Pith. sign in

REVIEW 5 major objections 5 minor 70 references

Learning Efficient and Generalizable Graph Retriever for Knowledge-Graph Question Answering

T0 review · 5 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 The paper claims a lightweight graph retriever trained on rationalized path labels and line graph representations can outperform state-of-the-art KGQA pipelines by 2.66%–20.34% and shrink the gap between weak and strong LLM reasoners.

desk verdict Solid new method for KGQA graph retrieval, but the headline gains are not established because the budget selection and missing error bars leave the main comparison underdetermined. read the letter →

arxiv 2506.09645 v1 pith:U2VXSDRH submitted 2025-06-11 cs.CL cs.IRcs.LG

Tianjun Yao , Haoxuan Li , Zhiqiang Shen , Pan Li , Tongliang Liu , Kun Zhang This is my paper
classification cs.CLcs.IRcs.LG
keywords knowledgegraphquestionansweringretrieverretrieval-augmentedgenerationlinetransformationpath-basedreasoningrationalizedsupervisionlargelanguagemodelsneuralnetworks
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

RAPL is a graph retriever for knowledge-graph question answering that is designed to retrieve compact, causally grounded reasoning paths rather than noisy shortest paths. It generates candidate paths near the shortest path, asks GPT-4o to pick the rational ones, then trains a small GNN on those labels after transforming each triple into a node of a directed line graph. The authors report gains of 2.66%–20.34% over strong baselines when paired with moderately sized reasoners like GPT-4o-mini and Llama3.1-8B, and they show the gap between a small reasoner and GPT-4o shrinks dramatically. If these results hold, cheap retrievers could supply structured evidence that makes weaker LLMs nearly as reliable as much larger models.

What carries the argument

The directed line graph transformation turns each triple into a node and connects consecutive triples while preserving directionality, so reasoning paths in the original graph correspond one-to-one to paths in the line graph. This lets path-based reasoning become node-sequence prediction, lets a standard GCN refine triple-level representations, and avoids over-reliance on unknown-entity embeddings. The two-stage label rationalizer, which bounds candidate paths by shortest-path distance and then uses GPT-4o to select the rational ones, is what injects causal grounding into the supervision; the STOP node and look-ahead bidirectional message passing are what let the retriever learn when to stop and to use future context when choosing the next triple.

What would settle it

Compute the recall of human-annotated rational paths within the [dmin, dmin+2] candidate window on a benchmark with ground-truth rationales; if a substantial share of correct chains are longer, RAPL's supervision is incomplete by construction. Concretely, measure performance on a curated set of questions whose ground-truth reasoning distance exceeds dmin+2, where the method should fail if the window is the binding constraint.

Watch

Extended reading notes

Core claim

The central claim is that a lightweight, parameter-efficient graph retriever can outperform much heavier retrieval pipelines if it learns from causally grounded path supervision on a line-graph representation. RAPL's labeling pipeline first generates candidate paths within a bounded length window around the shortest path, then uses GPT-4o to select the rational paths, replacing the pure shortest-path heuristic with an LLM-rationalized teacher. The retriever itself is a bidirectional GCN operating on the directed line graph of the subgraph, where each triple is a node, so message passing captures both intra-triple and inter-triple semantics while preserving path directionality. Training on path-based objectives with a STOP node lets the retriever decide when to stop; at inference it returns compact paths the LLM reasoner can consume directly. The paper argues these design choices jointly explain the reported gains and the reduced gaps between weak and strong reasoners and across datasets.

Load-bearing premise

The candidate path set is restricted to paths whose length is within two hops of the shortest path distance, so any correct reasoning chain longer than that is never a candidate and cannot be retrieved.

Editorial extensions

If this is right

  • Paired with Llama3.1-8B, RAPL reduces the Macro-F1 gap to GPT-4o from 14.78% for SubgraphRAG to 2.22% on CWQ.
  • Trained on WebQSP, RAPL drops only 3.1% Macro-F1 on CWQ versus 17.4% for SubgraphRAG, showing better cross-dataset generalization.
  • Path-structured retrieval inputs help reasoning: shuffling retrieved triples into an unordered bag lowers Macro-F1 for both GPT-4o and GPT-4o-mini.
  • The largest relative gains appear on ≥3-hop CWQ questions, where RAPL improves Macro-F1 by roughly 40% over SubgraphRAG when paired with Llama3.1-8B.
  • The retriever stays efficient, retrieving about 25–53 triples on WebQSP/CWQ with one LLM call at inference time.

