Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

CG-RAG: Research Question Answering by Citation Graph Retrieval-Augmented LLMs

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that entangling lexical and semantic retrieval signals inside a citation graph, instead of fusing them after each document is scored, improves retrieval and generation for research question answering.

desk verdict The architecture is coherent and potentially useful, but the central SOTA claim is undercut by an unspecified train/eval split and a flat efficiency error. read the letter →

arxiv 2501.15067 v1 pith:LSHVHJZR submitted 2025-01-25 cs.IR cs.LG

Yuntong Hu , Zhihan Lei , Zhongjie Dai , Allen Zhang , Abhinav Angirekula , Zheng Zhang , Liang Zhao This is my paper · ORCID
classification cs.IRcs.LG
keywords retrieval-augmentedgenerationresearchquestionansweringcitationgraphshybridretrievalgraphneuralnetworkslexical-semanticcontextual
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

The paper is trying to establish that retrieval for research question answering should operate inside the citation graph rather than on isolated papers. Its proposed method, LeSeGR (Lexical-Semantic Graph Retrieval), entangles keyword-match scores and semantic-embedding scores during graph message passing, then scores each chunk against the query; the paper claims this retrieves the correct chunk more often than standard sparse, dense, and hybrid retrievers, and that the graph context also improves LLM-generated answers. On a biomedical research-QA benchmark it reports Hit@1 of 0.961 versus 0.915 for the strongest baseline, and on a computer-science leaderboard benchmark it reports 0.835 accuracy versus 0.769 for the best hybrid baseline. The claim matters because scholarly question answering increasingly relies on RAG, and current systems typically ignore citation links when deciding what to retrieve.

What carries the argument

The load-bearing object is the contextual citation graph built at chunk level: each paper is split into fixed-length chunks, chunks inside one paper are linked by adjacency or explicit cross-reference, and chunks in citing or cited papers are linked to the top-n most similar chunks in the other paper. The carrying mechanism is LeSeGR's entangled message-passing update: a query-to-chunk sparse score $\delta_{q i}$, a chunk-to-chunk dense score $\alpha_{ij}$ computed by an MLP over embedding differences, and a graph encoder (a Graph Transformer in the default configuration) that aggregates the products of these scores over each chunk's neighborhood. Proposition 4.1 is the identity that ties the mechanism to prior work: with no neighbors, LeSeGR's relevance score becomes the additive fusion of sparse and dense relevance, so the construction is a strict generalization of post-retrieval hybrid search.

What would settle it

Run LeSeGR with empty graph neighborhoods ($\mathcal{N}(i)=\emptyset$) on the same two benchmarks while keeping the same sparse and dense encoders; Proposition 4.1 predicts the scores collapse to post-retrieval Score Fusion, and a mismatch within noise would show the reduction proof or implementation is wrong. Also rerun the comparison with a defined held-out split for the 10% gold-labeled chunks; if the margins shrink to baseline levels, the graph-entanglement advantage is not established.

Watch

Extended reading notes

Core claim

The central discovery is that sparse and dense retrieval signals do not have to be combined after each document is scored; they can be propagated through the citation graph before scoring. LeSeGR builds a chunk-level citation graph, gives each chunk a sparse relevance score to the query and a dense pairwise relevance score to its neighboring chunks, and then runs message passing so each chunk's final representation absorbs lexical and semantic evidence from nearby chunks. The final relevance score is a dense match between the query and that entangled representation. The paper proves that when a chunk has no graph neighbors, this procedure reduces to the standard post-retrieval fusion of sparse and dense scores, making the method a generalization of existing hybrid retrieval rather than a separate family. The reported experiments on two research-QA benchmarks support the claim that graph context, not just more powerful embeddings, is what drives the gains.

Load-bearing premise

The reported advantage assumes that the 10% of samples used to train the retrieval model do not overlap with the questions used for the reported Hit@1 and accuracy numbers, but the paper does not define that split.

