Hypercube-Based Retrieval-Augmented Generation for Scientific Question-Answering

Bowen Jin; Giri Narasimhan; Jiawei Han; Jimeng Shi; Runchu Tian; Shaowen Wang; Sizhe Zhou; Wei Hu

REVIEW 4 major objections 5 minor 1 cited by

Hypercube-Based Retrieval-Augmented Generation for Scientific Question-Answering

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Indexing documents by LLM-generated labels in a multi-dimensional hypercube makes retrieval-augmented generation more accurate, faster, and explainable for scientific question-answering.

desk verdict Solid RAG application of the text-cube idea, with evaluation gaps around threshold tuning and per-dataset dimensions that a serious referee should push on. read the letter →

arxiv 2505.19288 v2 pith:54UC34BR submitted 2025-05-25 cs.LG

Jimeng Shi , Sizhe Zhou , Bowen Jin , Wei Hu , Runchu Tian , Shaowen Wang , Giri Narasimhan , Jiawei Han This is my paper · ORCID

classification cs.LG

keywords retrieval-augmentedgenerationhypercubeindexingmulti-dimensionaltextcubescientificquestionansweringentitymatchingdenseretrievalexplainableRAGefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that standard RAG retrievers—sparse, dense, or graph-based—overlook the structured, multi-dimensional semantic information in documents, and that this omission hurts theme-specific scientific question-answering. To fix it, the authors propose Hypercube-RAG, which indexes documents into cells of a multi-dimensional cube according to fine-grained labels (location, event, theme, and similar dimensions) extracted by an LLM. Given a query, the system decomposes it into entities aligned with those dimensions, then retrieves documents by exact entity match plus dense similarity, returning documents that cover the query's key components. On three benchmarks across medicine, law, and geoscience, the method claims to outperform all tested baselines in answer accuracy and retrieval precision, while being one to two orders of magnitude faster than graph-based RAG and providing explainable retrieval through the cube labels.

What carries the argument

The central object is the Hypercube, a multi-dimensional text cube defined as a tensor product of dimensions $C_1 \otimes C_2 \otimes \dots \otimes C_m$, where each dimension is a set of fine-grained category labels (e.g., location, event, theme) and each document is assigned an $m$-tuple of labels by an LLM. Retrieval works by decomposing the query into entities aligned with the same dimensions, scoring documents by exact label matches plus a dense embedding similarity fallback, and ranking by coverage of the query's components. The cube's compact labels are what make retrieval constant-time, precise, and explainable.

What would settle it

Take a fresh scientific corpus and build the hypercube dimensions using only corpus entities (no peeking at test queries), then run Hypercube-RAG against e5 and HippoRAG 2 on a held-out QA set; if the F1 advantage drops below the reported 3.7% or reverses, the claim that the cube structure is the cause fails. A second, sharper test is to vary the K-means cluster count in dimension induction on the same corpus and check whether QA F1 swings by more than a point, which would show the reported consistency is not robust.

Watch

Extended reading notes

Core claim

Indexing documents into a hypercube of LLM-extracted labels allows a RAG system to retrieve more precise context than flat lexical or dense retrievers or graph-based methods, and to do so fast enough to scale. On three benchmarks spanning medicine, law, and geoscience, the authors report that Hypercube-RAG outperforms the strongest baselines by 3.7% average F1 and 5.3% average retrieval accuracy, with per-dataset retrieval precision gains of 3.9%, 3.7%, and 8.5%, while retrieving in time that stays near-constant in corpus size. Each retrieved document is justified by the cube labels that matched the query, giving an explainable retrieval trace.

Load-bearing premise

The method's accuracy depends on the automatically chosen hypercube dimensions matching the concepts that test questions actually ask about; if a new corpus yields dimensions that don't line up with the questions, exact entity matching will miss relevant documents and the reported advantage may disappear.

