliable and up-to-date external knowledge, providing huge convenience for numerous tasks. Particularly in the era of AI-Generated Content (AIGC), the powerful capacity of retrieval in providing additional knowledge enables RAG to assist existing generative AI in producing high-quality outputs. Recently, Large Language Models (LLMs) have demonstrated revolutionary abilities in language understanding and generation, while still facing inherent limitations, such as hallucinations and out-of-date internal knowledge. Given the powerful abilities of RAG in providing the latest and helpful auxiliary information, Retrieval-Augmented Large Language Models (RA-LLMs) have emerged to harness external and authoritative knowledge bases, rather than solely relying on the model's internal knowledge, to augment the generation quality of LLMs. In this survey, we comprehensively review existing research studies in RA-LLMs, covering three primary technical perspectives: architectures, training strategies, and applications. As the preliminary knowledge, we briefly introduce the foundations and recent advances of LLMs. Then, to illustrate the practical significance of RAG for LLMs, we systematically review mainstream relevant work by their architectures, training strategies, and application areas, detailing specifically the challenges of each and the corresponding capabilities of RA-LLMs. Finally, to deliver deeper insights, we discuss current limitations and several promising directions for future research. Updated information about this survey can be found at https://advanced-recommender-systems.github.io/RAG-Meets-LLMs/},
urldate = {2026-01-09},
publisher = {arXiv},
author = {Fan, Wenqi and Ding, Yujuan and Ning, Liangbo and Wang, Shijie and Li, Hengyun and Yin, Dawei and Chua, Tat-Seng and Li, Qing},
month = jun,
year = {2024},
note = {arXiv:2405.06211},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language, Computer Science - Information Retrieval},
}
@article{lewis_retrieval-augmented_2020,
title = {Retrieval-{Augmented} {Generation} for {Knowledge}-{Intensive} {NLP} {Tasks}},
journal = {arXiv: Computation and Language},
author = {Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and Küttler, Heinrich and Lewis, Mike and Yih, Wen-tau and Rocktäschel, Tim and Riedel, Sebastian and Kiela, Douwe},
year = {2020},
}
@article{liu_hm-rag_2025,
title = {{HM}-{RAG}: {Hierarchical} {Multi}-{Agent} {Multimodal} {Retrieval} {Augmented} {Generation}},
doi = {10.48550/arxiv.2504.12330},
journal = {arXiv.org},
author = {Liu, Pei and Liu, Xin and Yao, Ruoyu and Liu, Junming and Meng, Siyuan and Wang, Ding and Ma, Jun},
Une paire (question, question) où la question est exploitée comme question mais aussi comme document
\paragraph{Perturbations appliquées}
\begin{verbatim}
drone -> engin autonome
limitation -> contrainte
système -> dispositif
technique -> spécialisé
drone -> vecteur
limitation -> bornage
système -> ensemble
technique -> difficile
drone->gazon
limitation->animation
système->panneau
technique->magique
drone->ver de verre
limitation->plastique anarchique
système->esperluette en vent
technique->feuille entubée
Full-replace Synonymes 1
drone -> engin autonome
limitation -> contrainte
système -> dispositif
technique -> spécialisé
Full-replace Polysémie 1
drone -> vecteur
limitation -> bornage
système -> ensemble
technique -> difficile
Full-replace Elephants 1
drone -> gazon
limitation -> animation
système -> panneau
technique -> magique
Full-replace Elephants Roses 1
drone -> ver de verre
limitation -> plastique anarchique
système -> esperluette en vent
technique -> feuille entubée
\end{verbatim}
\paragraph{Données représentées}
Le graphique de gauche représente les scores de pertinence de la question comme document pour chaque perturbation. Les questions non modifiées (représentées en bleu) sont juste paddées pour faire correspondre la taille du document non modifié avec celui perturbé (ici de la question perturbée, représentées en rouge). Les scores de pertinence pour chaque perturbation sont dans l'ordre des perturbations citées. Les lignes verticales séparent les perturbations par niveau (de 1 à 4 en allant de gauche à droite) puis regroupent les pertuabtion de type full\_replace (le plus à droite, de niveau 1 à 4 en allant de gauche à droite)
Le graphique de droite est un histogramme des scores (en bleu pour la question paddée et en rouge pour la question perturbée)
\paragraph{Commentaires}
On observe bien la différence de représentation entre les questions et les question perturbées. On constate aussi que le système de padding is en place par \cite{parry_mechir_2025} provoque une perturbation des documents. Certaines perturbations améliorent la similarité de la question avec elle même. Cela est surprenant
\subsection{Différence des scores de petinence pour question paddée et question perturbée}
Le graphique de gauche associe à chaque perturbation (dans l'ordre cité) la différence entre les scores de similarité à la question de la question perturbée et la question paddée pour la perturbation. Les points en rouge correspondnet à aux différences négatives et les points en vert aux différences positives. On attend des scores positifs car perturber un objet (en l'occurrence une question) réduit sa similarité à lui même. Les pints sont séparés pour identifier les pertubations de type replace de diférent niveau (1 à 4 de gauche à droite). Le cinquième (dernier) groupe correspond aux pertuabtions de type full\_replace (niveau 1 à 4 de gauche à droite)\\
Le grphique de droite représente la distribution de ces différences.
\paragraph{Analyse des graphiques}
Pour le premier graphique, on constate une augmentation de da différence pour les niveau les plus élevés concernnat les pertubations de type replace. Les perturbation de type full\_replace ont un effet de perturbation moins important que le padding induit par le modèle.