Editorial extensions

If this is right

  • RAG systems over scholarly literature should carry retrieval signals through citation links during scoring, not fuse document scores after retrieval.
  • Queries whose answer spans a chain of papers become reachable, because message passing lets relevant evidence propagate across citation links rather than requiring a single paper to match the query.
  • Retrieval gains translate into generation gains: the reported coherence, consistency, and relevance scores improve alongside Hit@1, so the retriever is the lever.
  • Existing hybrid retrievers can be upgraded to this entangled form without swapping their sparse or dense encoders, since the paper proves post-retrieval fusion is a special case.

Reading between the lines

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

  • The same entangled-retrieval idea would likely transfer to any linked document collection, such as legal opinions, clinical guidelines, or multi-page encyclopedias, where an answer lives across connected documents; the paper only demonstrates scholarly citation graphs, so this transfer is untested.
  • The ablation points to the dense encoder as the bottleneck, with LeSeGR's Hit@1 dropping from 0.961 with MiniLM to 0.838 with mContriever; pairing the graph entanglement with a stronger dense encoder than MiniLM is a natural next experiment not reported.
  • The generation step makes one LLM summarize graph context, compress it, and write the answer in a single prompt; decoupling these roles with separate models or longer context windows is an unexplored variant.
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.

Referee Report

4 major / 6 minor

Summary. The paper introduces CG-RAG, a retrieval-augmented generation framework for answering research questions over citation graphs. It constructs a chunk-level contextual citation graph with intra- and inter-document edges, proposes LeSeGR, which entangles sparse and dense retrieval scores through a graph encoder (message passing with query-relevance and edge-relevance weights), and feeds retrieved contextual subgraphs (summarized by the LLM) to GPT-4 for answer generation. The evaluation on PubMedQA-1k and PapersWithCodeQA reports Hit@1 and generation quality improvements over sparse, dense, and hybrid retrieval baselines, and the paper also claims a theoretical result that post-retrieval hybrid fusion is a special case of LeSeGR.

Significance. If the results are valid, the idea of entangled sparse-dense graph retrieval with chunk-level contextualization is a useful step for research QA, and the proposed contextual citation graph and graph-context summarization are sensible components. The paper also offers efficiency measurements and an ablation analysis. However, the current evidence is not sufficient to support the strong state-of-the-art claims: the train/evaluation protocol is ambiguous, the theoretical generality proof has a gap, and the quantitative comparisons lack statistical support. The strengths are the clear problem formulation and the comprehensive set of baselines, but the missing experimental rigor and the proof issue are decisive for the current version.