Editorial extensions

If this is right

On scientific QA benchmarks, switching from graph-based or dense-only RAG to Hypercube-RAG would raise answer F1 by about 3.7 points on average.
Retrieval precision improves enough that a system can return fewer documents with higher recall, so downstream LLM answers are less noisy.
Retrieval time becomes nearly independent of corpus size, since the search is over cube labels rather than over the full document set, making the approach usable for large noisy corpora.
The cube labels give an audit trail: a user can see exactly which entities matched, addressing the opacity problem of dense retrieval.
Because dimensions are per-dataset and LLM-derived, the same framework can adapt to new scientific domains without manual ontology construction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the dimension-induction step is the real source of gain, the advantage should transfer to any corpus with coherent entity clusters; a direct test would be to fix the hypercube dimensions on a training set, then apply to a held-out test set of queries from the same domain to see if the F1 gain persists.
The method's precision suggests a natural extension to multi-hop QA: decompose the question into sub-questions, retrieve from different cube cells, and combine answers, a direction the paper only sketches in its appendix as 'multiple hypercubes'.
The hypercube could serve as a lightweight, interpretable memory layer for continual learning, since adding new documents only requires mapping them into existing cube cells, similar to HippoRAG's memory framing but with explicit dimensions.
The reported efficiency gains depend on the cube being sparse (most cells empty); a corpus with uniformly distributed entity combinations might fill the cube and degrade the speed advantage, so a stress test with a topic-diverse corpus is warranted.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

This is a reasonable paper that does one genuinely new thing: it takes the text-cube concept, which existed in the data-mining literature, and turns it into a working RAG retrieval index. The method is simple, clearly described, and the reported efficiency gains over graph-based RAG are large and plausible. I also appreciate that the authors are upfront about the cube being prior work; the new part is the application and the query-decomposition/retrieval protocol.

On the positive side, the experiments cover three domains, the ablations show each component contributes, and the code and data are available. The case studies illustrate the explainability benefit nicely.

The soft spots are in the evaluation protocol. There are no error bars or significance tests anywhere, which matters because the reported QA gains are around 3-4% on average. More importantly, the semantic threshold tau is swept on the test data (Figure 7), and the paper never states what tau value was used for the main results in Tables 2 and 3. That is a straightforward reporting gap. The hypercube dimensions are also built separately for each dataset, and while this is disclosed, there is no validation protocol to show the method would work without this per-dataset configuration. The ablation that removes dimensions one at a time on SciDCC does not address generalization. These are not fatal flaws, but they mean the central claim should be read as "this works when you can tailor the cube to the corpus and pick tau on the benchmark," which is weaker than the paper's phrasing.

I don't think the stress-test note overstates the risk, but I also wouldn't call it load-bearing in the sense of invalidating the work. The method is coherent, and the gains are consistent across three datasets, so it's more likely that a properly held-out evaluation would shrink the margins rather than eliminate them.

Bottom line: this deserves a serious referee. A careful reviewer should ask for significance tests, a validation split for tau and dimension selection, and a statement of the actual tau used. If those are added, it's a solid subfield contribution.

Referee Report

4 major / 5 minor

Summary. The paper introduces Hypercube-RAG, a retrieval-augmented generation framework that organizes documents in a multi-dimensional cube structure. The hypercube dimensions are induced from the corpus by extracting entities, clustering them with K-means, and asking an LLM to summarize each cluster; documents are then indexed by LLM-extracted labels along each dimension. At query time, an LLM decomposes the question into entity/phrase components aligned with the pre-defined dimensions, and retrieval combines exact lexical matching of query entities against document labels with dense semantic matching controlled by a similarity threshold tau. Retrieved documents are ranked by full coverage of query components, falling back to partial coverage. Experiments on SciFact, LegalBench, and SciDCC report QA F1 scores of 91.5, 40.5, and 46.4, respectively, with retrieval precision improvements over the strongest baselines, and the paper also reports lower retrieval latency than graph-based RAG and provides qualitative explainability through cube labels.

Significance. If the reported gains hold under a stricter evaluation protocol, Hypercube-RAG is a valuable addition to structured RAG: it is simple, provides an interpretable retrieval rationale via cube labels, and shows a clear efficiency advantage over graph-based methods. The authors ship code and data, and they evaluate on three domains with multiple QA and retrieval metrics. However, the current evidence is weakened by two evaluation-protocol issues: the semantic threshold tau is tuned on the test sets, and all results are point estimates without variance or significance testing. These issues directly affect the central claim that Hypercube-RAG 'consistently outperforms' all tested baselines, so the paper needs a revision that separates method from configuration and quantifies uncertainty.

major comments (4)

