Pith. sign in

REVIEW 4 major objections 6 minor 4 cited by

E^2GraphRAG: Streamlining Graph-based RAG for High Efficiency and Effectiveness

T0 review · 4 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 A hybrid tree-and-graph RAG cuts indexing time 10x and retrieval 100x.

desk verdict A practical, honest RAG efficiency paper whose speedup claims are credible; the QA-parity hinge is a plausible but unvalidated co-occurrence proxy, so it deserves review with requests for variance and retrieval-level validation. read the letter →

arxiv 2505.24226 v4 pith:M4HY4ZHZ submitted 2025-05-30 cs.AI

Yibo Zhao , Jiapeng Zhu , Ye Guo , Kangkang He , Xiang Li This is my paper · ORCID
classification cs.AI
keywords E2GraphRAGgraph-basedretrieval-augmentedgenerationsummarytreeentitygraphadaptiveretrievalco-occurrencelong-documentquestionansweringindexingefficiency
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

E2GraphRAG aims to show that graph-based retrieval-augmented generation can be made fast enough for practical use on very long documents without sacrificing answer quality. It replaces the expensive LLM-driven entity and relation extraction of GraphRAG-style systems with a SpaCy entity graph built from sentence-level co-occurrence, and it links that graph to an LLM-built summary tree through bidirectional entity–chunk indexes. Retrieval then becomes a graph lookup plus optional dense search, with an adaptive rule that picks local retrieval when query entities are closely connected and global retrieval otherwise. The paper reports up to 10x faster indexing than GraphRAG, up to 100x faster retrieval than LightRAG, and competitive accuracy on NovelQA and InfiniteBench subsets.

What carries the argument

The central mechanism is a pair of complementary structures and two indexes connecting them. The summary tree, built by recursively summarizing consecutive chunks with an LLM, provides multi-granularity global context, while the SpaCy-built entity graph, with undirected weighted edges for sentence-level co-occurrence, captures fine-grained local relations. The entity-to-chunk index $I_{e\to c}(\cdot)$ and chunk-to-entity index $I_{c\to e}(\cdot)$ let the retrieval stage jump from query entities to the chunks that mention them, and the graph-filtering rule $P_h = \{(e^i_q,e^j_q) : i<j,\ \mathrm{Dist}_G(e^i_q,e^j_q)\le h\}$ decides whether a query is local (entity pairs within $h$ hops) or global. This design moves almost all retrieval work out of the LLM and into indexed lookups.

What would settle it

Construct or select a set of questions whose supporting evidence requires relating two entities that never appear in the same sentence anywhere in the document; under E2GraphRAG's local path the entity-pair intersection will be empty for those questions, so the system will either switch to global retrieval or fail to surface the passage, and accuracy on that set should drop measurably below a system that uses LLM-extracted relations.

Watch

Extended reading notes

Core claim

The paper's central claim is that the costly components of graph-based RAG—LLM entity and relation extraction, community clustering, and LLM-driven retrieval—can be replaced with cheaper components without losing much accuracy. E2GraphRAG constructs a hierarchical summary tree with a minimal LLM summarization prompt and a weighted entity graph from SpaCy extractions, where an edge between two entities means they co-occur in the same sentence. Two one-to-many indexes map every entity to the chunks containing it and every chunk to its entities, so local questions can be answered by intersecting the chunk sets of query entities. A graph-filtering step keeps only query entity pairs within $h$ hops, and if no such pair exists the system switches to global dense retrieval over the tree with occurrence-based reranking. The result, the paper argues, is a single adaptive pipeline that matches or beats prior graph RAG methods on long-document QA while reducing indexing and retrieval time by orders of magnitude.

Load-bearing premise

The system assumes that two entities co-occurring in the same sentence are semantically related, and that questions needing local evidence will involve such co-occurring pairs; if that assumption fails, the entity-to-chunk lookup retrieves the wrong chunks and accuracy must fall back on the global path.

Editorial extensions

If this is right

  • Graph-based RAG becomes feasible for roughly 200k-token documents on a single GPU with open-weight 7B-8B models, since indexing takes about $n/(g-1)$ LLM calls and retrieval uses no LLM calls.
  • Users no longer need to manually choose between local and global query modes, because the connectivity of query entities in the co-occurrence graph makes the choice automatically.
  • Retrieval latency drops to the level of plain dense retrieval, because the local path is index intersections and the global path is Faiss lookup plus occurrence counting.
  • The approach depends only on an off-the-shelf entity tagger and an embedding model, making it portable across domains without JSON-schema prompts or retries.
  • Reported accuracy on NovelQA with Qwen2.5 and on InfiniteQA with both backbones indicates that the cheaper components do not automatically hurt answer quality, although gains vary by dataset.