Reading between the lines

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

  • The [dmin, dmin+2] candidate window is the hidden capacity limit: if correct reasoning chains frequently exceed two hops beyond the shortest path, RAPL cannot retrieve them regardless of training quality; an easy test is expanding the window and measuring recall on longer-hop subsets.
  • Because GPT-4o is the labeler, the supervision ceiling is set by the annotator's reasoning ability; the ablation showing GPT-4o-mini labels perform no better than shortest-path heuristics suggests the method's gains may shrink with a weaker annotator.
  • The line-graph and path-based machinery is task-agnostic, so the same recipe could transfer to other retrieval-heavy structured tasks such as multi-hop fact verification or entity linking where rational subgraph selection matters.
  • A hybrid retriever that combines the graph retriever's efficiency with an LLM retriever's flexibility, a direction the paper's limitations section acknowledges, would be the natural next step if the goal is to keep high recall on long chains without losing RAPL's compactness.
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

5 major / 5 minor

Summary. The paper proposes RAPL, a lightweight GNN-based graph retriever for knowledge-graph question answering. The method has three main components: (1) a two-stage labeling strategy in which a heuristic enumerates candidate paths within [dmin, dmin+2] and GPT-4o selects 'rational' paths as supervision; (2) a line-graph transformation that converts each triple into a node so that GNN message passing captures intra- and inter-triple interactions; and (3) a path-based training objective with bidirectional ('look-ahead') GCN encoders and a dedicated STOP node. Experiments are reported on WebQSP and CWQ with three retrieval budgets and three reasoners (Llama3.1-8B, GPT-4o-mini, GPT-4o). The paper claims improvements of 2.66%-20.34% over state-of-the-art methods, reduced gaps between weak and strong reasoners, and smaller cross-dataset performance drops, together with code release and qualitative examples.

Significance. If the reported results hold, RAPL addresses a real bottleneck in retrieve-then-reason KGQA: the quality of supervision for graph-based retrievers and their generalization to unseen questions. The line-graph transformation is a clean, model-agnostic idea, and the paper provides useful supporting material: an algorithmic pseudo-code, a theoretical bijection between paths in the original graph and the line graph, hop-level breakdowns, efficiency comparisons, and qualitative retrieval examples. The code release is another strength. However, the current empirical evaluation has several protocol-level weaknesses that make the central quantitative claim of superiority over existing methods not yet established.

major comments (5)
  1. [§5.1, Table 1; Abstract] The headline improvement range of 2.66%-20.34% is not tied to a single reproducible system. Table 1 reports three retrieval budgets (K,M) = (60,80), (80,120), and (120,200), and the best result per dataset/reasoner is selected from those configurations. The lower and upper bounds of the claimed range come from different configurations and different reasoners, so the range does not describe one setting. No standard deviations, multiple seeds, or significance tests are reported for any of the numbers. I ask the authors to report per-configuration results with variance across seeds, and to either pre-specify the budget-selection rule or apply a multiple-comparison correction before claiming the 2.66%-20.34% range.
  2. [§4.1, Table 2] There is a labeler-reasoner overlap confound. GPT-4o generates the rational path labels in Section 4.1, and GPT-4o or GPT-4o-mini are the main downstream reasoners in the evaluation. Table 2 shows that GPT-4o labels outperform GPT-4o-mini and shortest-path labels, but all labelers are from the same API family, so the retriever may be specialized to GPT-4o-family reasoning preferences rather than to 'rational' paths in general. The paper should include a labeler from a different model family (e.g., a strong open-weight model) and/or evaluate with a non-OpenAI reasoner to disentangle this effect.
  3. [§4.3, Algorithm 1, Eq. (5)-(6)] The STOP node is introduced to determine when to terminate reasoning, but its role in training and inference is not fully specified. Equation (5) defines its representation, yet Eq. (6) does not include STOP in the softmax denominator, and Algorithm 1 never mentions STOP. Appendix I credits RAPL's compact retrieval to learning when to stop, but the paper does not state how STOP is added to the candidate set, how it competes with graph nodes, or how the stopping criterion is applied during sampling. Please specify this mechanism precisely, otherwise the efficiency claim and the path-termination behavior are not reproducible.
  4. [§4.2, Table 3] The contribution of the line-graph transformation is not directly ablated. Table 3 compares a 1-layer GCN, a 2-layer GCN without look-ahead, and a 2-layer GCN with look-ahead, but all conditions appear to operate on the line graph. Since the line-graph transformation is one of the paper's three main design contributions and is claimed to enhance representational capacity, the paper should include an ablation that trains the same GCN on the original graph (without the transformation) to isolate this design choice. Without that condition, the empirical support for the line-graph claim is missing.
  5. [§4.1] The candidate path window [dmin, dmin+2] is an untested assumption. If GPT-4o's rational paths sometimes exceed dmin+2, they never appear in the candidate set and cannot be selected, so the supervision is biased toward shorter paths. The paper does not report the length distribution of the rational paths, nor does it test sensitivity to the window offset. I ask for a quantitative analysis (e.g., the proportion of training examples whose selected rational paths hit the window boundary) and, if feasible, an experiment with a larger window offset such as [dmin, dmin+3].
minor comments (5)
  1. [§5.5] The text refers to 'Table 3b', but the generalization results appear to be in Figure 3(b); the cross-reference should be fixed.
  2. [Appendix D, Appendix I] There are typos: 'upadte' should be 'update' in Appendix D, and 'effiency' should be 'efficiency' in Appendix I.
  3. [Table 1 caption] The caption says red, blue, and green shading indicate the best-performing result within each retrieval configuration, but the configurations are not explicitly labeled with those colors in the table body; please clarify which color corresponds to which (K,M) setting.
  4. [§5.2] The claim that 'for GPT-4o, RAPL slightly underperforms SubgraphRAG and GNN-RAG' is not accurate on WebQSP, where Ours+GPT-4o achieves 80.7 Macro-F1 versus 76.4 for SubgraphRAG+GPT-4o and 71.3 for GNN-RAG; the sentence should be qualified to the datasets or metrics where underperformance occurs.
  5. [Abstract and Conclusion] The phrase 'causally grounded supervision' overstates what the labels are: the retriever is trained on GPT-4o rationales, which are LLM preferences rather than verified causal ground truth. Suggest using 'rationalized supervision' or 'LLM-rationalized labels' to avoid overclaiming.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: RAPL's retrieval training and evaluation form an empirical pipeline with external benchmark answers, not a definitional or self-citation chain.

