Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

Knowledge Graph-Guided Retrieval Augmented Generation

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

Pith's one-line read KG2RAG claims that linking chunks with an LLM-built knowledge graph and expanding retrieval one hop along that graph improves answer F1 from 0.587 to 0.631 on HotpotQA-Full and from 0.380 to 0.419 on MuSiQue.

desk verdict Solid KG-RAG system paper with real engineering value, but the causal role of the KG is underdetermined by missing candidate-pool controls. read the letter →

arxiv 2502.06864 v1 pith:ONGIS45P submitted 2025-02-08 cs.CL cs.AI

classification cs.CLcs.AI
keywords retrieval-augmentedgenerationknowledgegraphmulti-hopquestionansweringgraph-guidedretrievalexpansioncontextorganizationHotpotQALLMtripletextractionRAGbaselines
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 ordinary semantic retrieval misses the fact-level connections between text chunks, and that a knowledge graph built from the documents can supply those connections. It proposes KG2RAG, which retrieves seed chunks by embedding similarity, expands the set along one-hop entity links in an LLM-extracted graph, and then re-organizes the expanded chunks into paragraphs using maximum spanning trees of each connected component. On HotpotQA and its shuffled variants the method reports higher response F1 and retrieval F1 than semantic, hybrid, and graph-based RAG baselines, with the largest gains in the fullwiki setting. The point of the exercise is that a cheap, offline graph index can make retrieved context more diverse and coherent, which matters for multi-hop questions whose answer requires stitching facts across documents. A reader should care because multi-hop open-domain QA is a standing weakness of plain RAG.

What carries the argument

The load-bearing object is the chunk-knowledge-graph association $G = \{(h, r, t, c) \mid c \in D\}$, a set of LLM-extracted triplets each tied to the chunk that supplied it. This graph lets the method replace 'physically adjacent or textually similar' expansion with entity-connected expansion: the retrieval module reads out all chunks attached to triplets within an $m$-hop neighborhood of the seed subgraph, and the organization module turns each connected component of the weighted subgraph into a maximum spanning tree, roots the tree at the highest-weight edge, and walks it depth-first to produce a coherent paragraph for the LLM prompt. The $m=1$ choice in the reported experiments is what keeps expansion broad enough to find missing facts without flooding the context window.

What would settle it

Count how many of the gold supporting sentences in HotpotQA are represented in the extracted KG as entity-overlapping triplets; if a large share of answer-relevant chunk pairs are never linked in the graph, then the reported response improvements cannot transfer to corpora where extraction is weaker.

Watch

Extended reading notes

Core claim

The central claim is that fact-level structure, not just lexical or semantic similarity, should drive retrieval expansion and context ordering in RAG. Concretely, KG2RAG constructs a graph whose edges are triplets (head, relation, tail) extracted from chunks by an LLM, with each triplet linked to its source chunk. After semantic retrieval produces seed chunks, a BFS traversal over the graph's entities with $m=1$ pulls in chunks that share entities or relations with the seeds even when their embeddings are not close to the query. The expanded subgraph is then split into connected components, reduced to maximum spanning trees weighted by query-chunk similarity, and serialized into paragraphs by depth-first search, so the final prompt contains a small number of organized passages rather than a long concatenation of isolated chunks. The paper's experiments report that this two-stage design beats all tested RAG baselines on response F1 and retrieval F1 across HotpotQA-Dist, HotpotQA-Full, shuffled variants, MuSiQue, and TriviaQA.

Load-bearing premise

The entire gain rests on the LLM extracting accurate and complete triplets from the documents: if extraction misses or corrupts the facts that connect an answer's supporting chunks, the one-hop graph expansion has nothing to walk and the method degrades toward plain semantic retrieval.