Reading between the lines

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

  • Not claimed in the paper: if sentence-level co-occurrence is a weak proxy for semantic relations, the local path will systematically miss cross-sentence relations; a testable extension is to augment edge weights with embedding similarity or dependency-based relations.
  • The adaptive $h$-hop threshold is a single global hyperparameter, so per-query or learned thresholds could improve precision on questions that involve many loosely related entities.
  • Because the indexes are cheap to build, the same entity graph could be reused for tasks beyond QA, such as summarization or fact verification, and could be updated incrementally as new chunks are added.
  • A direct comparison of entity-pair recall against GraphRAG's LLM-extracted relations would isolate exactly where the co-occurrence proxy loses information.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes E2GraphRAG, a graph-based retrieval-augmented generation framework that combines an LLM-built summary tree with a SpaCy-based entity graph using sentence-level co-occurrence edges, along with entity-to-chunk and chunk-to-entity indexes. Retrieval adaptively selects between local mode (entity-pair graph filtering followed by index mapping) and global mode (dense retrieval plus occurrence ranking) based on whether query entities are connected within a hop threshold. Experiments on NovelQA, InfiniteChoice, and InfiniteQA with Qwen2.5-7B and Llama3.1-8B report indexing and querying times and QA accuracy; the paper claims up to 10x faster indexing than GraphRAG and 100x faster retrieval than LightRAG while maintaining competitive QA performance. The paper also provides ablations, a cost analysis, and a code repository.

Significance. The main contribution is practical efficiency: replacing LLM-based entity extraction with SpaCy co-occurrence drastically cuts indexing cost, and the adaptive retrieval removes manual query-mode selection. The paper provides code, uses three public baselines, and reports ablations on three long-document QA benchmarks. If the results hold, the method is a useful practical alternative for resource-constrained deployments. However, the empirical evidence for QA parity is incomplete: no variance estimates are reported, the core assumption behind the graph-based retrieval is not directly validated, and the algorithm description contains inconsistencies. These issues are fixable and do not invalidate the overall approach.

major comments (4)
  1. [Table 1] All accuracy and timing numbers in Table 1 appear to be single-run measurements with no standard deviations, confidence intervals, or significance tests. This is particularly relevant because GraphRAG-L outperforms E2GraphRAG by 2.38 points on NovelQA with Llama3.1 (43.64 vs 41.26) and by 2.8-3.9 points on InfiniteChoice with both backbones (46.72 vs 43.23 with Qwen; 43.66 vs 39.74 with Llama). The abstract and Section 5 state that E2GraphRAG 'maintains competitive QA performance'; this claim must be qualified with these shortfalls, and the speedup claims would be strengthened by reporting variance or repeated runs.
  2. [Section 3.1 and 3.2] The validity of the graph-based retrieval hinges on the assumption that same-sentence co-occurrence of SpaCy-extracted entities is a sufficient proxy for semantic relations (Section 3.1). This assumption is never directly tested. The ablation study in Table 2 reports only aggregate accuracy shifts; for NovelQA, removing Graph Filter changes accuracy by just 1.08 points (45.38 vs 44.30), which does not demonstrate that the graph is retrieving the correct evidence. I recommend adding a retrieval-level evaluation: for a sample of local questions, measure whether gold evidence chunks are among the retrieved set, or compare retrieval hit rates against an LLM-extracted relation graph. This would substantiate the mechanism behind the QA results.
  3. [Appendix A, Section 3.2] The retrieval algorithm is specified inconsistently. The main text says the hop threshold is reduced until the number of chunks drops below k, but Algorithm 1 line 11 uses the literal constant 25 as the threshold, and line 12 updates an undefined variable l ('h=h-1 or l=l+1'). Appendix C states 'we choose a relatively large value4' without specifying k. Please clarify the exact stopping criterion and define all variables; otherwise the method is not reproducible from the description.
  4. [Section 4.2 / Table 1] The experimental comparison omits two methods that motivate the work (FastGraphRAG and LazyGraphRAG), and the only dense-retrieval baseline appears in the ablation table rather than in the main comparison. Given that the paper's practical claim is about the trade-off between cost and QA quality, the authors should either include a standard dense RAG baseline in Table 1 or explicitly report its accuracy in the main table, so readers can see what the graph machinery adds over a plain summary-tree retriever.
minor comments (6)
  1. [Section 3.2, Eq. (2)] The variable name 'Cevidience' should be 'C_evidence' for readability.
  2. [Algorithm 1, line 17] The condition 'Count(^C==0)' should be 'Count(^C)==0'.
  3. [Appendix C] Please report the concrete hyperparameter values used in the experiments, including group size g, initial hop threshold h, k, and chunk overlap.
  4. [References] Reference [1] contains a duplicated protocol in the URL: 'http://https://www.microsoft.com/...' should be corrected.
  5. [Figure 2] The legend entry 'selected chunks' appears twice; one occurrence should likely read 'candidate chunks' or be removed.
  6. [Section 4.5] The anomalous improvement when removing Dense Retrieval on NovelQA (45.90 vs 45.38) is attributed to hallucination; this deserves a brief discussion of whether the global branch may sometimes hurt.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical and evaluated against external benchmarks, not derived from fitted inputs or self-citation.