[§5.5, Figure 7, and Tables 2–3] The semantic similarity threshold tau is swept over the test data (Figure 7), and the value used to produce the main results in Tables 2 and 3 is never reported. Because tau controls the balance between exact and dense matching, selecting it on the full test set makes the headline comparison vulnerable to selection bias. Please report the exact tau used, and select it on a validation split or through cross-validation before reporting test-set numbers.
[Tables 2–5, 8, and Figures 6–7] All results are reported as point estimates with no standard deviations, confidence intervals, or significance tests. Several advantages over the second-best baseline are small (e.g., SciFact F1 91.5 vs. 91.2; SciDCC Recall@5 92.4 vs. 91.2), so the claim of consistent improvement is not fully supported without repeated runs or a paired significance test. Please add variance estimates over multiple runs or a statistical test such as paired bootstrap or Wilcoxon signed-rank.
[§3.4, Eq. (7)] As written, D*_return = arg max_{d_i in D} |E(d_i)| selects documents with the largest total number of entities, independent of the query, so it cannot implement 'highest partial coverage' of the query components. The intended objective should involve the overlap with the query entities, such as |E(d_i) ∩ E(q)|. Please correct the equation and confirm that the implementation matches the described ranking.
[Appendix A.2 and §3.2, Table 9] The hypercube dimensions are constructed separately for each dataset (Table 9), and the query-decomposition prompt is conditioned on those dimensions. The paper does not analyze whether this dimension-selection procedure is stable with respect to the number of K-means clusters or whether the induced dimensions generalize to unseen query concepts. Since the exact-match term in Eq. (2) only fires when query entities coincide with the chosen dimensions, a stability analysis or a held-out protocol (e.g., inducing dimensions from a training corpus subset and evaluating on held-out questions) is needed to show that the reported advantage is not an artifact of favorable dimension choices.

minor comments (5)

[Appendix A.2, Table 9 caption] The caption of Table 9 is a copy-paste from Table 7 ('Document ID and content in the corpus...'); it should instead state that the table lists the hypercube dimensions used for each dataset.
[§6, Discussion] The text says 'four retrieval metrics and two QA metrics,' but Table 2 reports four QA metrics (F1, Semantic, Correctness, Completeness) and Table 3 reports four retrieval metrics (Precision@3, Recall@3, Precision@5, Recall@5). Please correct the count.
[§5.6, Case Study 2] There is a typo: 'LOCATON' should be 'LOCATION'.
[Appendix A.1] The step numbering in the hypercube design process lists '5 Clusters to Dimensions' and then '5 Document indexing,' skipping step 4; please renumber.
[Tables 2–3] Contriever is cited as [13] in the tables but appears as reference [14] in the bibliography; please align the citation numbering.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the empirical RAG comparison is externally grounded, with only mild benchmark-tuning concerns that do not rise to circularity.

full rationale

Hypercube-RAG is an empirical systems paper: the central claim is that the method outperforms baselines on three QA benchmarks, and this is supported by held-out test questions with GPT-4o generation rather than by an equation-level derivation. The hypercube dimensions (Table 9) are induced from each corpus by entity extraction, K-means clustering, and LLM summarization (Appendix A.1); they are not derived from the gold answers, and the retrieval scores in Eqs. (2)-(7) only compare query entities against document labels. The same-author citations for text cubes/Doc2Cube/Geospatial Knowledge Hypercube ([7,31,38]) are used as provenance for the structure, not as proof of the empirical advantage, and no uniqueness theorem is imported. Two evaluation choices warrant a correctness caveat but are not circular in the defined sense: the semantic threshold tau is studied on the same datasets in Section 5.5/Figure 7 rather than on a separate validation split, and the dimension sets are dataset-specific. These are hyperparameter- and index-design choices that could overstate generality, but they do not make the reported predictions equivalent to the method's inputs by construction; a held-out protocol for tau would strengthen the paper's external-validity claim.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

The method contributes an index structure and retrieval strategy rather than a physical model; its load-bearing assumptions are about LLM extraction quality and dimension coverage.

free parameters (3)

semantic similarity threshold tau = not stated (studied in Fig. 7)
Controls when dense matching triggers; the main experiments never report the chosen value, and Figure 7 tunes it on the test datasets.
hypercube dimension schema per dataset = e.g., SciDCC: location, person, event, organization, theme, date
Dimensions are induced per corpus with K-means and LLM summarization (Appendix A) and differ across datasets (Table 9), making the schema a hand- or LLM-chosen model component.
number of K-means clusters = not reported
K-means clustering is used to group extracted entities (Appendix A); the number of clusters k is not stated, affecting the final dimension set.

