Pith. sign in

REVIEW 3 major objections 6 minor 57 references

DocNavRAG: Document-Structured Graph RAG with Stateful Evidence Construction for Complex Document Question Answering

T0 review · 3 major / 6 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 DocNavRAG claims document-native graph navigation lifts answer quality 7.8% over the strongest baseline.

desk verdict Well-engineered, honestly written system paper whose headline gains are not statistically pinned down; worth a serious referee, and the authors should be pushed to release code and error bars. read the letter →

arxiv 2608.01565 v1 pith:MWTS2F4R submitted 2026-08-03 cs.CL

Dongyang Xie (1) , Yao Tian (2) , Hao Zhang (3) , Yifei Yuan (4) , Tieyun Qian (1) , Ming Zhong (1) , Jiawei Jiang (1) , Yuanyuan Zhu (1) ((1) School of Computer Science
show 5 more authors
Wuhan University (2) The Hong Kong University of Science Technology (3) The Chinese University of Hong Kong (4) ETH Zurich)
This is my paper · ORCID
classification cs.CL
keywords documentquestionansweringretrieval-augmentedgenerationgraph-basedretrievalagenticstructureevidenceconstructionmulti-documentQAlong-document
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

DocNavRAG claims that complex question answering over long document collections improves when an agent navigates the documents' own structure—headings, sections, reading order, citations—instead of repeatedly searching a flat corpus from scratch. The system organizes each document collection into a multi-granular graph whose nodes are paragraphs and structural scopes, adds lateral links for citations, shared entities, semantic similarity, and topical communities, and lets an LLM controller choose among locate, navigate, expand, and fetch operations. Retrieval is steered by an evolving evidence state that records what has been found, what is still missing, and which node to focus on next, stopping when the collected evidence is judged sufficient. Across four long- and multi-document QA benchmarks, the paper reports average improvements of 7.8% in answer quality and 17.7% in retrieved-context sufficiency over the strongest baseline, A-RAG. A sympathetic reader would care because DocNavRAG requires no fine-tuning, works across multiple LLM backbones, and suggests that the structure of the source corpus itself is a reusable lever for evidence assembly.

What carries the argument

The carrying object is the document graph $\mathcal{G} = (V, E)$ with $V = P \cup S$ (paragraph nodes plus structural nodes) and $E = E_{\mathrm{struct}} \cup E_{\mathrm{lat}}$, where structural edges preserve the recovered hierarchy and local reading order and lateral edges connect related regions across the hierarchy through citation, entity-overlap, semantic-similarity, and community-membership signals. Structural nodes are given level-conditioned profiles: broad routing scopes receive LLM-generated summaries, while local scopes keep lightweight bottom-up views built from child paragraphs. At query time, four graph actions—LOCATE, NAVIGATE, EXPAND, FETCH—provide the agent's only interface to the graph, and the retrieval state $s_t = (E_t, C_t, M_t, F_t)$ directs each next move until the evidence set is sufficient. This machinery separates navigation from grounding: the graph is a lightweight retrieval substrate rather than a fully induced semantic knowledge graph, and only source text materialized by FETCH can enter the final answer.

What would settle it

Rerun the four benchmark comparisons with multiple independent runs or bootstrap resampling over the 33, 30, 88, and 21 evaluation questions and report confidence intervals for the DocNavRAG-minus-A-RAG differences; the central claim would collapse if the intervals for the average 7.8% and 17.7% gains include zero.

Watch

Extended reading notes

Core claim

The paper's central claim is that the limiting factor in complex document QA is not just the retrieval algorithm but the interface an agent operates over. DocNavRAG therefore turns each source document into a rooted hierarchy tree of paragraph nodes and nested structural scopes, augments that hierarchy with lateral edges built from citation mentions, entity overlap, profile similarity, and community membership, and exposes a constrained action space—LOCATE, NAVIGATE, EXPAND, FETCH—through which an LLM controller gathers evidence over multiple rounds. The controller maintains a state $s_t = (E_t, C_t, M_t, F_t)$ holding accumulated source evidence, candidate nodes for later exploration, remaining evidence needs, and the current focus node, and revises that state after every action until the needs are met or the budget runs out. A distinctive design boundary is that only fetched source text can ground the final answer: navigation artifacts such as outlines, profiles, and candidate snippets guide retrieval but are excluded from answer evidence. The reported outcome is consistent gains in answer quality and retrieved-context sufficiency over passage-based, graph-based, and agentic baselines on all four benchmarks.