full rationale

The paper's central claims are efficiency and effectiveness measurements, not predictions derived from fitted parameters. Indexing speed is measured as wall-clock time on external long-document corpora and is separately supported by counting LLM calls from the stated tree construction rule (Section 4.4, Appendix D); that count is a direct consequence of the definition of the tree, not an input that secretly encodes the speedup. Retrieval speed is similarly measured. The competitive-QA claim is evaluated with external benchmarks (NovelQA, InfiniteChoice, InfiniteQA) against baselines such as GraphRAG, LightRAG, and RAPTOR. The retrieval design uses sentence-level co-occurrence of SpaCy entities as a proxy for relations, and the ablation study in Table 2 tests variants against the full system; this tests a design assumption rather than fitting a parameter and then predicting the same quantity. Equations (1) and (2) define the graph-filtering and index-mapping operations, but the QA accuracy they help produce is measured externally, so there is no reduction of an output to an input. The paper contains no uniqueness theorem or load-bearing self-citation: prior work is cited for standard components such as SpaCy, Faiss, BGE-M3, and baselines. Appendix E concedes that the retrieval design is intuitive and not exhaustive, which is a stated limitation rather than a circularity. Overall, the derivation chain is self-contained: every claimed result is either a direct algebraic consequence of stated construction rules or an empirical observation against independent benchmarks, so no circular step is present.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central efficiency claim rests mainly on implementation choices (SpaCy instead of LLM extraction, no clustering) rather than on free parameters. The QA-effectiveness claim rests on several domain assumptions about SpaCy coverage, co-occurrence as relation, summary fidelity, and baseline fairness. No new theoretical entities are postulated.

free parameters (6)
  • chunk_size = 1200 tokens
    Section 3.1 sets chunk size to 1200 tokens; results may shift with chunking granularity.
  • chunk_overlap = 100 tokens
    Section 3.1 uses 100-token overlap between adjacent chunks to mitigate sentence fragmentation.
  • group_size_g = not reported
    Recursive summarization groups g consecutive segments per level; g is not reported and controls the number of LLM calls and tree shape.
  • hop_threshold_h = adaptive, starts unspecified
    Section 3.2 and Algorithm 1 decrement h until the candidate chunk count falls below 25, so h is a hand-designed adaptive schedule.
  • max_chunks_k = likely 4; ambiguous 'value4' in Appendix C
    Top-k retrieved chunks are set by GPU memory in Appendix C, but the paper does not define k cleanly.
  • candidate_threshold_25 = 25
    Algorithm 1 loops while the candidate chunk count exceeds 25; the threshold appears only in pseudocode and is not justified or swept.
assumptions (6)
  • domain assumption SpaCy noun/NER extraction identifies the entities that matter for QA queries.
    Section 3.1 uses SpaCy outputs as entities; Section 3.2 uses SpaCy on the query and discards unmapped entities as noise. If SpaCy misses key entities, local retrieval will not trigger.
  • domain assumption Sentence-level co-occurrence is a sufficient proxy for semantic relations between entities.
    Section 3.1 builds weighted edges from co-occurrence in the same sentence. Relational queries that require typed or multi-sentence relations may not be captured.
  • domain assumption Recursive LLM summarization preserves the global information needed for global questions.
    Section 3.1 constructs the summary tree via recursive summarization with a minimal prompt; the global retrieval path depends on these summaries being accurate.
  • domain assumption Queries are adequately served by one of two modes: dense global retrieval or entity-pair local retrieval.
    Section 3.2 classifies every query into local or global based on graph connectivity; there is no third mode for multi-hop questions that mix both.
  • domain assumption BGE-M3 embedding similarity is a valid relevance signal for dense retrieval.
    Section 3.1 stores embeddings in Faiss and Section 3.2 uses vector similarity for top-k selection.
  • domain assumption Nano-GraphRAG plus LightRAG's HuggingFace integration faithfully represents official GraphRAG.
    Section 4.2 and Appendix C use this modified open-source implementation for the GraphRAG baseline; if the integration changes behavior, speed and accuracy comparisons shift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of E^2GraphRAG: Streamlining Graph-based RAG for High Efficiency and Effectiveness." pith.science (2026). https://pith.science/paper/M4HY4ZHZ