assumptions (2)

domain assumption LLM-extracted entities and query decompositions are accurate and consistently align with the pre-defined dimensions
The entire retrieval mechanism relies on exact matches between query labels and document labels; any systematic extraction failure breaks the sparse path.
ad hoc to paper The automatically induced dimension set covers the concepts in the test questions
Dimensions are built per dataset from the corpus (Section 3.2); there is no guarantee of coverage for new domains, and the paper does not test dimension transfer.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hypercube-Based Retrieval-Augmented Generation for Scientific Question-Answering." pith.science (2026). https://pith.science/paper/54UC34BR

@misc{pith2026250519288,
  author       = {Pith},
  title        = {Pith review of: Hypercube-Based Retrieval-Augmented Generation for Scientific Question-Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/54UC34BR}},
  note         = {Machine review of arXiv:2505.19288}
}

read the original abstract

Large language models (LLMs) often need to incorporate external knowledge to solve theme-specific problems. Retrieval-augmented generation (RAG) has shown its high promise, empowering LLMs to generate more qualified responses with retrieved external data and knowledge. However, most RAG methods retrieve relevant documents based on either sparse or dense retrieval methods or their combinations, which overlooks the essential, multi-dimensional, and structured semantic information present in documents. This structured information plays a critical role in finding concise yet highly relevant information for domain knowledge-intensive tasks, such as scientific question-answering (QA). In this work, we introduce a multi-dimensional (cube) structure, Hypercube, which can index and allocate documents in a pre-defined multi-dimensional space. Built on the hypercube, we further propose Hypercube-RAG, a novel RAG framework for precise and efficient retrieval. Given a query, Hypercube-RAG first decomposes it based on its entities, phrases, and topics along with pre-defined hypercube dimensions, and then retrieves relevant documents from cubes by aligning these decomposed components with corresponding dimensions. Experiments on three datasets across different domains demonstrate that our method improves response accuracy by 3.7% and retrieval accuracy by 5.3% over the strongest RAG baseline. It also boosts retrieval efficiency (speed) by one or two magnitudes faster than graph-based RAG. Notably, our Hypercube-RAG inherently offers explainability by revealing those underlying dimensions used for retrieval. The code and data are available at https://github.com/JimengShi/Hypercube-RAG.

Figures

Figures reproduced from arXiv: 2505.19288 by the authors.

**Figure 2.** Structure comparison: Graph vs. Hypercube. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 4.** Hypercube construction on a corpus. We present [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (8 more)

**Figure 3.** Figure 3: Hypercube-RAG framework. 1 Input: Input the query and the prompt into a LLM; 2 Decomposition: LLM decomposes the query into different dimensions; 3 Retrieval: according to these dimensions, we use Hypercube-RAG to retrieve relevant documents; 4 Augmentation: query is a…

**Figure 6.** Figure 6: Ablation study on model components. Hypercube Dimensions. We also conduct an ablation study on the hypercube dimensions: Location, Event, Date, Organization, Person, and Theme. To study their effectiveness, we remove each of the dimensions, which are represented as No…

**Figure 5.** Figure 5: Comparison of different RAG methods on the same query. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 7.** Figure 7: Performance vs. similarity threshold. Hypercube-RAG is proposed to innovate the knowledge retrieval based on those cube labels. The cube label-based retrieval integrates sparse lexical and dense semantic strategies, making it accurate, efficient, and inherently explai…

**Figure 8.** Figure 8: Hypercube Structure Design and Construction. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 9.** Figure 9: Access one cube cell in one hypercube. ✓ represents the touched cube cells. C.2 Case 2: long query with multiple themes In cases where a query is very diverse, they may need to access multiple cube cells such that the query information can be covered as much as possibl…

**Figure 10.** Figure 10: Access multiple cube cells in one hypercube. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

**Figure 11.** Figure 11: Access multiple cube cells in multiple hypercubes. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Benchmarking Knowledge-Extraction Attack and Defense on Retrieval-Augmented Generation
cs.CR 2026-02 conditional novelty 5.0 of 10

A unified benchmark comparing RAG knowledge-extraction attacks and defenses, showing query diversity boosts extraction, embedding attacks fail to transfer, and graph indexing raises per-token leakage.