major comments (4)
  1. [Section 5.1 (Implementations, Datasets)] The training protocol is critically underspecified. The sentence "Training is conducted using CrossEntropy, with 10% of the samples labeled with gold-standard retrieval" does not state which samples are labeled or how the train/evaluation split is defined. Since PubMedQA-1k is described as containing "human-labeled gold-standard retrieval" for its 1,000 questions and PapersWithCodeQA's generative questions were created by first supplying LLMs with "the most relevant contexts labeled by humans," the natural reading is that the 10% labeled samples are drawn from the very same benchmark queries whose Hit@1 and accuracy are then reported in Tables 2 and 3. Under that reading, LeSeGR receives supervised labels from a subset of the evaluation queries, while every baseline (BM25, ColBERT, BGE-M3, ScoreFusion) is used off-the-shelf, making the comparison asymmetric and the claimed superiority an artifact of the protocol. The paper must specify the exact split (e.g., a held-out query set disjoint from the labeled samples, or a time/venue split of the citation graph) and confirm that no test query or its gold retrieval label was used in training. This is load-bearing for the central claims in the abstract and Section 5.2.
  2. [Section 4.3, Proposition 4.1 (Eqs. 10-17)] The claimed reduction of post-retrieval hybrid fusion to LeSeGR is not established. Equations (14)-(17) derive log s ∝ log f_sparse + log f_dense, i.e., a product of sparse and dense scores in the original score space. This is not the additive score fusion used by post-retrieval methods such as ScoreFusion, which combine f_sparse and f_dense by summation (with weights) in the original space. Recovering true additive fusion would require redefining the relevance functions as log-transformed scores, which the proof does not do. Furthermore, the proof assumes N(i)=∅ for every chunk, but the contextual citation graph in Section 4.2 always includes intra-document adjacency/cross-reference edges and inter-document Top-n edges; the "no contextual information" case therefore does not apply to the proposed setting. The theoretical generality claim in the abstract and Section 4.3 should either be corrected with a valid reduction (e.g., by explicitly zeroing neighbor messages via α_ij=0 or δ_qj=0) or substantially weakened.
  3. [Section 5.2, Tables 2-3] All retrieval and generation results are reported as point estimates with no error bars, standard deviations, or significance tests. The abstract and Section 5.2 repeatedly use "significantly outperforms" and "significantly surpasses" (e.g., the discussion following Table 2). With no indication of variance across runs or a paired statistical test, these statements are unsupported. The authors should report mean ± std over at least three runs (or bootstrap confidence intervals) and a paired test (e.g., Wilcoxon signed-rank) for the retrieval metrics, especially given that some absolute gains, such as PubMedQA Acc 0.778 vs 0.724 for ColBERT in Table 2, are modest and could be within run-to-run variability.
  4. [Section 5.4, Table 5] The ablation study selects the graph encoder, Top-n context size, and sparse/dense encoders based on Hit@1 on the full PubMedQA benchmark, which is the same benchmark used to report the headline results in Tables 2 and 3. Without a held-out validation split, hyperparameter and component selection is performed on the test set, which further compromises the reported superiority. The authors should tune on a validation split and report final results on an untouched test split.
minor comments (6)
  1. [Section 5.3, Table 4] The sentence "LeSeGR outperforms CLEAR in query speed while maintaining similar memory usage" is contradicted by Table 4, where CLEAR has a query latency of 16.07 ms versus LeSeGR's 403.94 ms. The text and the table should be reconciled.
  2. [Algorithm 1, line 13] The expression "𝑓sparseqsparse, csparse 𝑖 )" is missing an opening parenthesis; it should read "𝑓sparse(qsparse, csparse 𝑖)".
  3. [Figure 1 caption] The caption contains a blank space before "represents the dense embedding" where a symbol appears to have been dropped; the missing icon reference should be restored.
  4. [Table 4 header] The header "CPU & GPU Memory (MB) (MB) (ms) (ms)" has duplicated units and does not clearly identify which columns correspond to CPU memory, GPU memory, indexing time, and query latency. The column labels should be made explicit.
  5. [Equation (1)] Equation (1) uses arg max over θ to define the generation procedure, which is nonstandard because θ is a fixed pretrained parameter at inference. The equation should be rewritten as a conditional probability factorization without the arg max.
  6. [Template placeholders] The manuscript still contains ACM template placeholders, including the conference acronym, "Conference acronym 'XX," the year 2018, and "Received 20 February 2007." These should be removed or replaced before submission.

Circularity Check

1 steps flagged · score 5.0 of 10

Central SOTA claim rests on an unstated train/eval split: the same benchmark questions supply LeSeGR's gold-standard retrieval training labels and the Hit@1/accuracy being scored.

  1. fitted input called prediction [Section 5.1 (Implementations), with results in Tables 2 and 3]
    "Training is conducted using CrossEntropy, with 10% of the samples labeled with gold-standard retrieval, and optimized using the AdamW optimizer. ... PubMedQA-1k ... comprises 1,000 question-answer pairs ... with human-labeled gold-standard retrieval and answer annotations."