full rationale

I walked the claimed derivation chain: candidate paths are constrained to [dmin, dmin+2] (Sec. 4.1); GPT-4o annotates rational paths; a line-graph GNN is trained with path and question-triple losses (Sec. 4.3); retrieved paths are passed to GPT-4o, GPT-4o-mini, or Llama3.1-8B; answers are compared to benchmark ground truth. No equation or algorithmic step makes a predicted output equal to a fitted input by construction. The labeler (GPT-4o) and some downstream reasoners (GPT-4o/GPT-4o-mini) are from the same model family, which is an external-validity/independence concern, but it is not a formal circularity: the retriever is evaluated on held-out questions with fixed answers, and Table 2 shows GPT-4o labels also improve Llama3.1-8B, while Sec. 5.5 reports cross-dataset training (WebQSP→CWQ) with relatively small drops. The [dmin, dmin+2] window could exclude longer rational paths, but this is an unvalidated coverage assumption, not a reduction of the prediction to the input. The headline '2.66%–20.34%' aggregates the best of three retrieval budgets without error bars; this is a statistical-reporting / multiple-comparisons issue, not circularity. Self-citations sharing author Pan Li (e.g., [34], [35], [69]) are used as baseline or related-work citations and are not load-bearing for RAPL's central claim. Therefore there is no step where Eq. X = Eq. Y by construction or where a fitted parameter is renamed as a prediction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The ledger shows that the central claim rests mainly on the assumption that the LLM annotator provides correct rationales and that the shortest-path window contains those rationales. The STOP node is an architectural device, not a new empirical entity. Retrieval budgets and loss weights are hand-chosen hyperparameters that affect reported magnitudes.

free parameters (3)
  • Candidate path length window offset = 2
    Section 4.1 constrains candidate paths to [dmin, dmin+2]; chosen by hand, and if rational paths are longer they are excluded from supervision.
  • Retrieval budget (K, M) = e.g., (60,80), (80,120), (120,200); cross-dataset (120,180), (800,1200)
    Inference hyperparameters; the paper reports three configurations and highlights the best results, so the headline gains depend on configuration choice.
  • Loss weights lambda_q, lambda_path = 1.0
    Algorithm 1 sets both default to 1.0; no sensitivity analysis is provided.
assumptions (4)
  • domain assumption Question entity linking is correct: eq is identified and linked in Gq.
    Stated in Section 2: 'we assume that eq are correctly identified and linked in Gq'. If linking fails, retrieval and answer fail.
  • domain assumption GPT-4o's rational path selection is causally grounded and correct.
    The two-stage labeling in Section 4.1 delegates ground truth to LLM judgment; incorrect or biased labels propagate to the retriever.
  • domain assumption The LLM reasoner used for answer verification, GPT-4o-mini, judges correctness reliably.
    Hits and Macro-F1 include model-based verification as described in Appendix H and L; a biased verifier would systematically alter scores.
  • standard math Line graph transformation properties from Propositions 1 and 2 hold on Freebase subgraphs.
    The propositions are proven for directed graphs, but Freebase has multi-edges and loops; the paper does not analyze these cases.
invented entities (1)
  • STOP node in the path decoder
    purpose: Learned termination signal during path-based reasoning, deciding when to stop extending a path (Section 4.3, Eq. 5).
    Model-internal construct, not a postulated physical or external entity; no independent evidence outside the paper's algorithm.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Efficient and Generalizable Graph Retriever for Knowledge-Graph Question Answering." pith.science (2026). https://pith.science/paper/U2VXSDRH