Reference graph

Works this paper leans on

51 extracted references · 28 canonical work pages · cited by 1 Pith paper

[1]

Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations

2023
[2]

Isabelle Augenstein, Timothy Baldwin, Meeyoung Cha, Tanmoy Chakraborty, Giovanni Luca Ciampaglia, David Corney, Renee DiResta, Emilio Ferrara, Scott Hale, Alon Halevy, et al. 2024. Factuality challenges in the era of large language models and opportunities for fact-checking. Nature Machine Intelligence 6, 8 (2024), 852–863

2024
[3]

Dhivya Chandrasekaran and Vijay Mago. 2021. Evolution of semantic similar- ity—a survey. Acm Computing Surveys (Csur) 54, 2 (2021), 1–37

work page 2021
[4]

Hejie Cui, Alyssa Unell, Bowen Chen, Jason Alan Fries, Emily Alsentzer, Sanmi Koyejo, and Nigam Shah. 2025. TIMER: Temporal Instruction Modeling and Evaluation for Longitudinal Clinical Records. In Will Synthetic Data Finally Solve the Data Access Problem? https://openreview.net/forum?id=uBCAtA6M73

work page 2025
[5]

José Devezas and Sérgio Nunes. 2021. A review of graph-based models for entity-oriented search. SN Computer Science 2, 6 (2021), 437

work page 2021
[6]

Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130 (2024)

arXiv 2024
[7]

Wei Feng, Chao Zhang, Wei Zhang, Jiawei Han, Jianyong Wang, Charu Aggarwal, and Jianbin Huang. 2015. STREAMCUBE: Hierarchical spatio-temporal hash- tag clustering for event exploration over the Twitter stream. In 2015 IEEE 31st international conference on data engineering . IEEE, 1561–1572

work page 2015
[8]

Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. Lightrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779 (2024)

arXiv 2024

Show all 51 references

[9]

Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su
[10]

Bernal Jiménez Gutiérrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. 2025. From RAG to Memory: Non-Parametric Continual Learning for Large Language Models. arXiv preprint arXiv:2502.14802 (2025)

2025 arXiv
[11]

Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, Ma- hantesh Halappanavar, Ryan A Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al
[12]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2025. A survey on hallucination in large language models: Principles, taxonomy, chal- lenges, and open questions. ACM Transactions on In...

2025
[13]

arXiv preprint arXiv:2501.00309 (2024)

Retrieval-augmented generation with graphs (graphrag). arXiv preprint arXiv:2501.00309 (2024)

2024 arXiv
[14]

Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bo- janowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised Dense Infor- mation Retrieval with Contrastive Learning. Transactions on Machine Learning Research (2022). https://openreview.net/forum?i...

2022
[15]

Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bo- janowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense in- formation retrieval with contrastive learning. arXiv preprint arXiv:2112.09118 (2021)

2021 arXiv
[16]

Bowen Jin, Jinsung Yoon, Jiawei Han, and Sercan O Arik. 2025. Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG. In The Thir- teenth International Conference on Learning Representations . https://openreview. net/forum?id=oU3tpaR8fm

2025
[17]

Xiaonan Ji, Han-Wei Shen, Alan Ritter, Raghu Machiraju, and Po-Yin Yen. 2019. Visual exploration of neural document embedding in information retrieval: Se- mantics and feature selection. IEEE transactions on visualization and computer graphics 25, 6 (2019), 2181–2192

2019
[18]

SeongKu Kang, Shivam Agarwal, Bowen Jin, Dongha Lee, Hwanjo Yu, and Jiawei Han. 2024. Improving retrieval in theme-specific applications using a corpus topical taxonomy. In Proceedings of the ACM Web Conference 2024 . 1497–1508

2024
[19]

Ammar Ismael Kadhim. 2019. Term weighting for feature extraction on Twitter: A comparison between BM25 and TF-IDF. In 2019 international conference on advanced science and engineering (ICOASE) . IEEE, 124–128

2019
[20]

Kalpesh Krishna, Aurko Roy, and Mohit Iyyer. 2021. Hurdles to Progress in Long-form Question Answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technologies, Kristina Toutanova, Anna...