Editorial extensions

If this is right

  • If the reported gains hold, plain semantic RAG systems can be upgraded by one offline pass that extracts triplets and builds a chunk-to-triplet index, without changing the generator.
  • The fullwiki results (0.631 vs 0.587 response F1) suggest graph-guided expansion matters most when the candidate pool is large and semantic similarity alone cannot distinguish the useful documents.
  • The ablation shows context organization contributes a large share of retrieval precision: removing it drops retrieval F1 from 0.436 to 0.259 while the chunk count doubles, implying the MST filter is what keeps prompts clean.
  • The shuffled-dataset gains indicate that part of the advantage does not depend on the LLM having memorized Wikipedia, since entity renaming forces reliance on retrieved context.
  • Low sensitivity to $k$ and $m$ in the reported experiments means the method should transfer without much per-dataset tuning.

Reading between the lines

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

  • Beyond the paper, the same graph index could be reused per query for answer justification: the MST path that connects the seed to an expanded chunk is a ready-made chain of evidence explaining why a retrieved fact was included.
  • A testable extension would be to replace the single-hop BFS with entity-similarity or relation-type-aware expansion, so that queries requiring multi-relational chains (not just shared entities) can be answered without growing $m$.
  • If triplet extraction quality is the hidden bottleneck, then a cheaper extractor or a distilled model might preserve most of the gains, since the graph itself, not the generator, does the retrieval work.
  • The method's gains on TriviaQA, a single-hop dataset with long documents, suggest the organizational half (MST paragraphs) may matter independently of multi-hop reasoning; an ablation that runs only the organizer on semantic seeds would test this.
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 / 5 minor

Summary. The paper proposes KG2RAG, a retrieval-augmented generation framework that builds an offline knowledge graph by prompting an LLM to extract triplets from document chunks (Eq. 1), retrieves seed chunks by semantic similarity (Eq. 2), expands the seed set through BFS over the KG (Eqs. 4-5), then filters and arranges the expanded chunks into paragraphs via maximum spanning trees and cross-encoder reranking (Eqs. 6-8) before feeding them to an LLM. Experiments on HotpotQA, two shuffled variants, MuSiQue, and TriviaQA compare KG2RAG to Semantic RAG, Hybrid RAG, LightRAG, and GraphRAG in response quality and retrieval quality, with additional ablations, sensitivity analyses for top-k and m, robustness to random triplet deletion, and efficiency measurements. The central claim is that using LLM-extracted fact-level KG structure for chunk expansion and context organization consistently outperforms semantic- and hybrid-retrieval RAG baselines.

Significance. If the empirical claim is robust, the paper makes a useful and timely contribution: it shows a concrete way to use automatically extracted knowledge graphs to improve multi-hop open-domain QA beyond standard semantic/hybrid retrieval, and it provides release code and constructed datasets. The paper's strengths include evaluation across multiple datasets and settings, ablations isolating the two proposed modules, sensitivity analyses for k and m, efficiency comparisons, and a self-acknowledged limitation section. However, the significance is currently tempered by three gaps: KG extraction quality is never measured, the causal role of graph structure is not isolated from candidate-pool size, and the reported gains are not accompanied by statistical reliability evidence. These gaps are fixable within the manuscript's scope, but they are load-bearing for the central claim.