    LeSeGR's graph encoder and edge MLP are trained on gold-standard retrieval labels, yet the paper defines no held-out split or separate training set. The only labeled retrieval samples described are the PubMedQA-1k and PapersWithCodeQA questions whose Hit@1 and accuracy are then reported in Tables 2 and 3. As written, the 10% labeled samples can therefore be a subset of the very benchmark queries being scored, so the reported Hit@1=0.961 and Acc=0.835 would be partly fitted to the evaluation set rather than predicted. All baselines are used off-the-shelf with no benchmark-specific training, making the comparison asymmetric. The 'significantly outperforms' claim thus depends on an unstated separation between training and evaluation labels that the paper never establishes.

full rationale

The theoretical derivation is not circular: Proposition 4.1 explicitly exhibits a reduction of LeSeGR to post-retrieval hybrid fusion in the no-context case using algebraic identities, and that proposition is not used to justify the empirical SOTA numbers. Self-citations such as GRAG [12] appear only in related work and are not load-bearing for the main claim. The one circularity-relevant issue is the evaluation protocol: the paper trains with '10% of the samples labeled with gold-standard retrieval' without defining a train/test split, while the datasets it introduces are the same 1,000 PubMedQA and 924 PapersWithCodeQA questions whose retrieval and QA metrics are later reported. If the 10% labeled samples are drawn from those benchmarks, the reported superiority of LeSeGR over untrained baselines is partly a fit to the evaluation set. Because the text does not rule this out, the central claim is not self-contained as a clean prediction; it is a fitted-input-called-prediction risk rather than a formally demonstrated equivalence. If a proper held-out split exists, the remaining derivation is independent, which is why the score is moderate rather than severe.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on trained GNN and MLP parameters fitted to gold labels drawn from the same benchmarks used for evaluation, on domain assumptions about label correctness, citation extraction fidelity, and the UniEval evaluator, and on an unstated linearity assumption in the Proposition 4.1 proof. No new physical or conceptual entities are introduced beyond the methods themselves.

free parameters (5)
  • Graph Transformer parameters phi (GNN weights) = not reported
    Trained with CrossEntropy on 10% of benchmark samples labeled with gold-standard retrieval (Section 5.1); the reported Hit@1 and accuracy reflect these fitted parameters.
  • Edge MLP parameters phi (for alpha_ij) = not reported
    The MLP in Equation 8 computes dense chunk-chunk relevance and is trained as part of the retrieval model on the same gold labels (Sections 4.3, 5.1).
  • Top-n inter-document context per chunk = 4
    Selected via ablation on the PubMedQA evaluation graph (Table 5); n=2 gives 0.931 and n=8 gives 0.950, so this hyperparameter is tuned on the evaluation set.
  • Architecture hyperparameters = 2 layers, 4 heads, 1024 hidden, chunk length 8192 tokens
    Chosen by hand with no stated tuning protocol (Section 5.1, Implementations).
  • Fraction of samples labeled with gold-standard retrieval = 10%
    Controls how much supervised signal the retriever receives from the same benchmark used for evaluation (Section 5.1, Implementations).
assumptions (5)
  • domain assumption The gold-standard retrieval annotations in PubMedQA-1k and the human-labeled contexts in PapersWithCodeQA are correct and complete.
    All retrieval and generation metrics are computed against these labels (Section 5.1), and errors in them would propagate directly into Hit@1 and accuracy.
  • domain assumption Citation edges extracted from PubMed references and arXiv bibliographies faithfully represent true paper relationships.
    Inter-document edges are built from these extractions without manual validation (Section 5.1, Datasets).
  • ad hoc to paper The message function MSG in Equations 5 and 6 is linear or the identity.
    Proposition 4.1 factors the sparse score delta_qi out of the message to reduce LeSeGR to hybrid fusion; this step requires MSG to be linear, which the paper never states (Section 4.3).
  • domain assumption UniEval scores are a valid proxy for generated answer quality.
    Coherence, Consistency, and Relevance are measured by UniEval without calibration against human ratings (Section 5.1, Metrics).
  • standard math Sparse and dense relevance scores can be mapped to non-negative values before applying log.
    The proof of Proposition 4.1 requires f_sparse and f_dense to be in R+ for the log transformation, and the paper states this requirement explicitly (Section 4.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of CG-RAG: Research Question Answering by Citation Graph Retrieval-Augmented LLMs." pith.science (2026). https://pith.science/paper/LSHVHJZR