Load-bearing premise

The load-bearing premise is that the reported improvements are real rather than sampling noise, since each benchmark is evaluated on a single run using 33, 30, 88, and 21 questions respectively, with no significance testing or variance estimates reported.

Editorial extensions

If this is right

  • If the gains hold, the advantage is not tied to one LLM: DocNavRAG beats A-RAG on answer correctness in 19 of 24 backbone-benchmark comparisons, including under lighter DeepSeek, Qwen, and GPT-5.4-mini configurations.
  • The dominant capability differs by corpus setting: hierarchy search contributes most in multi-document scientific collections, state control and evidence fetch contribute most in long-document settings, and lateral navigation matters most when evidence is spread across heterogeneous government documents.
  • The system sits between lightweight retrieval baselines and LLM-intensive semantic graph pipelines in aggregate LLM usage, indicating that the reported quality gains do not come from the highest-cost regime measured in the paper.
  • Only fetched source text can ground the answer, so the retrieved-context sufficiency measure reflects source material rather than structural associations such as outlines, candidate snippets, and routing profiles.

Reading between the lines

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

  • If the reported gains are real, document-native structure becomes a cheap and reusable indexing axis: a layout-parsed hierarchy plus four lightweight link types can be built without LLM-extracted knowledge graphs, so the approach should transfer to any corpus with recoverable headings, reading order, and cross-references.
  • The intermediate cost profile suggests a testable extension: using a smaller, cheaper controller model for routing while keeping a stronger answer generator might preserve most of the quality gain, since the graph already constrains the search space.
  • The fetch-only grounding boundary makes a verifiable prediction: allowing unfetched navigation snippets into the generation context should degrade answer faithfulness, and the paper's own w/o Evidence Fetch ablation is the first confirmation of that prediction on these benchmarks.
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

3 major / 6 minor

Summary. DocNavRAG is a training-free agentic retrieval system for complex document question answering (CDQA). At indexing time, it builds a document-native graph that preserves each source document's hierarchy at document/section/paragraph granularity and adds lateral edges derived from citations, entity overlap, semantic similarity, and Leiden community membership (Section 4.2). At query time, an LLM controller performs LOCATE, NAVIGATE, EXPAND, and FETCH operations over this graph, maintaining a retrieval state containing evidence, candidates, remaining needs, and a current focus, until a step budget is exhausted or the evidence is judged sufficient (Section 4.3). The paper evaluates the system on ScholarQA, MDAQA, LongBenchV2-UserGuide, and LongBenchV2-Gov against passage, graph, and agentic baselines, reporting average gains of 7.8% in answer quality and 17.7% in context sufficiency over the strongest baselines (Table 3), together with ablations (Figure 3), backbone generalization across six LLMs (Table 4), and an LLM-cost analysis (Figure 4, Appendix Table 9).

Significance. If the reported gains are reliable, DocNavRAG makes a solid empirical contribution: it demonstrates that exposing document-native structure as a navigable, multi-granular search space—rather than only as flat units or induced knowledge graphs—improves evidence assembly in CDQA, with a clearly specified grounding boundary between navigation artifacts and answer evidence. The paper is transparent about sample sizes (Appendix Table 5), provides a high level of implementation detail (Appendices B and C), cleanly separates the ablation variants (Table 15), decouples the LLM evaluator (Claude-Sonnet-4-6) from the generator backbone (DeepSeek-v4-Flash), and reports complete end-to-end LLM usage (Table 9), all of which strengthen the auditability of the claims. The main uncertainty is statistical: the headline magnitudes rest on very small evaluation sets, and the backbone-robustness claim shows losses on two of six backbones. These issues are significant but appear addressable within the manuscript's scope.

