pith. machine review for the scientific record. sign in

arxiv: 2604.24623 · v1 · submitted 2026-04-27 · 💻 cs.AI · cs.IR· cs.LG

Recognition: unknown

XGRAG: A Graph-Native Framework for Explaining KG-based Retrieval-Augmented Generation

Ha Linh Hong Tran Nguyen, Maxim Romanovsky, Valeria Bladinieres, Zhuoling Li

Pith reviewed 2026-05-08 03:23 UTC · model grok-4.3

classification 💻 cs.AI cs.IRcs.LG
keywords explainable AIknowledge graphsretrieval-augmented generationgraph perturbationscausal attributionsLLM interpretabilityGraphRAG
0
0 comments X

The pith

Graph perturbations attribute model answers to specific components in knowledge graphs for retrieval-augmented generation.

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

The paper introduces a framework that generates explanations for systems using knowledge graphs to augment language model retrieval. It works by applying targeted changes to the graph structure and measuring how each change shifts the final answer produced by the model. A sympathetic reader would care because existing explanation techniques for retrieval systems operate only on unstructured text and therefore cannot isolate the role of connected facts and relations. If the approach holds, it would allow users to see which pieces of structured knowledge actually drive a given response rather than treating the entire graph as an opaque input.

Core claim

XGRAG generates causally grounded explanations for GraphRAG systems by employing graph-based perturbation strategies to quantify the contribution of individual graph components on the model answer. These explanations align with the original answers produced by the system and capture the structural properties of the underlying knowledge graph.

What carries the argument

Graph-based perturbation strategies, which systematically alter parts of the knowledge graph and observe resulting changes in the generated answer to measure each component's influence.

If this is right

  • Explanations become available that reflect the relational structure among knowledge components rather than treating retrieved text as a flat collection.
  • The same perturbation approach can be applied across different question types, narrative datasets, and underlying language models.
  • Generated explanations will tend to highlight elements that occupy central positions in the knowledge graph.
  • The framework scales to larger graphs because it operates directly on the graph representation instead of requiring separate text-based analysis.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the perturbations isolate genuine causal roles, the method could be used to diagnose why a GraphRAG system gives an incorrect answer by pointing to the specific graph links responsible.
  • The technique opens a route to auditing knowledge graphs for systematic biases that consistently steer model outputs in particular directions.
  • One testable extension would be to apply the same perturbations during training or fine-tuning so the model learns to rely more on high-contribution graph elements.

Load-bearing premise

That altering parts of the graph and tracking shifts in the answer truly identifies causal contributions rather than only correlations.

What would settle it

Manually remove or alter one fact known to be required for a correct answer, then check whether the explanation method flags that fact as important precisely when the answer changes.

Figures

Figures reproduced from arXiv: 2604.24623 by Ha Linh Hong Tran Nguyen, Maxim Romanovsky, Valeria Bladinieres, Zhuoling Li.

Figure 1
Figure 1. Figure 1: XGRAG vs. XAI framework for text-based RAG. Standard approaches (bottom) perturb unstructured text retrieved from a vector store to assess importance. Our framework XGRAG (top) operates on a KG, perturbing subgraphs to identify the key graph components for the LLM’s answer. Our key contributions are: (1) a novel XAI framework using graph-based perturbation to quantify the influence of graph components on L… view at source ↗
Figure 2
Figure 2. Figure 2: The XGRAG architecture. The GraphRAG backbone retrieves a subgraph, which is then view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison on questions with different cognitive levels. These results com view at source ↗
Figure 4
Figure 4. Figure 4: Performance comparison of XGRAG against baseline RAG-Ex across different narra view at source ↗
Figure 5
Figure 5. Figure 5: Breakdown of correlation strengths for statistically significant results (p view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative analysis for the query about Della’s gift. The top image visualizes the retrieved view at source ↗
Figure 7
Figure 7. Figure 7: Impact of generation temperature on explanation accuracy. A low temperature (0.1) im view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity analysis of similarity threshold ( view at source ↗
read the original abstract