@misc{pith2026250115067,
  author       = {Pith},
  title        = {Pith review of: CG-RAG: Research Question Answering by Citation Graph Retrieval-Augmented LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSHVHJZR}},
  note         = {Machine review of arXiv:2501.15067}
}
read the original abstract

Research question answering requires accurate retrieval and contextual understanding of scientific literature. However, current Retrieval-Augmented Generation (RAG) methods often struggle to balance complex document relationships with precise information retrieval. In this paper, we introduce Contextualized Graph Retrieval-Augmented Generation (CG-RAG), a novel framework that integrates sparse and dense retrieval signals within graph structures to enhance retrieval efficiency and subsequently improve generation quality for research question answering. First, we propose a contextual graph representation for citation graphs, effectively capturing both explicit and implicit connections within and across documents. Next, we introduce Lexical-Semantic Graph Retrieval (LeSeGR), which seamlessly integrates sparse and dense retrieval signals with graph encoding. It bridges the gap between lexical precision and semantic understanding in citation graph retrieval, demonstrating generalizability to existing graph retrieval and hybrid retrieval methods. Finally, we present a context-aware generation strategy that utilizes the retrieved graph-structured information to generate precise and contextually enriched responses using large language models (LLMs). Extensive experiments on research question answering benchmarks across multiple domains demonstrate that our CG-RAG framework significantly outperforms RAG methods combined with various state-of-the-art retrieval approaches, delivering superior retrieval accuracy and generation quality.

Figures

Figures reproduced from arXiv: 2501.15067 by the authors.

Figure 1
Figure 1. Illustration of retrieval-augmented research question answering using: (a) sparse retrieval based on lexical matches, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of Contextualized Graph Retrieval [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

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

  1. KAMR: Grounding Generation via Knowledge-Aligned Multi-hop Retrieval

    cs.IR 2026-07 conditional novelty 6.0 of 10

    Partial-alignment contrastive pretraining plus anchor-then-expand graph retrieval improves multi-hop KG evidence recovery and downstream QA over strong dense and graph RAG baselines.

Reference graph

Works this paper leans on

51 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [1]

    Sören Auer, Allard Oelen, Muhammad Haris, Markus Stocker, Jennifer D’Souza, Kheir Eddine Farfar, Lars Vogt, Manuel Prinz, Vitalis Wiens, and Mohamad Yaser Jaradeh. 2020. Improving access to scientific literature with knowledge graphs. Bibliothek Forschung und Praxis 44, 3 (2020), 516–529

    work page 2020

  2. [2]

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al . 2024. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology 15, 3 (2024), 1–45

    2024

  3. [3]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216 (2024)

    arXiv 2024

  4. [4]

    Xilun Chen, Kushal Lakhotia, Barlas Oğuz, Anchit Gupta, Patrick Lewis, Stan Peshterliev, Yashar Mehdad, Sonal Gupta, and Wen-tau Yih. 2021. Salient phrase aware dense retrieval: can a dense retriever imitate a sparse one? arXiv preprint arXiv:2110.06918 (2021)

    arXiv 2021

  5. [5]

    Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

    arXiv 2018

  6. [6]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, 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]

    Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2020. Language-agnostic BERT sentence embedding. arXiv preprint arXiv:2007.01852 (2020)

    arXiv 2020

  8. [8]

    Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Benjamin Van Durme, and Jamie Callan. 2021. Complement lexical retrieval model with semantic residual embeddings. In Advances in Information Retrieval: 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28–April 1, 2021, Proceedings, Part I

    work page 2021

Show all 51 references
  1. [9]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023)

    2023 arXiv

  2. [10]

    Hamed Babaei Giglou, Tilahun Abedissa Taffa, Rana Abdullah, Aida Usmanova, Ricardo Usbeck, Jennifer D’Souza, and Sören Auer. 2024. Scholarly Question Answering using Large Language Models in the NFDI4DataScience Gateway. arXiv preprint arXiv:2406.07257 (2024)