major comments (3)
  1. [Section 5.2, Table 3; Appendix A.1, Table 5; Figure 3] The headline claim that DocNavRAG 'improves answer quality and context sufficiency over the strongest baseline by 7.8% and 17.7% on average' is not statistically established. The four benchmarks consist of 33, 30, 88, and 21 evaluation units (Appendix Table 5), the results come from single runs with no significance tests or confidence intervals, and the per-benchmark gains over A-RAG correspond to about 2 of 33 questions (ScholarQA, +6.7 points), 0.66 of 30 (MDAQA, +2.2), 7 of 88 (UserGuide, +8.0), and 3 of 21 (Gov, +14.3). The ablation drops in Figure 3 are equally fragile: on Gov, drops of 9.52, 23.81, and 28.57 points correspond to 2, 5, and 6 of 21 questions, so the per-component attribution stated in Section 5.3 is not robust at these sample sizes. Additionally, the 88 UserGuide units are option-level instances drawn from only 22 manuals and are therefore not independent, so clustering at the manual level further reduces the effective sample size. Since Correctness and Sufficient Context are LLM-judged, judge variance is also unquantified. The authors should add paired significance tests (McNemar or exact binomial tests on the per-question binary outcomes for the accuracy benchmarks, and bootstrap or paired tests for the judged metrics), report per-question outcome counts, or temper the 'improves ... on average' claims to a level consistent with the available evidence.
  2. [Section 5.4, Table 4; Appendix Table 8] The conclusion that 'the improvement is not specific to the main DeepSeek-v4-Flash configuration' is only partially supported by the data. DocNavRAG is behind A-RAG on two of the six backbones: Qwen3.5-Plus (2/4 Correctness wins, mean -1.14 points, including a -9.52-point loss on Gov) and GPT-5.5 (1/4 wins, mean -1.72 points, including a -3.11-point loss on ScholarQA). The robustness claim should be qualified to state that gains are consistently observed for the Flash/mini-class controllers, while the benefit is not established for the stronger backbone family; the Gov losses in particular should be discussed rather than subsumed in the '19/24' aggregate.
  3. [Section 4.2.2, Eq. (6); Appendices B.1-B.2] Several design parameters are never instantiated, and no sensitivity analysis is reported. These include the routing-scope thresholds tau_h and tau_l, the entity-overlap and semantic-similarity thresholds for lateral edges, the Leiden community resolution and community-node choice, and the retrieval step budget B. Without the concrete values used in the main experiments, readers cannot assess how much of the reported gain depends on particular settings, and the label 'training-free' does not exempt these choices from tuning. Please report the values used, and ideally include a small sensitivity study on at least tau_h/tau_l and B.
minor comments (6)
  1. [Section 1 (Introduction)] The sentence 'improving answer quality over the strongest baseline by% points on average' is missing its numeric value; it should state 7.8 percentage points.
  2. [Abstract and Section 5.2] The averaged claim of '7.8% and 17.7%' mixes comparison baselines: the answer-quality average is always relative to A-RAG, while the Sufficient Context average uses the per-dataset strongest baseline (e.g., Dense for MDAQA, which yields a 33.3-point gap). This convention should be stated explicitly, since a reader would otherwise assume both averages are relative to the same strongest baseline.
  3. [Section 5.2, observation (1)] The observation that graph-based RAG 'leads only on ScholarQA Correctness by 0.1 percentage points' rests on the difference between LightRAG (0.413) and BM25 (0.412) over 33 questions, which is a fraction of one question; this supports the qualitative point but is too fragile to carry the exact '0.1 points' framing.
  4. [Section 1, contribution 1] The claim of introducing 'the first training-free agentic Graph-based RAG system for CDQA' is risky given the agentic graph systems already cited in Section 2 (e.g., GeAR); please soften the novelty claim or justify it with a concrete comparison.
  5. [Section 5.5 and Appendix Table 9] Because the headline comparisons are against A-RAG and DocNavRAG uses 3.1x the LLM calls (4,118 vs. 1,325) and 6.4x the input tokens (94.4M vs. 14.7M) of A-RAG, consider adding a per-question cost-adjusted comparison or an explicit discussion of this quality-cost trade-off.
  6. [Figures 1, 2, 5; Sections 5.2 and 5.4] There are several typos and grammatical slips: 'benifits' (Section 5.2), 'Offeline' (Figure 2), 'hirerachical' (Figure 5), 'Complentary' (Figure 1), and 'the generally larger gains under the lighter configurations such as DeepSeek, Qwen3.5-Flash and gpt 5.4-mini configuration' (Section 5.4).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DocNavRAG's central claims are empirical comparisons against external benchmarks, with metrics judged externally on fetched evidence.