2021 doi
[21]

Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering.. In EMNLP (1). 6769–6781

2020
[22]

Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2021. Learning Dense Representations of Phrases at Scale. In ACL

2021
[23]

Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. 2025. NV-Embed: Improved Techniques for Train- ing LLMs as Generalist Embedding Models. InThe Thirteenth International Confer- ence on Learning Representations . https://op...

2025
[24]

Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics 12 (2024), 157–173

2024
[25]

Xi Victoria Lin, Xilun Chen, Mingda Chen, Weijia Shi, Maria Lomeli, Richard James, Pedro Rodriguez, Jacob Kahn, Gergely Szilvasy, Mike Lewis, et al. 2023. Ra-dit: Retrieval-augmented dual instruction tuning. In The Twelfth International Conference on Learning Representations

2023
[26]

Prakamya Mishra and Rohan Mittal. 2021. Neuralnere: Neural named entity rela- tionship extraction for end-to-end climate change knowledge graph construction. In Tackling climate change with machine learning workshop at ICML

2021
[27]

Jean Maillard, Barlas Oguz, and Wen-tau Yih. 2021. Multi-Faith Retrieval for Question Answering. In NAACL

2021
[28]

Nicholas Pipitone and Ghita Houir Alami. 2024. Legalbench-rag: A bench- mark for retrieval-augmented generation in the legal domain. arXiv preprint arXiv:2408.10343 (2024)

2024 arXiv
[29]

Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921 (2024)

2024 arXiv
[30]

Stephen E Robertson and Steve Walker. 1994. Some simple effective approxi- mations to the 2-poisson model for probabilistic weighted retrieval. In SIGIR’94: Proceedings of the Seventeenth Annual International ACM-SIGIR Conference on Research and Development in Information Retr...

1994
[31]

Tejaskumar Pujari, Anil Kumar Pakina, and Anshul Goel. 2023. Explainable AI and governance: Enhancing transparency and policy frameworks through retrieval-augmented generation (RAG). IOSR Journal of Computer Engineering (2023)

2023
[32]

Andrew Trotman, Antti Puurula, and Blake Burgess. 2014. Improvements to BM25 and language models examined. In Proceedings of the 19th Australasian Document Computing Symposium. 58–65

2014
[33]

Fangbo Tao, Chao Zhang, Xiusi Chen, Meng Jiang, Tim Hanratty, Lance Kaplan, and Jiawei Han. 2018. Doc2cube: Allocating documents to text cube without labeled data. In 2018 IEEE International Conference on Data Mining (ICDM) . IEEE, 1260–1265

2018
[34]

Gelan Wang, Yu Liu, Shukai Liu, Ling Zhang, and Liqun Yang. 2025. REMFLOW: RAG-enhanced multi-factor rainfall flooding warning in sponge airports via large language model. International Journal of Machine Learning and Cybernetics (2025), 1–21

2025
[35]

David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or fiction: Verifying scientific claims. arXiv preprint arXiv:2004.14974 (2020)

2020 arXiv
[36]

Shijie Wang, Jiani Huang, Zhikai Chen, Yu Song, Wenzhuo Tang, Haitao Mao, Wenqi Fan, Hui Liu, Xiaorui Liu, Dawei Yin, et al. 2024. Graph machine learning in the era of large language models (llms).ACM Transactions on Intelligent Systems and Technology (2024)

2024
[37]

Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672 (2024)

2024 arXiv
[38]

Zhaonan Wang, Bowen Jin, Wei Hu, Minhao Jiang, Seungyeon Kang, Zhiyuan Li, Sizhe Zhou, Jiawei Han, and Shaowen Wang. 2023. Geospatial Knowledge Hypercube. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems . 1–4

2023
[39]

Yi Wang, Wenhao Yu, Yining Wang, Chenguang Li, Soroush Liu, and Hoifung Poon. 2022. Text Embeddings by Weakly-Supervised Contrastive Pre-training. arXiv preprint arXiv:2212.03533 (2022)

2022 arXiv
[40]

Bennett, Junaid Ahmed, and Arnold Overwijk

Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate Nearest Neighbor Neg- ative Contrastive Learning for Dense Text Retrieval. In International Conference on Learning Representations. https://openre...