@misc{pith2026250609645,
  author       = {Pith},
  title        = {Pith review of: Learning Efficient and Generalizable Graph Retriever for Knowledge-Graph Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U2VXSDRH}},
  note         = {Machine review of arXiv:2506.09645}
}
abstract

Large Language Models (LLMs) have shown strong inductive reasoning ability across various domains, but their reliability is hindered by the outdated knowledge and hallucinations. Retrieval-Augmented Generation mitigates these issues by grounding LLMs with external knowledge; however, most existing RAG pipelines rely on unstructured text, limiting interpretability and structured reasoning. Knowledge graphs, which represent facts as relational triples, offer a more structured and compact alternative. Recent studies have explored integrating knowledge graphs with LLMs for knowledge graph question answering (KGQA), with a significant proportion adopting the retrieve-then-reasoning paradigm. In this framework, graph-based retrievers have demonstrated strong empirical performance, yet they still face challenges in generalization ability. In this work, we propose RAPL, a novel framework for efficient and effective graph retrieval in KGQA. RAPL addresses these limitations through three aspects: (1) a two-stage labeling strategy that combines heuristic signals with parametric models to provide causally grounded supervision; (2) a model-agnostic graph transformation approach to capture both intra- and inter-triple interactions, thereby enhancing representational capacity; and (3) a path-based reasoning strategy that facilitates learning from the injected rational knowledge, and supports downstream reasoner through structured inputs. Empirically, RAPL outperforms state-of-the-art methods by $2.66\%-20.34\%$, and significantly reduces the performance gap between smaller and more powerful LLM-based reasoners, as well as the gap under cross-dataset settings, highlighting its superior retrieval capability and generalizability. Codes are available at: https://github.com/tianyao-aka/RAPL.

Figures

Figures reproduced from arXiv: 2506.09645 by the authors.

Figure 1
Figure 1. Overall framework of RAPL. The generalization ability of RAPL arises from the label rationalizer, line graph transformation, and the path-based reasoning paradigm. 3 Related Work We discuss the relevant literature on the retrieval-then-reasoning paradigm and knowledge graph￾based agentic RAG in detail in Appendix B. 4 Empowering Graph Retrievers with Enhanced Generalization for KGQA In this section, we discuss how t… view at source ↗
Figure 2
Figure 2. Impact of path-formatted in￾puts on reasoning performance. ❸ Finally, for 2-hop samples in the CWQ dataset, we ob￾serve that when using Llama3.1-8B as the reasoner, RAPL outperforms SubgraphRAG. However, when paired with GPT-4o-mini, RAPL underperforms in terms of Macro-F1, despite retrieving the same set of triples. This suggests that GPT-4o-mini possesses stronger denoising capabil￾ities and is better able to reor… view at source ↗
Figure 3
Figure 3. Illustration of faithfulness and generalization performance. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (26 more)
Figure 4
Figure 4. Figure 4: Motivating example on not all shortest paths are rational paths. [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 16-25
Figure 16-25. Figure 16-25: L Prompt Template We provide the prompt template in this section, for rational paths identification, relation targeting, and hallucination detection, as shown in [PITH_FULL_IMAGE:figures/full_fig_p019_16-25.png]
Figure 5
Figure 5. Figure 5: Motivating example on not all shortest paths are rational paths. [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Motivating example on not all shortest paths are rational paths. [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Motivating example on not all shortest paths are rational paths. [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Motivating example on not all shortest paths are rational paths. [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: Motivating example on not all shortest paths are rational paths. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Motivating example on shortest paths may not be rational paths. [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Motivating example on shortest paths may not be rational paths. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Motivating example on shortest paths may not be rational paths. [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]
Figure 13
Figure 13. Figure 13: Motivating example on shortest paths may not be rational paths. [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]
Figure 14
Figure 14. Figure 14: Motivating example on shortest paths may not be rational paths. [PITH_FULL_IMAGE:figures/full_fig_p025_14.png]
Figure 15
Figure 15. Figure 15: Motivating example on shortest paths may not be rational paths. [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p026_16.png]
Figure 17
Figure 17. Figure 17: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p026_17.png]
Figure 18
Figure 18. Figure 18: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p027_18.png]
Figure 19
Figure 19. Figure 19: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p027_19.png]
Figure 20
Figure 20. Figure 20: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p027_20.png]
Figure 21
Figure 21. Figure 21: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p028_21.png]
Figure 22
Figure 22. Figure 22: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p028_22.png]
Figure 23
Figure 23. Figure 23: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p029_23.png]
Figure 24
Figure 24. Figure 24: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p030_24.png]
Figure 25
Figure 25. Figure 25: Example on the retrieved reasoning paths by [PITH_FULL_IMAGE:figures/full_fig_p030_25.png]
Figure 26
Figure 26. Figure 26: Prompt template for retrieving rational reasoning paths. [PITH_FULL_IMAGE:figures/full_fig_p031_26.png]
Figure 27
Figure 27. Figure 27: Prompt template for potential relation targeting. [PITH_FULL_IMAGE:figures/full_fig_p031_27.png]
Figure 28
Figure 28. Figure 28: Prompt template for hallucination detection. [PITH_FULL_IMAGE:figures/full_fig_p032_28.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 28 canonical work pages