major comments (4)
  1. [§3.2, Tables 1 and 2] The paper's central claim of 'consistent outperformance' rests on small absolute differences in several conditions, yet no error bars, significance tests, or multiple-seed results are reported. For example, on HotpotQA-Dist the response F1 gap between KG2RAG (0.663) and Semantic RAG + rerank (0.652) is 0.011, and on HotpotQA-Full the retrieval F1 gap is 0.310 versus 0.306. Without variance estimates or paired significance tests, these differences may not be reliable. Please report confidence intervals, bootstrap tests, or results over multiple generation seeds, at least for the main tables.
  2. [§2.2, Tables 1 and 3] The graph-guided expansion is only compared against semantic and hybrid baselines that retrieve and rerank exactly k=10 chunks, while the full KG2RAG pipeline considers a much larger candidate pool before organization (16.76 chunks on average in the ablation, before filtering to 8.11). This leaves open the possibility that the observed gains come from a larger candidate pool plus reranking rather than from graph structure. Please add a control that retrieves a larger top-k by embedding similarity (e.g., top-50), reranks with the same cross-encoder, and feeds the top-10 to the same LLM. A second useful control is a random-graph or edge-shuffled graph with the same degree distribution; without such a control, the ablation in Table 3 shows sensitivity to component removal but does not establish that fact-level graph content, rather than additional candidates, drives the improvement.
  3. [§2.1, Fig. 4, and Table 6] The pipeline depends critically on the quality of LLM-extracted triplets: Eq. (5) retrieves chunks only through triplets in the expanded subgraph, and Fig. 4 shows that many chunks have very few triplets. However, the paper never measures extraction precision or recall, entity resolution quality, or relation validity. The robustness analysis in Table 6 randomly drops 5% or 10% of triplets, which tests insensitivity to deletion but not whether systematically missing or incorrect facts undermine the approach. Please evaluate extraction on a sample against a gold KG or human annotations, or at least report precision/recall and entity-resolution statistics for the constructed KG; this is needed to support the claim that fact-level relationships, rather than unrelated graph connectivity, provide the benefit.
  4. [§3.1, Shuffle-HotpotQA construction] The Shuffle-HotpotQA variants are created by randomly replacing each entity with another entity in the same category and then propagating the replacement to queries, triplets, and documents. This construction can create internally inconsistent facts (e.g., a replacement entity receiving attributes from the original entity), and it is not clear whether the same category constraint is sufficient to preserve answerability and relation validity. Please describe how entity replacement was validated or filtered, and report statistics on how many questions remain answerable after shuffling; this affects the interpretation of the 'reliance on RAG rather than prior knowledge' framing.
minor comments (5)
  1. [Table 2] In the HotpotQA-Full row for Semantic RAG + rerank, the precision value is printed as '0339' instead of '0.339'; please fix this formatting error.
  2. [§2.2, Eq. (6)] The definition of U^m_q as a set of edges with attached meta-information is notationally unusual for an undirected weighted graph; please define the vertex set explicitly and clarify whether multiple edges between the same entity pair are retained before MST construction.
  3. [Appendix A.1] The additional experiments on MuSiQue and TriviaQA are reported without the setup details used for KG construction, chunk size, top-k, and the m value on those datasets; please state whether the same hyperparameters and extraction prompt were used.
  4. [§1 and §2.2] The phrase 'Shed light by such insights' is grammatically awkward; consider replacing with 'Motivated by these insights.' Also, the discussion of alternative expansion methods mentions context-window expansion but does not provide a quantitative comparison with it; consider adding this as a baseline or clarifying why it is omitted.
  5. [Fig. 5] The sensitivity plots for varying k would be easier to interpret if the y-axis ranges were consistent across subplots and if error bars or at least raw numerical values were included; currently the small differences at k=10 and k=15 are hard to assess visually.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KG2RAG's central claims rest on held-out empirical comparisons, not on definitions or fitted predictions.

full rationale

The paper's central derivation is empirical: KG2RAG constructs a query-independent KG via LLM-based triplet extraction (Sec. 2.1, Eq. 1), retrieves seed chunks by semantic similarity (Eq. 2), expands via BFS over the KG (Eqs. 3-5), organizes contexts via MST and cross-encoder reranking (Eqs. 6-8), and feeds the result to an LLM for generation. Response and retrieval metrics are computed on held-out labels from HotpotQA, Shuffle-HotpotQA, MuSiQue, and TriviaQA. The KG construction uses no answer or query information, and the hyperparameters k and m are fixed or swept rather than fitted to maximize test scores. No equation defines a predicted quantity in terms of the target labels, nor is any fitted parameter renamed as a prediction. The use of the same LLM for triplet extraction and generation is a shared component across all methods, not a circular reduction: extraction is unsupervised and applied uniformly to all documents, and the Shuffle variants reduce reliance on parametric memory. Self-citations to the authors' prior work (e.g., entity linkage, KG completion) appear only as background examples and are not load-bearing for the main claim. The absence of a larger-candidate-pool baseline for Semantic or Hybrid RAG is an experimental-control limitation rather than circularity, because the reported comparisons remain genuine out-of-sample evaluations. No specific circular step can be quoted from the paper, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method is an empirical pipeline; its soundness depends on the unmeasured quality of the automatically constructed knowledge graph and on the assumption that entity-connected chunks are useful context. The only manually set hyperparameters are top-k and the BFS depth m, both examined in sensitivity analyses. No new physical or mathematical entities are introduced.