2021
[41]

Geemi P Wellawatte, Huixuan Guo, Magdalena Lederbauer, Anna Borisova, Matthew Hart, Marta Brucka, and Philippe Schwaller. 2025. ChemLit-QA: a human evaluated dataset for chemistry RAG tasks. Machine Learning: Science and Technology 6, 2 (2025), 020601. Conference acronym ’XX, ...

2025
[42]

date”, ‘location

Yunyi Zhang, Ruozhen Yang, Xueqiang Xu, Rui Li, Jinfeng Xiao, Jiaming Shen, and Jiawei Han. 2025. Teleclass: Taxonomy enrichment and llm-enhanced hierarchical text classification with minimal supervision. In Proceedings of the ACM on Web Conference 2025. 2032–2042. Hypercube-B...

2025
[43]

Qinggang Zhang, Shengyuan Chen, Yuanchen Bei, Zheng Yuan, Huachi Zhou, Zijin Hong, Junnan Dong, Hao Chen, Yi Chang, and Xiao Huang. 2025. A Sur- vey of Graph Retrieval-Augmented Generation for Customized Large Language Models. arXiv preprint arXiv:2501.13958 (2025)

2025
[45]

Comprehend the given question: understand what the question asks, how to answer it step by step, and all concepts, aspects, or directions that are relevant to each step
[46]

How much rainfall did Melbourne Beach, Florida receive from Tropical Storm Fay?

Compose queries to retrieve documents for answering the question: each document is indexed by the entities or phrases occurred inside and those entities or phrases lie within the following dimensions: {dimensions}. For each of the above dimensions, synthesize queries that are ...
[47]

query_dimension: ‘location’; query_content: ‘Atlantic’
[48]

query_dimension: ‘event’; query_content: ‘Tropical Storm Fay’
[49]

How many

query_dimension: ’theme’; query_content: ‘rainfall’ Prompt template for Question-Answering Answer the question based on the given retrieved documents. Question: {question}, Retrieved documents: {retrieved document}, Output requirements: - If the query asks the quantitative ana...
[50]

- Score 0 if it has factual errors or misrepresents key information

Correctness (0 or 1): - Score 1 if the predicted answer is generally accurate and aligns with the key points in the reference answer. - Score 0 if it has factual errors or misrepresents key information
[51]

correctness

Completeness (0 or 1): - Score 1 if the predicted answer covers the main points present in the reference answer. - Score 0 if it misses essential information or fails to address the core of the question. Output your evaluation in the following JSON format: {“correctness”: int,...

2018
[2024]

In The Thirty-eighth Annual Conference on Neural Information Processing Systems

Hipporag: Neurobiologically inspired long-term memory for large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations

2023

[2] [2]

Isabelle Augenstein, Timothy Baldwin, Meeyoung Cha, Tanmoy Chakraborty, Giovanni Luca Ciampaglia, David Corney, Renee DiResta, Emilio Ferrara, Scott Hale, Alon Halevy, et al. 2024. Factuality challenges in the era of large language models and opportunities for fact-checking. Nature Machine Intelligence 6, 8 (2024), 852–863

2024

[3] [3]

Dhivya Chandrasekaran and Vijay Mago. 2021. Evolution of semantic similar- ity—a survey. Acm Computing Surveys (Csur) 54, 2 (2021), 1–37

work page 2021

[4] [4]

Hejie Cui, Alyssa Unell, Bowen Chen, Jason Alan Fries, Emily Alsentzer, Sanmi Koyejo, and Nigam Shah. 2025. TIMER: Temporal Instruction Modeling and Evaluation for Longitudinal Clinical Records. In Will Synthetic Data Finally Solve the Data Access Problem? https://openreview.net/forum?id=uBCAtA6M73

work page 2025

[5] [5]

José Devezas and Sérgio Nunes. 2021. A review of graph-based models for entity-oriented search. SN Computer Science 2, 6 (2021), 437

work page 2021

[6] [6]

Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130 (2024)

arXiv 2024

[7] [7]

Wei Feng, Chao Zhang, Wei Zhang, Jiawei Han, Jianyong Wang, Charu Aggarwal, and Jianbin Huang. 2015. STREAMCUBE: Hierarchical spatio-temporal hash- tag clustering for event exploration over the Twitter stream. In 2015 IEEE 31st international conference on data engineering . IEEE, 1561–1572

work page 2015

[8] [8]

Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. Lightrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779 (2024)

arXiv 2024