    2024 arXiv

  3. [11]

    Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh V Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. 2024. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. arXiv preprint arXiv:2402.07630 (2024)

    2024 arXiv

  4. [12]

    Yuntong Hu, Zhihan Lei, Zheng Zhang, Bo Pan, Chen Ling, and Liang Zhao. 2024. GRAG: Graph Retrieval-Augmented Generation. arXiv preprint arXiv:2405.16506 (2024)

    2024 arXiv

  5. [13]

    Yuntong Hu, Zheng Zhang, and Liang Zhao. 2023. Beyond Text: A Deep Dive into Large Language Models’ Ability on Understanding Graph Data. arXiv preprint arXiv:2310.04944 (2023)

    2023 arXiv

  6. [14]

    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

  7. [15]

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William W Cohen, and Xinghua Lu

  8. [16]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP...

    2020

  9. [17]

    Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 39–48

    2020

  10. [18]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  11. [19]

    Saar Kuzi, Mingyang Zhang, Cheng Li, Michael Bendersky, and Marc Najork. 2020. Leveraging semantic and lexical matching to improve the recall of document retrieval systems: A hybrid approach. arXiv preprint arXiv:2010.01195 (2020)

    2020 arXiv

  12. [20]

    Jens Lehmann, Antonello Meloni, Enrico Motta, Francesco Osborne, Diego Re- forgiato Recupero, Angelo Antonio Salatino, and Sahar Vahdati. 2024. Large Language Models for Scientific Question Answering: An Extensive Analysis of the SciQA Benchmark. In European Semantic Web Confe...

    2024

  13. [21]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing S...

    2020

  14. [22]

    Huayang Li, Yixuan Su, Deng Cai, Yan Wang, and Lemao Liu. 2022. A survey on retrieval-augmented text generation. arXiv preprint arXiv:2202.01110 (2022)

    2022 arXiv

  15. [23]

    I Loshchilov. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  16. [24]

    Alejandro Lozano, Scott L Fleming, Chia-Chun Chiang, and Nigam Shah. 2023. Clinfo. ai: An open-source retrieval-augmented large language model system for answering medical questions using scientific literature. In PACIFIC SYMPOSIUM ON BIOCOMPUTING 2024 . World Scientific, 8–23

    2023

  17. [25]

    Yi Luan, Jacob Eisenstein, Kristina Toutanova, and Michael Collins. 2021. Sparse, dense, and attentional representations for text retrieval. Transactions of the Association for Computational Linguistics 9 (2021), 329–345

    2021

  18. [26]

    Priyanka Mandikal and Raymond Mooney. 2024. Sparse Meets Dense: A Hybrid Approach to Enhance Scientific Document Retrieval. arXiv preprint arXiv:2401.04055 (2024)

    2024 arXiv

  19. [27]

    Kurnia Muludi, Kaira Milani Fitria, Joko Triloka, et al. 2024. Retrieval-Augmented Generation Approach: Document Question Answering using Large Language Model. International Journal of Advanced Computer Science & Applications 15, 3 (2024)

    2024

  20. [28]

    Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction. arXiv preprint arXiv:1904.08375 (2019)

    2019 arXiv

  21. [29]

    Vít Novotn`y and Michal Stefánik. 2022. Combining Sparse and Dense Information Retrieval.. In CLEF (Working Notes). 104–118

    2022

  22. [30]

    N Reimers. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT- Networks. arXiv preprint arXiv:1908.10084 (2019)

    2019 arXiv

  23. [31]

    Stephen Robertson, Hugo Zaragoza, et al . 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval 3, 4 (2009), 333–389. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Yuntong et al

    2009

  24. [32]

    Gerard Salton, Anita Wong, and Chung-Shu Yang. 1975. A vector space model for automatic indexing. Commun. ACM 18, 11 (1975), 613–620

    1975

  25. [33]

    Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjin Wang, and Yu Sun. 2020. Masked label prediction: Unified message passing model for semi- supervised classification. arXiv preprint arXiv:2009.03509 (2020)

    2020 arXiv

  26. [34]

    Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen, Tharindu Kalu- arachchi, Rajib Rana, and Suranga Nanayakkara. 2023. Improving the domain adaptation of retrieval augmented generation (RAG) models for open domain question answering. Transactions of the Association for Co...