free parameters (3)
  • top-k = 10
    Number of chunks selected or fed to the LLM; fixed across experiments with a sensitivity analysis in Fig. 5.
  • m-hop expansion depth = 1
    BFS depth for graph-guided expansion; set to 1 after sensitivity analysis in Table 5.
  • chunk size = predefined, not specified
    Documents are split into chunks of a predefined size; the exact size is not stated in the paper.
assumptions (5)
  • domain assumption Llama-3 triplet extraction produces accurate, faithful factual triplets from each chunk.
    Section 2.1 and Fig. 3; the KG quality is never evaluated against gold annotations.
  • domain assumption Chunks connected through shared entities within m-hop are useful context for answering multi-hop questions.
    Section 2.2 graph-guided expansion; this is the mechanism that finds additional chunks.
  • domain assumption HotpotQA evaluation script and F1 metrics adequately measure retrieval and response quality.
    Section 3.1 evaluation metrics.
  • domain assumption Entity replacement in Shuffle-HotpotQA preserves question answerability and difficulty distribution.
    Section 3.1 dataset construction; the paper notes that new triplets such as (Rick and Morty, language, French) appear.
  • domain assumption Cross-encoder reranking on triplet representations accurately scores query relevance of chunks.
    Section 2.3, Eq. (8).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge Graph-Guided Retrieval Augmented Generation." pith.science (2026). https://pith.science/paper/ONGIS45P

@misc{pith2026250206864,
  author       = {Pith},
  title        = {Pith review of: Knowledge Graph-Guided Retrieval Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ONGIS45P}},
  note         = {Machine review of arXiv:2502.06864}
}
abstract

Retrieval-augmented generation (RAG) has emerged as a promising technology for addressing hallucination issues in the responses generated by large language models (LLMs). Existing studies on RAG primarily focus on applying semantic-based approaches to retrieve isolated relevant chunks, which ignore their intrinsic relationships. In this paper, we propose a novel Knowledge Graph-Guided Retrieval Augmented Generation (KG$^2$RAG) framework that utilizes knowledge graphs (KGs) to provide fact-level relationships between chunks, improving the diversity and coherence of the retrieved results. Specifically, after performing a semantic-based retrieval to provide seed chunks, KG$^2$RAG employs a KG-guided chunk expansion process and a KG-based chunk organization process to deliver relevant and important knowledge in well-organized paragraphs. Extensive experiments conducted on the HotpotQA dataset and its variants demonstrate the advantages of KG$^2$RAG compared to existing RAG-based approaches, in terms of both response quality and retrieval quality.

Figures

Figures reproduced from arXiv: 2502.06864 by the authors.

Figure 1
Figure 1. A comparison among LLM-only, Semantic RAG, and Graph RAG paradigms. incorporating it into the prompts of LLMs for re￾sponse generation. Existing studies in RAG (Lewis et al., 2020; Yu, 2022; Purwar and Sundar, 2023; Gao et al., 2023; Ziletti and D’Ambrosi, 2024), as shown in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Workflow of the proposed KG2RAG. the document offline processing aligns with what vanilla RAG does, KG2RAG naturally supports adding new documents to or removing documents from the existing knowledge base and KG effi￾ciently. 2.2 KG-enhanced Chunk Retrieval Given the chunks D and the associated KG G, the proposed KG2RAG suggests a two-stage retrieval process, including semantic-based retrieval and graph-guided expan… view at source ↗
Figure 3
Figure 3. The prompt for triplet extraction. be given as G m q = traverse(G, G 0 q , m), (4) where traverse(·) can be implemented with the breadth-first search (BFS) algorithm, serving as a function that captures all entities in G 0 q , correspond￾ing m-hop neighboring entities, and all edges link￾ing these entities to form an expanded subgraph. Given the expanded subgraph G m q , we can read￾out all the chunks associated wit… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Statistics of triplet extraction. HotpotQA-Full. In the distractor setting, a total of ten documents are provided as supporting ma￾terials, including all useful knowledge as well as some irrelevant content. In the fullwiki setting, it is required to identify useful kno…
Figure 5
Figure 5. Figure 5: Experimental results with varying top-k on HotpotQA in distractor setting. Response Quality Retrieval Quality F1 Precision Recall F1 Precision Recall #Avg. m = 1 0.663 0.690 0.683 0.436 0.301 0.908 8.11 m = 2 0.656 0.681 0.674 0.420 0.291 0.917 8.53 m = 3 0.658 0.678 0…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. KGRAG-Ex: Explainable Retrieval-Augmented Generation with Knowledge Graph-based Perturbations