full rationale

This paper is an empirical systems paper rather than a formal derivation; there is no chain of equations or cited theorem that reduces the headline result to its own inputs. The headline claim ('improves answer quality and context sufficiency... by 7.8% and 17.7%') is measured on four external CDQA benchmarks against baselines with fixed shared preprocessing. Answer quality is scored by option accuracy (LongBenchV2 subsets) or by RAGAS Answer Correctness and Semantic Similarity, and Sufficient Context is judged by Claude-Sonnet-4-6 on the grounding boundary defined in Appendix A.3, which explicitly excludes navigation-only artifacts unless materialized through fetch_content. The method's document graph, agent actions, and evidence state are not defined in terms of the evaluation metrics; ablations (Appendix C) vary only controller-visible operations while keeping the index, embeddings, backend, and evaluator fixed. No parameter is fitted to a subset and then reported as a prediction, and no load-bearing self-citation or imported uniqueness theorem appears. Potential concerns about small sample sizes and lack of significance testing (e.g., Gov being 21 questions) are correctness/robustness risks, not circularity, and do not affect the circularity score.

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

The method relies on several domain assumptions about document structure, paragraph granularity, controller reliability, and LLM-judged metrics; the headline gains rest on threshold parameters whose values are not reported.

free parameters (5)
  • routing-scope thresholds tau_h and tau_l = not reported
    Select which structural nodes become high-level routing scopes in Eq. 6; affects where LLM summaries are allocated and, therefore, retrieval behavior.
  • entity-overlap and semantic-similarity thresholds for lateral edges = not reported
    Determine which nodes get lateral links in Table 1; the agent's cross-region reach depends on these cutoffs.
  • Leiden community resolution and community node choice = not reported
    Controls topical community granularity used for Ecomm edges.
  • paragraph chunk bounds = 200-400 tokens
    Preprocessing merges or splits paragraphs to this range; evidence granularity and fetch cost depend on it.
  • retrieval step budget B = not reported
    Terminates retrieval when the budget is exhausted; cost and recall tradeoff.
assumptions (5)
  • domain assumption MinerU layout parsing reliably recovers the true document hierarchy from headings and reading order.
    The whole graph structure is built on this hierarchy (Section 4.2.1, Appendix B.1). If headings are noisy, routing scopes misalign.
  • domain assumption Paragraphs (200 to 400 tokens) are self-contained enough to serve as answer-grounding evidence units.
    Fetching and evidence admission operate at paragraph level (Appendix B.1, A.2).
  • domain assumption The LLM controller can reliably judge evidence sufficiency from the retrieval note (can_answer flag) and update missing needs.
    Stopping and need revision depend on this judgment (Appendix B.2).
  • domain assumption Claude-Sonnet-4-6 LLM judgments of Sufficient Context and Answer Correctness are valid proxies for retrieval sufficiency and answer quality.
    Primary evidence-completeness and answer-quality metrics are LLM-judged (Appendix A.3).
  • domain assumption Lateral signals (citation, entity overlap, semantic similarity, community) connect regions that are actually complementary for a question.
    EXPAND assumes these links lead to useful evidence (Section 4.2.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DocNavRAG: Document-Structured Graph RAG with Stateful Evidence Construction for Complex Document Question Answering." pith.science (2026). https://pith.science/paper/MWTS2F4R

@misc{pith2026260801565,
  author       = {Pith},
  title        = {Pith review of: DocNavRAG: Document-Structured Graph RAG with Stateful Evidence Construction for Complex Document Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MWTS2F4R}},
  note         = {Machine review of arXiv:2608.01565}
}
read the original abstract

Answering complex questions over large document collections requires assembling complementary evidence across sections and documents. GraphRAG offers structured retrieval but typically uses fixed traversal, while agentic RAG operates over weakly structured interfaces. Our key insight is that agents should navigate document structure within and across documents rather than repeatedly search from scratch. We introduce DocNavRAG, which organizes document hierarchies and cross-region relations into a navigable graph, exposes graph operations for locating, navigating, expanding, and fetching, and maintains an evolving evidence state to guide retrieval until sufficient evidence is collected. Across four long- and multi-document QA benchmarks, DocNavRAG improves answer quality and context sufficiency over the strongest baseline by 7.8\% and 17.7\% on average.

Figures

Figures reproduced from arXiv: 2608.01565 by the authors.