@misc{pith2026250524226,
  author       = {Pith},
  title        = {Pith review of: E^2GraphRAG: Streamlining Graph-based RAG for High Efficiency and Effectiveness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M4HY4ZHZ}},
  note         = {Machine review of arXiv:2505.24226}
}
read the original abstract

Graph-based RAG methods like GraphRAG have shown promising global understanding of the knowledge base by constructing hierarchical entity graphs. However, they often suffer from inefficiency and rely on manually pre-defined query modes, limiting practical use. In this paper, we propose E^2GraphRAG, a streamlined graph-based RAG framework that improves both Efficiency and Effectiveness. During the indexing stage, E^2GraphRAG constructs a summary tree with large language models and an entity graph with SpaCy based on document chunks. We then construct bidirectional indexes between entities and chunks to capture their many-to-many relationships, enabling fast lookup during both local and global retrieval. For the retrieval stage, we design an adaptive retrieval strategy that leverages the graph structure to retrieve and select between local and global modes. Experiments show that E^2GraphRAG achieves up to 10 times faster indexing than GraphRAG and 100 times speedup over LightRAG in retrieval while maintaining competitive QA performance.

Figures

Figures reproduced from arXiv: 2505.24226 by the authors.

Figure 1
Figure 1. Overview of the indexing stage of E2GraphRAG . The left part shows the indexing tasks, the center presents the four data structures, and the right part displays the two constructed indexes. contextual flow, and the clustering process is also time-consuming. Further, RAPTOR adopts the traditional RAG-style vector-based retrieval, which may lead to inaccurate retrieval results [40]. To improve the efficiency of graph-… view at source ↗
Figure 2
Figure 2. The retrieval stage of E2GraphRAG. Operations belonging to the local retrieval are highlighted in light yellow, while those for global retrieval are highlighted in light green. ⋆If no entities are identified, we cannot utilize the entities to support meaningful retrieval. In such cases, the query is treated as a global query, and Dense Retrieval is performed over the summary tree. Specifically, we adopt a collapsed-… view at source ↗
Figure 3
Figure 3. Time cost as a function of document token count for each method. The statistic is based on [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Distribution of questions across different entity count buckets. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Agents-K1: Towards Agent-native Knowledge Orchestration

    cs.AI 2026-06 unverdicted novelty 6.0 of 10

    Agents-K1 is an end-to-end pipeline with a multimodal parser, 4B GRPO-trained extractor, and agent CLI that builds scientific knowledge graphs from full papers and was run on 2.46 million documents to produce Scholar-KG.

  2. MG$^2$-RAG: Multi-Granularity Graph for Multimodal Retrieval-Augmented Generation

    cs.IR 2026-04 unverdicted novelty 6.0 of 10

    MG²-RAG proposes a multi-granularity graph RAG framework that constructs hierarchical multimodal nodes via entity-driven visual grounding and performs structured retrieval, delivering SOTA results on four multimodal t...

  3. Youtu-GraphRAG: Vertically Unified Agents for Graph Retrieval-Augmented Complex Reasoning

    cs.IR 2025-08 conditional novelty 6.0 of 10

    A schema-bounded agentic framework unifies graph construction and retrieval, reporting SOTA accuracy and up to 90.71% token-cost savings across six benchmarks, plus a new anonymized QA dataset.