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Semantic parsing on freebase from question-answer pairs

    Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. Semantic parsing on freebase from question-answer pairs. InProceedings of the 2013 conference on empirical methods in natural language processing, pages 1533–1544, 2013

    2013

  3. [3]

    Freebase: a collaboratively created graph database for structuring human knowledge

    Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase: a collaboratively created graph database for structuring human knowledge. InProceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 1247–1250, 2008

    2008

  4. [4]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877–1901, 2020

    1901

  5. [5]

    Sparks of artificial general intelligence: Early experiments with gpt-4.arXiv preprint arXiv:2303.12712, 2023

    Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with gpt-4.arXiv preprint arXiv:2303.12712, 2023

    arXiv 2023

  6. [6]

    Springer Science & Business Media, 2008

    Michel Chein and Marie-Laure Mugnier.Graph-based knowledge representation: computational founda- tions of conceptual graphs. Springer Science & Business Media, 2008

    work page 2008

  7. [7]

    Plan-on-graph: Self-correcting adaptive planning of large language model on knowledge graphs.arXiv preprint arXiv:2410.23875, 2024

    Liyi Chen, Panrong Tong, Zhongming Jin, Ying Sun, Jieping Ye, and Hui Xiong. Plan-on-graph: Self-correcting adaptive planning of large language model on knowledge graphs.arXiv preprint arXiv:2410.23875, 2024

    arXiv 2024

  8. [8]

    EWEK-QA : Enhanced web and efficient knowledge graph retrieval for citation-based question answering systems

    Mohammad Dehghan, Mohammad Alomrani, Sunyam Bagga, David Alfonso-Hermelo, Khalil Bibi, Abbas Ghaddar, Yingxue Zhang, Xiaoguang Li, Jianye Hao, Qun Liu, Jimmy Lin, Boxing Chen, Prasanna Parthasarathi, Mahdi Biparva, and Mehdi Rezagholizadeh. EWEK-QA : Enhanced web and efficient knowledge graph retrieval for citation-based question answering systems. In Lun...

    work page 2024

Show all 70 references
  1. [9]

    Hardmath: A benchmark dataset for challenging problems in applied mathematics.arXiv preprint arXiv:2410.09988, 2024

    Jingxuan Fan, Sarah Martinson, Erik Y Wang, Kaylie Hausknecht, Jonah Brenner, Danxian Liu, Nianli Peng, Corey Wang, and Michael P Brenner. Hardmath: A benchmark dataset for challenging problems in applied mathematics.arXiv preprint arXiv:2410.09988, 2024

    arXiv

  2. [10]

    Fast graph representation learning with pytorch geometric.arXiv preprint arXiv:1903.02428, 2019

    Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geometric.arXiv preprint arXiv:1903.02428, 2019

    1903 arXiv

  3. [11]

    Towards foundation models for knowledge graph reasoning.arXiv preprint arXiv:2310.04562, 2023

    Mikhail Galkin, Xinyu Yuan, Hesham Mostafa, Jian Tang, and Zhaocheng Zhu. Towards foundation models for knowledge graph reasoning.arXiv preprint arXiv:2310.04562, 2023

    2023 arXiv

  4. [12]

    Double permutation equivariance for knowledge graph completion.arXiv preprint arXiv:2302.01313, 2023

    Jianfei Gao, Yangze Zhou, and Bruno Ribeiro. Double permutation equivariance for knowledge graph completion.arXiv preprint arXiv:2302.01313, 2023