Figure 1
Figure 1. Comparison of conventional agentic retrieval [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Framework overview of DocNavRAG. Source documents are first organized into a multi-granular [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Per-benchmark ablation effects, measured as absolute drops (percentage points) in each benchmark’s [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Aggregate end-to-end LLM usage across the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Detailed query-time retrieval workflow in DocNavRAG. The controller maintains a retrieval note, selects [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 25 canonical work pages

  1. [1]

    arXiv preprint arXiv:2501.05366 , year =

    Search-o1: Agentic Search-Enhanced Large Reasoning Models , author =. arXiv preprint arXiv:2501.05366 , year =

    arXiv

  2. [2]

    arXiv preprint arXiv:2507.21892 , year =

    Graph-R1: Towards Agentic GraphRAG Framework via End-to-end Reinforcement Learning , author =. arXiv preprint arXiv:2507.21892 , year =

    arXiv

  3. [3]

    International Conference on Learning Representations , year =

    Sufficient Context: A New Lens on Retrieval Augmented Generation Systems , author =. International Conference on Learning Representations , year =

    work page

  4. [4]

    Es, Shahul and James, Jithin and Espinosa-Anke, Luis and Schockaert, Steven , journal =

    work page

  5. [5]

    and Artzi, Yoav , booktitle =

    Zhang, Tianyi and Kishore, Varsha and Wu, Felix and Weinberger, Kilian Q. and Artzi, Yoav , booktitle =

    work page

  6. [6]

    Saad-Falcon, Jon and Khattab, Omar and Potts, Christopher and Zaharia, Matei , journal =

    work page

  7. [7]

    arXiv preprint arXiv:2405.07437 , year =

    Evaluation of Retrieval-Augmented Generation: A Survey , author =. arXiv preprint arXiv:2405.07437 , year =

    arXiv

  8. [8]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. arXiv preprint arXiv:2005.11401 , year =

    arXiv 2005

Show all 57 references
  1. [9]

    From Local to Global: A Graph

    Edge, Darren and Trinh, Ha and Cheng, Newman and Bradley, Joshua and Chao, Alex and Mody, Apurva and Truitt, Steven and Metropolitansky, Dasha and Ness, Robert Osazuwa and Larson, Jonathan , journal =. From Local to Global: A Graph

  2. [10]

    Guo, Zirui and Xia, Lianghao and Yu, Yanhua and Ao, Tu and Huang, Chao , journal =

  3. [11]

    arXiv preprint arXiv:2405.14831 , year =

    Guti. arXiv preprint arXiv:2405.14831 , year =

    arXiv

  4. [12]

    Zhuang, Luyao and Chen, Shengyuan and Xiao, Yilin and Zhou, Huachi and Zhang, Yujing and Chen, Hao and Zhang, Qinggang and Huang, Xiao , journal =

  5. [13]

    Du, Mingxuan and Xu, Benfeng and Zhu, Chiwei and Wang, Shaohan and Wang, Pengyu and Wang, Xiaorui and Mao, Zhendong , journal =

  6. [14]

    Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , booktitle =. Self-

  7. [15]

    Verma, Prakhar and Midigeshi, Sukruta Prakash and Sinha, Gaurav and Solin, Arno and Natarajan, Nagarajan and Sharma, Amit , journal =

  8. [16]

    Luo, Haoran and E, Haihong and Chen, Guanting and Zheng, Yandan and Wu, Xiaobao and Guo, Yikai and Lin, Qika and Feng, Yu and Kuang, Zemin and Song, Meina and Zhu, Yifan and Tuan, Luu Anh , journal =

  9. [17]

    arXiv preprint arXiv:2508.11247 , year =

    Cross-Granularity Hypergraph Retrieval-Augmented Generation for Multi-hop Question Answering , author =. arXiv preprint arXiv:2508.11247 , year =

    arXiv

  10. [18]

    Feng, Yifan and Hu, Hao and Hou, Xingliang and Liu, Shiquan and Ying, Shihui and Du, Shaoyi and Hu, Han and Gao, Yue , journal =. Hyper-

  11. [19]

    Asai, Akari and He, Jacqueline and Shao, Rulin and Shi, Weijia and Singh, Amanpreet and Chang, Joseph Chee and Lo, Kyle and Soldaini, Luca and Feldman, Sergey and D'Arcy, Mike and Wadden, David and Latzke, Matt and Tian, Minyang and Ji, Pan and Liu, Shengyan and Tong, Hao and ...