  4. DyG-RAG: Dynamic Graph Retrieval-Augmented Generation with Event-Centric Reasoning

    cs.IR 2025-07 conditional novelty 5.0 of 10

    DyG-RAG builds a dynamic event graph from time-anchored event units and uses timeline retrieval with Time-CoT prompting to answer temporal questions.

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages · cited by 4 Pith papers

  1. [1]

    http://https://www.microsoft.com/en-us/research/blog/ lazygraphrag-setting-a-new-standard-for-quality-and-cost/

    Lazygraphrag: Setting a new standard for quality and cost. http://https://www.microsoft.com/en-us/research/blog/ lazygraphrag-setting-a-new-standard-for-quality-and-cost/

    work page

  2. [2]

    Fastgraphrag.https://github.com/circlemind-ai/fast-graphrag, 2024

    work page 2024

  3. [3]

    J. Chen, S. Xiao, P. Zhang, K. Luo, D. Lian, and Z. Liu. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In L.-W. Ku, A. Martins, and V . Srikumar, editors,Findings of the Association for Computational Linguistics: ACL 2024, pages 2318–2335, Bangkok, Thailand, Aug. 2024. Associati...

    work page 2024

  4. [4]

    X. Chen, X. Chen, B. He, T. Wen, and L. Sun. Analyze, generate and refine: Query expansion with LLMs for zero-shot open-domain QA. In L.-W. Ku, A. Martins, and V . Srikumar, editors, Findings of the Association for Computational Linguistics: ACL 2024, pages 11908–11922, Bangkok, Thailand, Aug. 2024. Association for Computational Linguistics

    work page 2024

  5. [5]

    T. Dao, D. Y . Fu, S. Ermon, A. Rudra, and C. Ré. Flashattention: fast and memory-efficient exact attention with io-awareness. InProceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY , USA, 2022. Curran Associates Inc

    work page 2022

  6. [6]

    Douze, A

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazaré, M. Lomeli, L. Hosseini, and H. Jégou. The faiss library. 2024

    work page 2024

  7. [7]

    X. Du, C. Xiao, and Y . Li. Haloscope: Harnessing unlabeled llm generations for hallucination detection. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume 37, pages 102948–102972. Curran Associates, Inc., 2024

    work page 2024

  8. [8]

    D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, D. Metropolitansky, R. O. Ness, and J. Larson. From local to global: A graph rag approach to query-focused summarization, 2025

    work page 2025

Show all 48 references
  1. [9]

    W. Fan, Y . Ding, L. Ning, S. Wang, H. Li, D. Yin, T.-S. Chua, and Q. Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, page 6491–6501, New York,...

    2024

  2. [10]

    Grattafiori, A

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, A. Yang, A. Fan, A. Goyal, A. Hartshorn, A. Yang, A. Mitra, A. Sravankumar, A. Korenev, A. Hinsvark, A. Rao, A. Zhang, A. Rodriguez, A. Gregerson, A. Spataru,...

    2024

  3. [11]

    Z. Guo, L. Xia, Y . Yu, T. Ao, and C. Huang. Lightrag: Simple and fast retrieval-augmented generation, 2024

    2024

  4. [12]

    B. J. Gutiérrez, Y . Shu, Y . Gu, M. Yasunaga, and Y . Su. Hipporag: Neurobiologically inspired long-term memory for large language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing S...

    2024

  5. [13]

    X. He, Y . Tian, Y . Sun, N. V . Chawla, T. Laurent, Y . LeCun, X. Bresson, and B. Hooi. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  6. [14]

    Honnibal, I

    M. Honnibal, I. Montani, S. Van Landeghem, and A. Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020

    2020

  7. [15]

    Jiang, L

    Z. Jiang, L. Zhong, M. Sun, J. Xu, R. Sun, H. Cai, S. Luo, and Z. Zhang. Efficient knowledge infusion via KG-LLM alignment. In L.-W. Ku, A. Martins, and V . Srikumar, editors,Findings of the Association for Computational Linguistics: ACL 2024, pages 2986–2999, Bangkok, Thailan...