    2023 arXiv

  5. [13]

    Two-stage gen- erative question answering on temporal knowledge graph using large language models.arXiv preprint arXiv:2402.16568, 2024

    Yifu Gao, Linbo Qiao, Zhigang Kan, Zhihua Wen, Yongquan He, and Dongsheng Li. Two-stage gen- erative question answering on temporal knowledge graph using large language models.arXiv preprint arXiv:2402.16568, 2024

    2024 arXiv

  6. [14]

    Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2024

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2024

    2024 arXiv

  7. [15]

    Rela- tional message passing for fully inductive knowledge graph completion

    Yuxia Geng, Jiaoyan Chen, Jeff Z Pan, Mingyang Chen, Song Jiang, Wen Zhang, and Huajun Chen. Rela- tional message passing for fully inductive knowledge graph completion. In2023 IEEE 39th international conference on data engineering (ICDE), pages 1221–1233. IEEE, 2023

    2023

  8. [16]

    Exploring network structure, dynamics, and function using networkx

    Aric Hagberg, Pieter J Swart, and Daniel A Schult. Exploring network structure, dynamics, and function using networkx. Technical report, Los Alamos National Laboratory (LANL), Los Alamos, NM (United States), 2008

    2008

  9. [17]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 10

    2017

  10. [18]

    G-retriever: Retrieval-augmented generation for textual graph understanding and question answering.Advances in Neural Information Processing Systems, 37:132876–132907, 2024

    Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering.Advances in Neural Information Processing Systems, 37:132876–132907, 2024

    2024

  11. [19]

    Knowledge graphs

    Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia d’Amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, et al. Knowledge graphs. ACM Computing Surveys (Csur), 54(4):1–37, 2021

    2021

  12. [20]

    Towards reasoning in large language models: A survey

    Jie Huang and Kevin Chen-Chuan Chang. Towards reasoning in large language models: A survey. In Findings of the Association for Computational Linguistics: ACL 2023, pages 1049–1065, 2023

    2023

  13. [21]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232, 2023

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:231...

    2023 arXiv

  14. [22]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. InInternational conference on machine learning, pages 448–456. pmlr, 2015

    2015

  15. [23]

    Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12), 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12), 2023

    2023

  16. [24]

    Structgpt: A general framework for large language model to reason over structured data

    Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Wayne Xin Zhao, and Ji-Rong Wen. Structgpt: A general framework for large language model to reason over structured data. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9237–9251, 2023

    2023

  17. [25]

    Kg-agent: An efficient autonomous agent framework for complex reasoning over knowledge graph.arXiv preprint arXiv:2402.11163, 2024

    Jinhao Jiang, Kun Zhou, Wayne Xin Zhao, Yang Song, Chen Zhu, Hengshu Zhu, and Ji-Rong Wen. Kg-agent: An efficient autonomous agent framework for complex reasoning over knowledge graph.arXiv preprint arXiv:2402.11163, 2024

    2024 arXiv

  18. [26]

    Unikgqa: Unified retrieval and reasoning for solving multi-hop question answering over knowledge graph

    Jinhao Jiang, Kun Zhou, Xin Zhao, and Ji-Rong Wen. Unikgqa: Unified retrieval and reasoning for solving multi-hop question answering over knowledge graph. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  19. [27]

    Graph chain-of-thought: Augmenting large language models by reasoning on graphs.arXiv preprint arXiv:2404.07103, 2024

    Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al. Graph chain-of-thought: Augmenting large language models by reasoning on graphs.arXiv preprint arXiv:2404.07103, 2024

    2024 arXiv

  20. [28]

    Smith, Yejin Choi, and Kentaro Inui

    Jungo Kasai, Keisuke Sakaguchi, yoichi takahashi, Ronan Le Bras, Akari Asai, Xinyan Velocity Yu, Dragomir Radev, Noah A. Smith, Yejin Choi, and Kentaro Inui. Realtime QA: What’s the answer right now? InThirty-seventh Conference on Neural Information Processing Systems Datasets...

    2023

  21. [29]

    KG-GPT: A general framework for reasoning on knowledge graphs using large language models

    Jiho Kim, Yeonsu Kwon, Yohan Jo, and Edward Choi. KG-GPT: A general framework for reasoning on knowledge graphs using large language models. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9410...

    2023

  22. [30]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  23. [31]

    Semi-supervised classification with graph convolutional networks

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

    2016 arXiv

  24. [32]

    Ingram: Inductive knowledge graph embedding via relation graphs

    Jaejun Lee, Chanyoung Chung, and Joyce Jiyoung Whang. Ingram: Inductive knowledge graph embedding via relation graphs. InInternational Conference on Machine Learning, pages 18796–18809. PMLR, 2023

    2023

  25. [33]

    Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems...