Graph-based Retrieval-Augmented Generation (GraphRAG) extends traditional RAG by using knowledge graphs (KGs) to give large language models (LLMs) a structured, semantically coherent context, yielding more grounded answers. However, GraphRAG reasoning process remains a black-box, limiting our ability to understand how specific pieces of structured knowledge influence the final output. Existing explainability (XAI) methods for RAG systems, designed for text-based retrieval, are limited to interpreting an LLM response through the relational structures among knowledge components, creating a critical gap in transparency and trustworthiness. To address this, we introduce XGRAG, a novel framework that generates causally grounded explanations for GraphRAG systems by employing graph-based perturbation strategies, to quantify the contribution of individual graph components on the model answer. We conduct extensive experiments comparing XGRAG against RAG-Ex, an XAI baseline for standard RAG, and evaluate its robustness across various question types, narrative structures and LLMs. Our results demonstrate a 14.81% improvement in explanation quality over the baseline RAG-Ex across NarrativeQA, FairyTaleQA, and TriviaQA, evaluated by F1-score measuring alignment between generated explanations and original answers. Furthermore, XGRAG explanations exhibit a strong correlation with graph centrality measures, validating its ability to capture graph structure. XGRAG provides a scalable and generalizable approach towards trustworthy AI through transparent, graph-based explanations that enhance the interpretability of RAG systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces XGRAG, a framework for explaining GraphRAG systems via graph-based perturbation strategies that quantify the contribution of individual KG components to LLM answers. It claims these yield causally grounded explanations, demonstrated by a 14.81% F1-score improvement over the RAG-Ex baseline (measuring alignment between explanations and original answers) across NarrativeQA, FairyTaleQA, and TriviaQA, plus a strong correlation with graph centrality measures.

Significance. If the perturbation method can be shown to isolate causal effects and the F1 metric validated as a faithful proxy, XGRAG would address a genuine gap in interpretability for structured RAG, offering a scalable graph-native XAI approach with potential to enhance trustworthiness in KG-augmented LLMs.

major comments (3)
  1. [Abstract and §3] Abstract and §3 (method): the central claim that graph-based perturbations produce 'causally grounded' explanations is not supported by any description of the perturbation algorithm, implementation details (e.g., node/edge removal, path isolation), or controls for confounders such as LLM priors, indirect paths, or correlated signals. This detail is load-bearing for the causality interpretation.
  2. [§4] §4 (experiments): the reported 14.81% F1 improvement and correlation with centrality lack any mention of statistical tests, error bars, number of runs, ablation studies, or controls for confounding factors, making it impossible to evaluate whether the gains reflect genuine attribution quality rather than dataset artifacts or baseline weaknesses.
  3. [Evaluation] Evaluation section: F1-score alignment between generated explanations and original model answers is used as the quality metric, but this can be high for post-hoc rationalizations without establishing faithfulness or causality; no synthetic ground-truth recovery, do-calculus application, or independent causal validation is provided to support the 'causally grounded' claim.
minor comments (2)
  1. Define all acronyms (KG, LLM, GraphRAG, XAI) on first use and ensure consistent notation for graph components throughout.
  2. Clarify how 'explanation quality' is operationalized beyond F1 alignment and discuss limitations of this proxy.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback, which identifies key areas where the manuscript can be strengthened in terms of methodological clarity, statistical rigor, and validation of claims. We address each major comment below and will make corresponding revisions to the paper.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (method): the central claim that graph-based perturbations produce 'causally grounded' explanations is not supported by any description of the perturbation algorithm, implementation details (e.g., node/edge removal, path isolation), or controls for confounders such as LLM priors, indirect paths, or correlated signals. This detail is load-bearing for the causality interpretation.