    2024

  8. [16]

    Kirstein, T

    F. Kirstein, T. Ruas, R. Kratel, and B. Gipp. Tell me what I need to know: Exploring LLM-based (personalized) abstractive multi-source meeting summarization. In F. Dernoncourt, D. Preo¸ tiuc- Pietro, and A. Shimorina, editors,Proceedings of the 2024 Conference on Empirical Met...

    2024

  9. [17]

    Koshkin, K

    R. Koshkin, K. Sudoh, and S. Nakamura. TransLLaMa: LLM-based simultaneous translation system. In Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, editors,Findings of the Association for Computational Linguistics: EMNLP 2024, pages 461–476, Miami, Florida, USA, Nov. 2024. Associatio...

    2024

  10. [18]

    Laban, A

    P. Laban, A. Fabbri, C. Xiong, and C.-S. Wu. Summary of a haystack: A challenge to long- context LLMs and RAG systems. In Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 9885–990...

    2024

  11. [19]

    Lewis, E

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela. Retrieval-augmented generation for knowledge- intensive nlp tasks. In H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, edi...

    2020

  12. [20]

    D. Li, S. Yang, Z. Tan, J. Y . Baik, S. Yun, J. Lee, A. Chacko, B. Hou, D. Duong-Tran, Y . Ding, H. Liu, L. Shen, and T. Chen. DALK: Dynamic co-augmentation of LLMs and KG to answer Alzheimer‘s disease questions with scientific literature. In Y . Al-Onaizan, M. Bansal, and Y ....

    2024

  13. [21]

    Z. Li, S. Fan, Y . Gu, X. Li, Z. Duan, B. Dong, N. Liu, and J. Wang. Flexkbqa: A flexible llm-powered framework for few-shot knowledge base question answering.Proceedings of the AAAI Conference on Artificial Intelligence, 38(17):18608–18616, Mar. 2024

    2024

  14. [22]

    C.-Y . Lin. ROUGE: A package for automatic evaluation of summaries. InText Summarization Branches Out, pages 74–81, Barcelona, Spain, July 2004. Association for Computational Linguistics. 11

    2004

  15. [23]

    J. Liu, C. Zhang, J. Guo, Y . Zhang, H. Que, K. Deng, Z. Bai, J. Liu, G. Zhang, J. Wang, Y . Wu, C. Liu, J. Wang, L. Qu, W. Su, and B. Zheng. Ddk: Distilling domain knowledge for efficient large language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. To...

    2024

  16. [24]

    Y . Lu, W. Zhu, L. Li, Y . Qiao, and F. Yuan. LLaMAX: Scaling linguistic horizons of LLM by enhancing translation capabilities beyond 100 languages. In Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, editors,Findings of the Association for Computational Linguistics: EMNLP 2024, pa...

    2024

  17. [25]

    Nakshatri, S

    N. Nakshatri, S. Liu, S. Chen, D. Roth, D. Goldwasser, and D. Hopkins. Using LLM for improving key event discovery: Temporal-guided news stream clustering with event summaries. In H. Bouamor, J. Pino, and K. Bali, editors,Findings of the Association for Computational Lin- guis...

    2023

  18. [26]

    L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web. Technical Report 1999-66, Stanford InfoLab, November 1999. Previous number = SIDL-WP-1999-0120

    1999

  19. [27]

    B. Peng, Y . Zhu, Y . Liu, X. Bo, H. Shi, C. Hong, Y . Zhang, and S. Tang. Graph retrieval- augmented generation: A survey, 2024

    2024

  20. [28]

    R. Pope, S. Douglas, A. Chowdhery, J. Devlin, J. Bradbury, J. Heek, K. Xiao, S. Agrawal, and J. Dean. Efficiently scaling transformer inference. In D. Song, M. Carbin, and T. Chen, editors, Proceedings of Machine Learning and Systems, volume 5, pages 606–624. Curan, 2023

    2023

  21. [29]

    H. Qian, Z. Liu, P. Zhang, K. Mao, D. Lian, Z. Dou, and T. Huang. Memorag: Boosting long context processing with global memory-enhanced retrieval augmentation. InProceed- ings of the ACM Web Conference 2025 (TheWebConf 2025), Sydney, Australia, 2025. ACM. arXiv:2409.05591

    2025 arXiv

  22. [30]

    Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, ...

    2025

  23. [31]

    Ramprasad, E

    S. Ramprasad, E. Ferracane, and Z. Lipton. Analyzing LLM behavior in dialogue summarization: Unveiling circumstantial hallucination trends. In L.-W. Ku, A. Martins, and V . Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguisti...