    cs.LG 2025-07 reject novelty 6.0 of 10

    KGRAG-Ex retrieves answer-relevant paths through a knowledge graph, turns them into natural-language paragraphs, and explains each answer by removing individual graph nodes, edges, or sub-paths and observing whether t...

  2. SCAIR: Schema-Conditioned Agentic Iterative Reasoning for Enterprise Knowledge Graphs

    cs.AI 2026-06 conditional novelty 5.0 of 10

    SCAIR, a training-free schema-conditioned agentic KG-RAG method, substantially outperforms existing KG-RAG approaches on a new enterprise CMDB benchmark, but the evaluation has notable confounds.

  3. Trusted Routing for Blockchain-Empowered UAV Networks via Multi-Agent Deep Reinforcement Learning

    eess.SY 2025-07 unverdicted novelty 5.0 of 10

    A blockchain-based trust management mechanism combined with multi-agent double deep Q-learning reportedly reduces delay in UAV networks with malicious nodes.

  4. A Query-Aware Multi-Path Knowledge Graph Fusion Approach for Enhancing Retrieval-Augmented Generation in Large Language Models

    cs.IR 2025-07 conditional novelty 5.0 of 10

    QMKGF builds multi-path knowledge graph subgraphs from LLM-extracted entities, fuses the highest-scoring subgraph with query-relevant triples, and expands the query to improve RAG answer quality.

Reference graph

Works this paper leans on

52 extracted references · 50 canonical work pages · cited by 4 Pith papers

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Nicholas Ampazis. 2024. Improving RAG quality for large language models with topic-enhanced reranking. In AIAI, volume 712, pages 74--87, Corfu, Greece. Springer

  4. [4]

    Arian Askari, Amin Abolghasemi, Gabriella Pasi, Wessel Kraaij, and Suzan Verberne. 2023. Injecting the BM25 score as text improves BERT -based re-rankers. CoRR

  5. [5]

    S \" o ren Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary G. Ives. 2007. DBpedia : A nucleus for a web of open data. In ISWC, volume 4825, pages 722--735, Busan, Korea. Springer

  6. [6]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert - Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litw...

  7. [7]

    Marco Calamo, Francesca De Luzi, Mattia Macr \` , Tommaso Mencattini, and Massimo Mecella. 2023. CICERO: A GPT 2-based writing assistant to investigate the effectiveness of specialized LLMs' applications in e-justice. In ECAI, volume 372, pages 3196--3203, Kraków, Poland. IOS

  8. [8]

    Wei Cheng, Yuhan Wu, and Wei Hu. 2024. Dataflow-guided retrieval augmentation for repository-level code completion. In ACL, pages 7957--7977, Bangkok, Thailand. ACL

Show all 52 references
  1. [9]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al - Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Z...