    Authors: We appreciate the referee highlighting the need for greater detail on this central aspect. While §3 presents the perturbation strategy at a conceptual level to quantify component contributions via graph interventions, we agree that implementation specifics are insufficiently elaborated. In the revised manuscript, we will expand §3 with a full algorithmic description, pseudocode for node/edge removal and path isolation, and explicit discussion of potential confounders including LLM priors, indirect paths, and correlated signals. The 'causally grounded' framing is intended in the interventional sense—measuring output changes under targeted graph perturbations—which follows standard practices in attribution-based XAI; we will clarify this and add any feasible controls. revision: yes

  2. Referee: [§4] §4 (experiments): the reported 14.81% F1 improvement and correlation with centrality lack any mention of statistical tests, error bars, number of runs, ablation studies, or controls for confounding factors, making it impossible to evaluate whether the gains reflect genuine attribution quality rather than dataset artifacts or baseline weaknesses.

    Authors: We acknowledge this gap in the experimental reporting. The current results present the 14.81% F1 improvement and centrality correlations but omit the requested statistical elements. In the revision, we will specify the number of runs, include error bars or standard deviations, report statistical significance tests (e.g., paired t-tests), add ablation studies on perturbation variants, and incorporate controls for confounding factors such as dataset artifacts or baseline-specific issues. This will provide a more robust evaluation of the gains. revision: yes

  3. Referee: [Evaluation] Evaluation section: F1-score alignment between generated explanations and original model answers is used as the quality metric, but this can be high for post-hoc rationalizations without establishing faithfulness or causality; no synthetic ground-truth recovery, do-calculus application, or independent causal validation is provided to support the 'causally grounded' claim.

    Authors: The F1 metric is employed as a proxy for explanation quality through alignment with answer-influencing components, and it is further supported by the observed correlation with graph centrality measures. We recognize its limitations in directly proving causality or faithfulness and will revise the evaluation section to discuss these explicitly, including potential for post-hoc rationalization. Full do-calculus application is challenging without a complete causal model of the LLM, which is not feasible here; however, we will explore adding synthetic ground-truth recovery experiments where possible to provide additional validation. revision: partial

Circularity Check

0 steps flagged

No circularity: framework evaluated on external benchmarks with independent metrics

full rationale

The paper introduces XGRAG as a perturbation-based framework for explanations in GraphRAG systems and evaluates it empirically on standard external QA datasets (NarrativeQA, FairyTaleQA, TriviaQA) against a named baseline (RAG-Ex). It reports F1-score improvements and correlation with graph centrality measures. No derivation chain, equations, or first-principles results are presented that reduce by construction to fitted parameters, self-definitions, or self-citation load-bearing premises. The 'causally grounded' framing is an interpretive label on the perturbation method rather than a derived claim that loops back to inputs. Evaluation uses independent data and proxies, satisfying the criteria for a self-contained, non-circular contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that targeted graph perturbations isolate causal contributions of individual components; no free parameters or new physical entities are mentioned in the abstract.

axioms (1)
  • domain assumption Graph-based perturbation strategies can isolate the causal contribution of individual graph components to an LLM's final answer.
    Invoked to justify that the generated explanations are causally grounded rather than correlational.
invented entities (1)
  • XGRAG framework no independent evidence
    purpose: To generate causally grounded explanations for GraphRAG via graph perturbations.
    New framework introduced in the paper; no independent evidence outside the reported experiments is provided in the abstract.

pith-pipeline@v0.9.0 · 5581 in / 1459 out tokens · 37958 ms · 2026-05-08T03:23:27.368344+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages · 10 internal anchors

  1. [1]

    Kgrag-ex: Explainable retrieval-augmented generation with knowledge graph-based perturbations, 2025

    Georgios Balanos, Evangelos Chasanis, Konstantinos Skianis, and Evaggelia Pitoura. Kgrag-ex: Explainable retrieval-augmented generation with knowledge graph-based perturbations, 2025. URL https://arxiv.org/abs/2507.08443

    work page arXiv 2025

  2. [2]