  12. [20]

    Bai, Yushi and Tu, Shangqing and Zhang, Jiajie and Peng, Hao and Wang, Xiaozhi and Lv, Xin and Cao, Shulin and Xu, Jiazheng and Hou, Lei and Dong, Yuxiao and Tang, Jie and Li, Juanzi , booktitle =

  13. [21]

    When to Use Graphs in

    Xiang, Zhishang and Wu, Chuanjie and Zhang, Qinggang and Chen, Shengyuan and Hong, Zijin and Huang, Xiao and Su, Jinsong , journal =. When to Use Graphs in

  14. [22]

    Li, Chuhan and Shangguan, Ziyao and Zhao, Yilun and Li, Deyuan and Liu, Yixin and Cohan, Arman , journal =

  15. [23]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =

    Towards Multi-Document Question Answering in Scientific Literature: Pipeline, Dataset, and Evaluation , author =. Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =. doi:10.18653/v1/2025.findings-emnlp.576 , pages =

    2025 doi

  16. [24]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =

    Retrieval-Augmented Generation with Hierarchical Knowledge , author =. Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =. doi:10.18653/v1/2025.findings-emnlp.321 , pages =

    2025 doi

  17. [25]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =

    Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =. doi:10.18653/v1/2023.acl-long....

    doi

  18. [26]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =

    Active Retrieval Augmented Generation , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =. doi:10.18653/v1/2023.emnlp-main.495 , pages =

    2023 doi

  19. [27]

    Lee, Myeonghwa and An, Seonho and Kim, Min-Soo , editor =. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , month = jun, year =. doi:10.18653/v1/2024.naacl-l...

    2024 doi

  20. [28]

    Computational Linguistics , volume =

    Measuring Attribution in Natural Language Generation Models , author =. Computational Linguistics , volume =. 2023 , publisher =. doi:10.1162/coli_a_00486 , url =

    2023 doi

  21. [29]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =

    Enabling Large Language Models to Generate Text with Citations , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =. doi:10.18653/v1/2023.emnlp-main.398 , pages =

    2023 doi

  22. [30]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , month = jun, year =

    A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers , author =. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , month = jun, year =. doi:10.18653/v1/...

    2021 doi

  23. [31]

    , editor =

    Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William and Salakhutdinov, Ruslan and Manning, Christopher D. , editor =. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , month = oct #. 2018 , address =. doi:1...

    2018 doi

  24. [32]

    Proceedings of the Third Workshop on Natural Language Generation, Evaluation, and Metrics (GEM) , month = dec, year =

    Amouyal, Samuel and Wolfson, Tomer and Rubin, Ohad and Yoran, Ori and Herzig, Jonathan and Berant, Jonathan , editor =. Proceedings of the Third Workshop on Natural Language Generation, Evaluation, and Metrics (GEM) , month = dec, year =

  25. [33]

    Transactions of the Association for Computational Linguistics , volume =

    Lost in the Middle: How Language Models Use Long Contexts , author =. Transactions of the Association for Computational Linguistics , volume =. 2024 , address =. doi:10.1162/tacl_a_00638 , pages =

    2024 doi

  26. [34]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =

    Searching for Best Practices in Retrieval-Augmented Generation , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2024.emnlp-main.981 , pages =

    2024 doi

  27. [35]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =

    Kim, Kiseung and Lee, Jay-Yoon , editor =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2024.emnlp-main.1236 , pages =

    2024 doi

  28. [36]

    Proceedings of the 3rd Workshop on Knowledge Augmented Methods for NLP , month = aug, year =

    Hwang, Taeho and Jeong, Soyeong and Cho, Sukmin and Han, SeungYoon and Park, Jong , editor =. Proceedings of the 3rd Workshop on Knowledge Augmented Methods for NLP , month = aug, year =. doi:10.18653/v1/2024.knowledgenlp-1.6 , pages =

    2024 doi

  29. [37]

    Qi, Zehan and Xu, Rongwu and Guo, Zhijiang and Wang, Cunxiang and Zhang, Hao and Xu, Wei , editor =. LONG^. Findings of the Association for Computational Linguistics: EMNLP 2024 , month = nov, year =. doi:10.18653/v1/2024.findings-emnlp.279 , pages =

    2024 doi

  30. [38]

    , editor =

    Shen, Zhili and Diao, Chenxin and Vougiouklis, Pavlos and Merita, Pascual and Piramanayagam, Shriram and Chen, Enting and Graux, Damien and Melo, Andre and Lai, Ruofei and Jiang, Zeren and Li, Zhongyang and Qi, Ye and Ren, Yang and Tu, Dandan and Pan, Jeff Z. , editor =. Findi...