    2020

  26. [34]

    Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation.arXiv preprint arXiv:2410.20724, 2024

    Mufei Li, Siqi Miao, and Pan Li. Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation.arXiv preprint arXiv:2410.20724, 2024

    2024 arXiv

  27. [35]

    Distance encoding: Design provably more powerful neural networks for graph representation learning.Advances in Neural Information Processing Systems, 33:4465–4478, 2020

    Pan Li, Yanbang Wang, Hongwei Wang, and Jure Leskovec. Distance encoding: Design provably more powerful neural networks for graph representation learning.Advances in Neural Information Processing Systems, 33:4465–4478, 2020. 11

    2020

  28. [36]

    Graph reasoning for question answering with triplet retrieval

    Shiyang Li, Yifan Gao, Haoming Jiang, Qingyu Yin, Zheng Li, Xifeng Yan, Chao Zhang, and Bing Yin. Graph reasoning for question answering with triplet retrieval. InFindings of the Association for Computational Linguistics: ACL 2023, pages 3366–3375, 2023

    2023

  29. [37]

    Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281, 2023

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning.arXiv preprint arXiv:2308.03281, 2023

    2023 arXiv

  30. [38]

    Explore then determine: A gnn-llm synergy framework for reasoning over knowledge graph.arXiv preprint arXiv:2406.01145, 2024

    Guangyi Liu, Yongqi Zhang, Yong Li, and Quanming Yao. Explore then determine: A gnn-llm synergy framework for reasoning over knowledge graph.arXiv preprint arXiv:2406.01145, 2024

    2024 arXiv

  31. [39]

    Dual reasoning: A gnn-llm collaborative framework for knowledge graph question answering, 2025

    Guangyi Liu, Yongqi Zhang, Yong Li, and Quanming Yao. Dual reasoning: A gnn-llm collaborative framework for knowledge graph question answering, 2025

    2025

  32. [40]

    Reasoning on graphs: Faithful and interpretable large language model reasoning

    Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. Reasoning on graphs: Faithful and interpretable large language model reasoning. InInternational Conference on Learning Representations, 2024

    2024

  33. [41]

    Think-on-graph 2.0: Deep and interpretable large language model reasoning with knowledge graph-guided retrieval.arXiv e-prints, pages arXiv–2407, 2024

    Shengjie Ma, Chengjin Xu, Xuhui Jiang, Muzhi Li, Huaren Qu, and Jian Guo. Think-on-graph 2.0: Deep and interpretable large language model reasoning with knowledge graph-guided retrieval.arXiv e-prints, pages arXiv–2407, 2024

    2024

  34. [42]

    Automated social science: Language models as scientist and subjects

    Benjamin S Manning, Kehang Zhu, and John J Horton. Automated social science: Language models as scientist and subjects. Technical report, National Bureau of Economic Research, 2024

    2024

  35. [43]

    Rearev: Adaptive reasoning for question answering over knowledge graphs

    Costas Mavromatis and George Karypis. Rearev: Adaptive reasoning for question answering over knowledge graphs. InFindings of the Association for Computational Linguistics: EMNLP 2022, pages 2447–2458, 2022

    2022

  36. [44]

    Gnn-rag: Graph neural retrieval for large language model reasoning.arXiv preprint arXiv:2405.20139, 2024

    Costas Mavromatis and George Karypis. Gnn-rag: Graph neural retrieval for large language model reasoning.arXiv preprint arXiv:2405.20139, 2024

    2024 arXiv

  37. [45]

    Build the future of ai with meta llama 3, 2024

    Meta. Build the future of ai with meta llama 3, 2024

    2024

  38. [46]

    Efficient estimation of word representations in vector space.arXiv preprint arXiv:1301.3781, 2013

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space.arXiv preprint arXiv:1301.3781, 2013

    2013 arXiv

  39. [47]

    Introducing chatgpt, 2022

    OpenAI. Introducing chatgpt, 2022

    2022

  40. [48]

    Hello gpt-4o, 2024

    OpenAI. Hello gpt-4o, 2024

    2024

  41. [49]

    Pytorch: An imperative style, high-performance deep learning library, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

    2019

  42. [50]

    O’Reilly Media, Inc

    Ian Robinson, Jim Webber, and Emil Eifrem.Graph databases: new opportunities for connected data. " O’Reilly Media, Inc.", 2015

    2015

  43. [51]

    Retrieval augmentation reduces hallucination in conversation

    Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval augmentation reduces hallucination in conversation. InFindings of the Association for Computational Linguistics: EMNLP 2021, pages 3784–3803, 2021

    2021

  44. [52]

    Dropout: a simple way to prevent neural networks from overfitting.The journal of machine learning research, 15(1):1929–1958, 2014

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting.The journal of machine learning research, 15(1):1929–1958, 2014

    1929

  45. [53]

    Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph

    Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel Ni, Heung- Yeung Shum, and Jian Guo. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. InThe Twelfth International Conference on Learning Representati...