  2. [10]

    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. CoRR

  3. [11]

    Matous Eibich, Shivay Nagpal, and Alexander Fred - Ojala. 2024. ARAGOG: A dvanced RAG output grading. CoRR

  4. [12]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat - Seng Chua, and Qing Li. 2024. A survey on RAG meeting LLM s: T owards retrieval-augmented large language models. In KDD, pages 6491--6501, Barcelona, Spain. ACM

  5. [13]

    Luyu Gao, Zhuyun Dai, Tongfei Chen, Zhen Fan, Benjamin Van Durme, and Jamie Callan. 2021. Complement lexical retrieval model with semantic residual embeddings. In ECIR, volume 12656, pages 146--160, Glasgow, UK. Springer

  6. [14]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Qianyu Guo, Meng Wang, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. CoRR

  7. [15]

    Glass, Gaetano Rossiello, Md

    Michael R. Glass, Gaetano Rossiello, Md. Faisal Mahbub Chowdhury, Ankita Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2G : R etrieve, rerank, generate. In NAACL, pages 2701--2715, Seattle, WA, USA. ACL

  8. [16]

    Mandalay Grems. 1962. A survey of languages and systems for information retrieval. Commun. ACM , 5(1):43--46

  9. [17]

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. LightRAG : S imple and fast retrieval-augmented generation. CoRR

  10. [18]

    Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and Philip S. Yu. 2022. A survey on knowledge graphs: R epresentation, acquisition, and applications. IEEE Trans. Neural Networks Learn. Syst. , 33(2):494--514

  11. [19]

    Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi - Yu, Yiming Yang, Jamie Callan, and Graham Neubig

    Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi - Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Active retrieval augmented generation. In EMNLP, pages 7969--7992, Singapore. ACL

  12. [20]

    Yishan Jiao. 2006. Maintaining stream statistics over multiscale sliding windows. ACM Trans. Database Syst. , 31(4):1305--1334

  13. [21]

    Weld, and Luke Zettlemoyer

    Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. TriviaQA : A large scale distantly supervised challenge dataset for reading comprehension. In ACL, pages 1601--1611, Vancouver, Canada. ACL

  14. [22]

    Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. CoRR

  15. [23]

    Tzu - Lin Kuo, Tzu - Wei Chiu, Tzung - Sheng Lin, Sheng - Yang Wu, Chao - Wei Huang, and Yun - Nung Chen. 2024. A survey of generative information retrieval. CoRR

  16. [24]

    u ttler, Mike Lewis, Wen - tau Yih, Tim Rockt \

    Patrick S. H. Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \" u ttler, Mike Lewis, Wen - tau Yih, Tim Rockt \" a schel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP task...

  17. [25]

    Jiawei Li, Yizhe Yang, Yu Bai, Xiaofeng Zhou, Yinghao Li, Huashan Sun, Yuhang Liu, Xingpeng Si, Yuhao Ye, Yixiao Wu, Yiguan Lin, Bin Xu, Ren Bowen, Chong Feng, Yang Gao, and Heyan Huang. 2024. Fundamental capabilities of large language models and their applications in domain s...

  18. [26]

    Xianming Li and Jing Li. 2024. AoE : A ngle-optimized embeddings for semantic textual similarity. In ACL, pages 1825--1839, Bangkok, Thailand. ACL

  19. [27]

    Yinheng Li. 2023. A practical survey on zero-shot prompt design for in-context learning. In RANLP, pages 641--647, Varna, Bulgaria. INCOMA

  20. [28]

    Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. 2024 a . A survey on hallucination in large vision-language models. CoRR

  21. [29]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024 b . Lost in the middle: H ow language models use long contexts. Trans. Assoc. Comput. Linguistics, 12:157--173

  22. [30]

    Yang Liu, Xiaobin Tian, Zequn Sun, and Wei Hu. 2024 c . Finetuning generative large language models with discrimination instructions for knowledge graph completion. In ISWC, Baltimore, MD, USA. Springer

  23. [31]

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting for retrieval-augmented large language models. CoRR

  24. [32]

    Laurent Mombaerts, Terry Ding, Adi Banerjee, Florian Felice, Jonathan Taws, and Tarik Borogovac. 2024. Meta knowledge for retrieval augmented large language models. CoRR

  25. [33]

    Morris, Brandon Duderstadt, and Andriy Mulyar

    Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. 2024. Nomic embed: T raining a reproducible long context text embedder. CoRR