    2025 doi

  31. [39]

    Findings of the Association for Computational Linguistics: ACL 2025 , month = jul, year =

    Tao, Wenyu and Xing, Xiaofen and Chen, Yirong and Huang, Linyi and Xu, Xiangmin , editor =. Findings of the Association for Computational Linguistics: ACL 2025 , month = jul, year =. doi:10.18653/v1/2025.findings-acl.20 , pages =

    2025 doi

  32. [40]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =

    Hierarchical Document Refinement for Long-context Retrieval-augmented Generation , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =. doi:10.18653/v1/2025.acl-long.176 , pages =

    doi

  33. [41]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , month = nov, year =

    Shin, Joongmin and Park, Chanjun and Park, Jeongbae and Seo, Jaehyung and Lim, Heuiseok , editor =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2025.emnlp-main.1062 , pages =

    2025 doi

  34. [42]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =

    Equipping Retrieval-Augmented Large Language Models with Document Structure Awareness , author =. Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =. doi:10.18653/v1/2025.findings-emnlp.1339 , pages =

    2025 doi

  35. [43]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =

    Shifting from Ranking to Set Selection for Retrieval Augmented Generation , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =. doi:10.18653/v1/2025.acl-long.861 , pages =

    doi

  36. [44]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = aug, year =

    Su, Weihang and Tang, Yichen and Ai, Qingyao and Wu, Zhijing and Liu, Yiqun , editor =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = aug, year =. doi:10.18653/v1/2024.acl-long.702 , pages =

    doi

  37. [45]

    arXiv preprint arXiv:2303.18223 , year =

    A Survey of Large Language Models , author =. arXiv preprint arXiv:2303.18223 , year =

    arXiv

  38. [46]

    arXiv preprint arXiv:2312.10997 , year =

    Retrieval-Augmented Generation for Large Language Models: A Survey , author =. arXiv preprint arXiv:2312.10997 , year =

    arXiv

  39. [47]

    arXiv preprint arXiv:2408.08921 , year =

    Graph Retrieval-Augmented Generation: A Survey , author =. arXiv preprint arXiv:2408.08921 , year =

    arXiv

  40. [48]

    2025 , url =

    Xu, Tianyang and Zheng, Haojie and Li, Chengze and Chen, Haoxiang and Liu, Yixin and Chen, Ruoxi and Sun, Lichao , journal =. 2025 , url =

    2025

  41. [49]

    2025 , address =

    Mavromatis, Costas and Karypis, George , booktitle =. 2025 , address =. doi:10.18653/v1/2025.findings-acl.856 , pages =

    2025 doi

  42. [50]

    2025 , address =

    Hu, Yuntong and Lei, Zhihan and Zhang, Zheng and Pan, Bo and Ling, Chen and Zhao, Liang , booktitle =. 2025 , address =. doi:10.18653/v1/2025.findings-naacl.232 , pages =

    2025 doi

  43. [51]

    2026 , doi =

    Chen, Boyu and Guo, Zirui and Yang, Zidan and Chen, Yuluo and Chen, Junze and Liu, Zhenghao and Shi, Chuan and Yang, Cheng , journal =. 2026 , doi =

    2026

  44. [52]

    2026 , address =

    Wu, Tianhao and Luo, Siqiang , booktitle =. 2026 , address =. doi:10.18653/v1/2026.findings-acl.1703 , pages =

    2026 doi

  45. [53]

    Reasoning

    Liang, Jintao and Sugang and Lin, Huifeng and Wu, You and Zhao, Rui and Li, Ziyue , booktitle =. Reasoning. 2025 , address =. doi:10.18653/v1/2025.findings-ijcnlp.122 , pages =

    2025 doi

  46. [54]

    The Probabilistic Relevance Framework:

    Robertson, Stephen and Zaragoza, Hugo , journal =. The Probabilistic Relevance Framework:. 2009 , publisher =

    2009

  47. [55]

    Information Processing & Management , volume =

    Term-Weighting Approaches in Automatic Text Retrieval , author =. Information Processing & Management , volume =. 1988 , publisher =

    1988

  48. [56]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Dense Passage Retrieval for Open-Domain Question Answering , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =. 2020 , publisher =. doi:10.18653/v1/2020.emnlp-main.550 , url =

    2020 doi

  49. [57]

    Sentence-

    Reimers, Nils and Gurevych, Iryna , booktitle =. Sentence-. 2019 , address =. doi:10.18653/v1/D19-1410 , url =

    2019 doi

Pith tools

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