    2024

  46. [54]

    The web as a knowledge-base for answering complex questions

    Alon Talmor and Jonathan Berant. The web as a knowledge-base for answering complex questions. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 641–651, 2018

    2018

  47. [55]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  48. [56]

    Graph attention networks.arXiv preprint arXiv:1710.10903, 2017

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks.arXiv preprint arXiv:1710.10903, 2017. 12

    2017 arXiv

  49. [57]

    Graph attention networks.stat, 1050(20):10–48550, 2017

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. Graph attention networks.stat, 1050(20):10–48550, 2017

    2017

  50. [58]

    Knowledge-driven cot: Exploring faithful reasoning in llms for knowledge-intensive question answering.arXiv preprint arXiv:2308.13259, 2023

    Keheng Wang, Feiyu Duan, Sirui Wang, Peiguang Li, Yunsen Xian, Chuantao Yin, Wenge Rong, and Zhang Xiong. Knowledge-driven cot: Exploring faithful reasoning in llms for knowledge-intensive question answering.arXiv preprint arXiv:2308.13259, 2023

    2023 arXiv

  51. [59]

    Knowledge graph prompting for multi-document question answering

    Yu Wang, Nedim Lipka, Ryan A Rossi, Alexa Siu, Ruiyi Zhang, and Tyler Derr. Knowledge graph prompting for multi-document question answering. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19206–19214, 2024

    2024

  52. [60]

    Chi, Quoc V Le, and Denny Zhou

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems, 2022

    2022

  53. [61]

    Mindmap: Knowledge graph prompting sparks graph of thoughts in large language models.arXiv preprint arXiv:2308.09729, 2023

    Yilin Wen, Zifeng Wang, and Jimeng Sun. Mindmap: Knowledge graph prompting sparks graph of thoughts in large language models.arXiv preprint arXiv:2308.09729, 2023

    2023 arXiv

  54. [62]

    A survey on large language models for recommendation.World Wide Web, 27(5):60, 2024

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. A survey on large language models for recommendation.World Wide Web, 27(5):60, 2024

    2024

  55. [63]

    Retrieve-rewrite-answer: A kg-to-text enhanced llms framework for knowledge graph question answering, 2023

    Yike Wu, Nan Hu, Sheng Bi, Guilin Qi, Jie Ren, Anhuan Xie, and Wei Song. Retrieve-rewrite-answer: A kg-to-text enhanced llms framework for knowledge graph question answering, 2023

    2023

  56. [64]

    How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InInternational Conference on Learning Representations, 2019

    2019

  57. [65]

    Qwen2 technical report.arXiv preprint arXiv:2407.10671, 2024

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng...

    2024 arXiv

  58. [66]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in Neural Information Processing Systems, 36, 2024

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in Neural Information Processing Systems, 36, 2024

    2024

  59. [67]

    The value of semantic parse labeling for knowledge base question answering

    Wen-tau Yih, Matthew Richardson, Christopher Meek, Ming-Wei Chang, and Jina Suh. The value of semantic parse labeling for knowledge base question answering. InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), page...

    2016

  60. [68]

    Subgraph retrieval enhanced model for multi-hop knowledge base question answering

    Jing Zhang, Xiaokang Zhang, Jifan Yu, Jian Tang, Jie Tang, Cuiping Li, and Hong Chen. Subgraph retrieval enhanced model for multi-hop knowledge base question answering. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    2022

  61. [69]

    Labeling trick: A theory of using graph neural networks for multi-node representation learning.Advances in Neural Information Processing Systems, 34:9061–9073, 2021

    Muhan Zhang, Pan Li, Yinglong Xia, Kai Wang, and Long Jin. Labeling trick: A theory of using graph neural networks for multi-node representation learning.Advances in Neural Information Processing Systems, 34:9061–9073, 2021

    2021

  62. [70]

    close to Russia

    Jincheng Zhou, Beatrice Bevilacqua, and Bruno Ribeiro. An ood multi-task perspective for link prediction with new relation types and nodes.arXiv preprint arXiv:2307.06046, 23, 2023. 13 Appendix Contents A Broad Impact 15 B Related Work 15 C Complexity Analysis 15 D More Discus...

    2023 arXiv

Pith tools

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