    D. J. Best and D. E. Roberts. Algorithm as 89: The upper tail probabilities of spearman's rho. Journal of the Royal Statistical Society. Series C (Applied Statistics), 24 0 (3): 0 377--379, 1975. ISSN 00359254, 14679876. URL http://www.jstor.org/stable/2347111

    work page arXiv 1975

  3. [3]

    Llms for explainable ai: A comprehensive survey, 2025

    Ahsan Bilal, David Ebert, and Beiyu Lin. Llms for explainable ai: A comprehensive survey, 2025. URL https://arxiv.org/abs/2504.00125

    work page arXiv 2025

  4. [4]

    Bloom, M

    Benjamin S. Bloom, M. D. Engelhart, E. J. Furst, W. H. Hill, and D. R. Krathwohl. Taxonomy of Educational Objectives: Handbook 1, Cognitive Domain. David McKay Company, New York, 1956

    1956

  5. [5]

    Mean Reciprocal Rank, pp.\ 1703--1703

    Nick Craswell. Mean Reciprocal Rank, pp.\ 1703--1703. Springer US, Boston, MA, 2009. ISBN 978-0-387-39940-9. doi:10.1007/978-0-387-39940-9_488. URL https://doi.org/10.1007/978-0-387-39940-9_488

    work page doi:10.1007/978-0-387-39940-9_488 2009

  6. [6]

    W.W. Daniel. Applied Nonparametric Statistics. Duxbury advanced series in statistics and decision sciences. PWS-KENT Pub., 1990. ISBN 9780534919764. URL https://books.google.de/books?id=0hPvAAAAMAAJ

    1990

  7. [7]

    From Local to Global: A Graph RAG Approach to Query-Focused Summarization

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization, 2024. URL https://arxiv.org/abs/2404.16130

    work page internal anchor Pith review arXiv 2024

  8. [8]

    Retrieval-Augmented Generation for Large Language Models: A Survey

    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, 2024. URL https://arxiv.org/abs/2312.10997

    work page internal anchor Pith review arXiv 2024

  9. [9]

    The Llama 3 Herd of Models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. URL https://arxiv.org/abs/2407.21783

    work page internal anchor Pith review arXiv 2024

  10. [10]

    LightRAG: Simple and Fast Retrieval-Augmented Generation

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Lightrag: Simple and fast retrieval-augmented generation, 2024. URL https://arxiv.org/abs/2410.05779

    work page internal anchor Pith review arXiv 2024

  11. [11]

    ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1–23

    Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, Mahantesh Halappanavar, Ryan A Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al. Retrieval-augmented generation with graphs (graphrag). arXiv preprint arXiv:2501.00309, 2024

    work page arXiv 2024

  12. [12]

    Haoyu Han, Harry Shomer, Yu Wang, Yongjia Lei, Kai Guo, Zhigang Hua, Bo Long, Hui Liu, and Jiliang Tang. Rag vs. graphrag: A systematic evaluation and key insights, 2025. URL https://arxiv.org/abs/2502.11371

    work page arXiv 2025

  13. [13]

    Mistral 7B

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. URL https://arxiv.org/abs/2310.06825

    work page internal anchor Pith review arXiv 2023

  14. [14]

    TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension

    Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension, 2017. URL https://arxiv.org/abs/1705.03551

    work page internal anchor Pith review arXiv 2017

  15. [15]

    Gemma 3 Technical Report

    Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Thomas Mesnard, et al. Gemma 3 technical report. arXiv preprint arXiv:2503.19786, 2025. URL https://arxiv.org/abs/2503.19786

    work page internal anchor Pith review arXiv 2025

  16. [16]

    Kocisk \' y , J

    Tomáš Kočiský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Edward Grefenstette. The narrativeqa reading comprehension challenge, 2017. URL https://arxiv.org/abs/1712.07040

    work page arXiv 2017

  17. [17]

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021. URL https://arxiv.org/abs/2005.11401

    work page internal anchor Pith review arXiv 2021

  18. [18]

    Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation

    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. In International Conference on Learning Representations, 2025