  26. [34]

    Anupam Purwar and Rahul Sundar. 2023. Keyword augmented retrieval: N ovel framework for information retrieval integrated with speech interface. In AIMLSystems, pages 58:1--58:5, Bangalore, India. ACM

  27. [35]

    Chawla, and Chao Huang

    Xubin Ren, Jiabin Tang, Dawei Yin, Nitesh V. Chawla, and Chao Huang. 2024. A survey of large language models for graphs. In KDD, pages 6616--6626, Barcelona, Spain. ACM

  28. [36]

    Priyanka Sen, Sandeep Mavadia, and Amir Saffari. 2023. Knowledge graph-augmented language models for complex question answering. In NLRSE, pages 1--8, Toronto, Canada. ACL

  29. [37]

    Alon Talmor and Jonathan Berant. 2018. The web as a knowledge-base for answering complex questions. In NAACL, pages 641--651, New Orleans, Louisiana, USA. ACL

  30. [38]

    Xiaobin Tian, Zequn Sun, and Wei Hu. 2024. Generating explanations to understand and repair embedding-based entity alignment. In ICDE, pages 2205--2217, Utrecht, Netherlands. IEEE

  31. [39]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie - Anne Lachaux, Timoth \' e e Lacroix, Baptiste Rozi \` e re, Naman Goyal, Eric Hambro, Faisal Azhar, Aur \' e lien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: O pen and effic...

  32. [40]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. MuSiQue : M ultihop questions via single-hop question composition. Trans. Assoc. Comput. Linguistics, 10:539--554

  33. [41]

    Rossi, Alexa F

    Yu Wang, Nedim Lipka, Ryan A. Rossi, Alexa F. Siu, Ruiyi Zhang, and Tyler Derr. 2024 a . Knowledge graph prompting for multi-document question answering. In AAAI, pages 19206--19214, Vancouver, Canada. AAAI

  34. [42]

    Yujing Wang, Hainan Zhang, Liang Pang, Binghui Guo, Hongwei Zheng, and Zhiming Zheng. 2024 b . MaFeRw : Q uery rewriting with multi-aspect feedbacks for retrieval-augmented large language models. CoRR

  35. [43]

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. 2023. AutoGen : E nabling next-gen LLM applications via multi-agent conversation framework. CoRR

  36. [44]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack : P ackaged resources to advance general Chinese embedding. CoRR

  37. [45]

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. 2024 a . Retrieval-augmented generation with knowledge graphs for customer service question answering. In SIGIR, pages 2905--2909, Washington DC, USA. ACM

  38. [46]

    Kankanhalli

    Ziwei Xu, Sanjay Jain, and Mohan S. Kankanhalli. 2024 b . Hallucination is inevitable: A n innate limitation of large language models. CoRR

  39. [47]

    Cohen, Ruslan Salakhutdinov, and Christopher D

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA : A dataset for diverse, explainable multi-hop question answering. In EMNLP, pages 2369--2380, Brussels, Belgium. ACL

  40. [48]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. ReAct : S ynergizing reasoning and acting in language models. In ICLR , Kigali, Rwanda. OpenReview.net

  41. [49]

    Wen - tau Yih, Matthew Richardson, Christopher Meek, Ming - Wei Chang, and Jina Suh. 2016. The value of semantic parse labeling for knowledge base question answering. In ACL, Berlin, Germany. ACL

  42. [50]

    Wenhao Yu. 2022. Retrieval-augmented generation across heterogeneous knowledge. In NAACL, pages 52--58, Seattle, WA, USA. ACL

  43. [51]

    Wenzheng Zhao, Yuanning Cui, and Wei Hu. 2023. Improving continual relation extraction by distinguishing analogous semantics. In ACL, pages 1162--1175, Toronto, Canada. ACL

  44. [52]

    Angelo Ziletti and Leonardo D'Ambrosi. 2024. Retrieval augmented text-to- SQL generation for epidemiological question answering using electronic health records. In NAACL, pages 47--53, Mexico City, Mexico. ACL

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.