    2023

  27. [35]

    Markus Stocker, Allard Oelen, Mohamad Yaser Jaradeh, Muhammad Haris, Omar Arab Oghli, Golsa Heidari, Hassan Hussein, Anna-Lena Lorenz, Salomon Kabenamualu, Kheir Eddine Farfar, et al. 2023. FAIR scientific information with the open research knowledge graph. FAIR Connect 1, 1 (...

    2023

  28. [36]

    Tilahun Abedissa Taffa and Ricardo Usbeck. 2023. Leveraging LLMs in Scholarly Knowledge Graph Question Answering.. In QALD/SemREC@ ISWC

    2023

  29. [37]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  30. [38]

    Chengrui Wang, Qingqing Long, Xiao Meng, Xunxin Cai, Chengjun Wu, Zhen Meng, Xuezhi Wang, and Yuanchun Zhou. 2024. BioRAG: A RAG-LLM Frame- work for Biological Question Reasoning. arXiv preprint arXiv:2408.01107 (2024)

    2024 arXiv

  31. [39]

    Haiwen Wang, Le Zhou, Weinan Zhang, and Xinbing Wang. 2021. LiteratureQA: A Qestion Answering Corpus with Graph Knowledge on Academic Literature. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 4623–4632

    2021

  32. [40]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533 (2022)

    2022 arXiv

  33. [41]

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou

  34. [42]

    Yuhao Wang, Ruiyang Ren, Junyi Li, Wayne Xin Zhao, Jing Liu, and Ji-Rong Wen. 2024. REAR: A Relevance-Aware Retrieval-Augmented Framework for Open-Domain Question Answering. arXiv preprint arXiv:2402.17497 (2024)

    2024 arXiv

  35. [43]

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 641–649

    2024

  36. [44]

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor nega- tive contrastive learning for dense text retrieval. arXiv preprint arXiv:2007.00808 (2020)

    2020 arXiv

  37. [45]

    Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong, Qi Liu, and Zhaofeng Liu. 2024. Evaluation of Retrieval-Augmented Generation: A Survey. arXiv preprint arXiv:2405.07437 (2024)

    2024 arXiv

  38. [46]

    Cyril Zakka, Rohan Shad, Akash Chaurasia, Alex R Dalal, Jennifer L Kim, Michael Moor, Robyn Fong, Curran Phillips, Kevin Alexander, Euan Ashley, et al. 2024. Almanac—retrieval-augmented language models for clinical medicine. NEJM AI 1, 2 (2024), AIoa2300068

    2024

  39. [47]

    Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, and Bin Cui. 2024. Retrieval-augmented generation for ai-generated content: A survey. arXiv preprint arXiv:2402.19473 (2024)

    2024 arXiv

  40. [48]

    Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2024. Dense text retrieval based on pretrained language models: A survey. ACM Transactions on Information Systems 42, 4 (2024), 1–60

    2024

  41. [49]

    Ming Zhong, Yang Liu, Da Yin, Yuning Mao, Yizhu Jiao, Pengfei Liu, Chenguang Zhu, Heng Ji, and Jiawei Han. 2022. Towards a unified multi-dimensional evalua- tor for text generation. arXiv preprint arXiv:2210.07197 (2022). Received 20 February 2007; revised 12 March 2009; accep...

    2022 arXiv

  42. [2019]

    arXiv preprint arXiv:1909.06146 (2019)

    Pubmedqa: A dataset for biomedical research question answering. arXiv preprint arXiv:1909.06146 (2019)

    2019 arXiv

  43. [2020]

    Advances in Neural Information Processing Systems 33 (2020), 5776–5788

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems 33 (2020), 5776–5788

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.