    2025

  19. [19]

    Visual Instruction Tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. URL https://arxiv.org/abs/2304.08485

    work page internal anchor Pith review arXiv 2023

  20. [20]

    Deepseek-r1 thoughtology: Let's think about llm reasoning

    Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, et al. Deepseek-r1 thoughtology: Let's think about llm reasoning. arXiv preprint arXiv:2504.07128, 2025. URL https://arxiv.org/abs/2504.07128

    work page arXiv 2025

  21. [21]

    Nomic embed: Training a reproducible long context text embedder.arXiv preprint arXiv:2402.01613, 2024

    Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. Nomic embed: Training a reproducible long context text embedder. arXiv preprint arXiv:2402.01613, 2024. URL https://arxiv.org/abs/2402.01613

    work page arXiv 2024

  22. [22]

    The pagerank citation ranking: Bringing order to the web

    Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical Report 1999-66, Stanford InfoLab, November 1999. URL http://ilpubs.stanford.edu:8090/422/. Previous number = SIDL-WP-1999-0120

    1999

  23. [23]

    Rage against the machine: Retrieval-augmented llm explanations, 2024

    Joel Rorseth, Parke Godfrey, Lukasz Golab, Divesh Srivastava, and Jaroslaw Szlichta. Rage against the machine: Retrieval-augmented llm explanations, 2024. URL https://arxiv.org/abs/2405.13000

    work page arXiv 2024

  24. [24]

    Rag-ex: A generic framework for explaining retrieval augmented generation

    Viju Sudhi, Sinchana Ramakanth Bhat, Max Rudat, and Roman Teucher. Rag-ex: A generic framework for explaining retrieval augmented generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '24, pp.\ 2776–2780, New York, NY, USA, 2024. Association for Computing Machinery. ISBN 97984...

    work page doi:10.1145/3626772.3657660 2024

  25. [25]

    Causalrag: Integrating causal graphs into retrieval-augmented generation, 2025

    Nengbo Wang, Xiaotian Han, Jagdip Singh, Jing Ma, and Vipin Chaudhary. Causalrag: Integrating causal graphs into retrieval-augmented generation, 2025. URL https://arxiv.org/abs/2503.19878

    work page arXiv 2025

  26. [26]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. URL https://arxiv.org/abs/2201.11903

    work page internal anchor Pith review arXiv 2023

  27. [27]

    Usable xai: 10 strategies towards exploiting explainability in the llm era, 2025

    Xuansheng Wu, Haiyan Zhao, Yaochen Zhu, Yucheng Shi, Fan Yang, Lijie Hu, Tianming Liu, Xiaoming Zhai, Wenlin Yao, Jundong Li, Mengnan Du, and Ninghao Liu. Usable xai: 10 strategies towards exploiting explainability in the llm era, 2025. URL https://arxiv.org/abs/2403.08946

    work page arXiv 2025

  28. [28]

    Fantastic questions and where to find them: Fairytaleqa -- an authentic dataset for narrative comprehension, 2022

    Ying Xu, Dakuo Wang, Mo Yu, Daniel Ritchie, Bingsheng Yao, Tongshuang Wu, Zheng Zhang, Toby Jia-Jun Li, Nora Bradford, Branda Sun, Tran Bao Hoang, Yisi Sang, Yufang Hou, Xiaojuan Ma, Diyi Yang, Nanyun Peng, Zhou Yu, and Mark Warschauer. Fantastic questions and where to find them: Fairytaleqa -- an authentic dataset for narrative comprehension, 2022. URL h...

    work page arXiv 2022

  29. [29]

    2024 , isbn =

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. Retrieval-augmented generation with knowledge graphs for customer service question answering. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, pp.\ 2905–2909. ACM, July 2024....

    work page doi:10.1145/3626772.3661370 2024

  30. [30]

    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 format.date year duplicate empty "emp...

  31. [31]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  32. [32]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  33. [33]

    tool\_used

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    work page arXiv 2007