    2024

  24. [32]

    Sahoo, P

    P. Sahoo, P. Meharia, A. Ghosh, S. Saha, V . Jain, and A. Chadha. A comprehensive survey of hallucination in large language, image, video and audio foundation models. In Y . Al- Onaizan, M. Bansal, and Y .-N. Chen, editors,Findings of the Association for Computational Linguist...

    2024

  25. [33]

    V . Sanh, A. Webson, C. Raffel, S. Bach, L. Sutawika, Z. Alyafeai, A. Chaffin, A. Stiegler, A. Raja, M. Dey, M. S. Bari, C. Xu, U. Thakker, S. S. Sharma, E. Szczechla, T. Kim, G. Chh- ablani, N. Nayak, D. Datta, J. Chang, M. T.-J. Jiang, H. Wang, M. Manica, S. Shen, Z. X. Yong...

    2022

  26. [34]

    Sarthi, S

    P. Sarthi, S. Abdullah, A. Tuli, S. Khanna, A. Goldie, and C. D. Manning. Raptor: Recursive abstractive processing for tree-organized retrieval. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  27. [35]

    Schimanski, J

    T. Schimanski, J. Ni, M. Kraus, E. Ash, and M. Leippold. Towards faithful and robust LLM specialists for evidence-based question-answering. In L.-W. Ku, A. Martins, and V . Srikumar, 12 editors,Proceedings of the 62nd Annual Meeting of the Association for Computational Linguis...

    1913

  28. [36]

    J. Shen, N. Tenenholtz, J. B. Hall, D. Alvarez-Melis, and N. Fusi. Tag-llm: repurposing general- purpose llms for specialized domains. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024

    2024

  29. [37]

    Sriramanan, S

    G. Sriramanan, S. Bharti, V . S. Sadasivan, S. Saha, P. Kattakinda, and S. Feizi. Llm-check: Investigating detection of hallucinations in large language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Inf...

    2024

  30. [38]

    J. Sun, C. Xu, L. Tang, S. Wang, C. Lin, Y . Gong, L. Ni, H.-Y . Shum, and J. Guo. Think-on- graph: Deep and responsible reasoning of large language model on knowledge graph. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  31. [39]

    Szymanski, N

    A. Szymanski, N. Ziems, H. A. Eicher-Miller, T. J.-J. Li, M. Jiang, and R. A. Metoyer. Limita- tions of the llm-as-a-judge approach for evaluating llm outputs in expert knowledge tasks. In Proceedings of the 30th International Conference on Intelligent User Interfaces, IUI ’25...

    2025

  32. [40]

    F. Tian, D. Ganguly, and C. Macdonald. Is relevance propagated fro retriever to generator in rag? InAdvances in Information Retrieval: 47th European Conference on Information Retrieval, ECIR 2025, Lucca, Italy, April 6–10, 2025, Proceedings, Part I, page 32–48, Berlin, Heidelberg,

    2025

  33. [41]

    K. Tian, E. Mitchell, A. Zhou, A. Sharma, R. Rafailov, H. Yao, C. Finn, and C. Manning. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. In H. Bouamor, J. Pino, and K. Bali, editors,Proceedings...

    2023

  34. [42]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin. Attention is all you need. In I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing...

    2017

  35. [43]

    C. Wang, R. Ning, B. Pan, T. Wu, Q. Guo, C. Deng, G. Bao, X. Hu, Z. Zhang, Q. Wang, and Y . Zhang. NovelQA: Benchmarking question answering on documents exceeding 200k tokens. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  36. [44]

    M. Wang, A. Stoll, L. Lange, H. Adel, H. Schütze, and J. Strötgen. Bring your own knowledge: A survey of methods for llm knowledge expansion, 2025

    2025

  37. [45]

    J. Wei, M. Bosma, V . Zhao, K. Guu, A. W. Yu, B. Lester, N. Du, A. M. Dai, and Q. V . Le. Finetuned language models are zero-shot learners. InInternational Conference on Learning Representations

  38. [46]

    J. Ye, Y . Wang, Y . Huang, D. Chen, Q. Zhang, N. Moniz, T. Gao, W. Geyer, C. Huang, P.-Y . Chen, N. V . Chawla, and X. Zhang. Justice or prejudice? quantifying biases in LLM-as-a-judge. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  39. [47]

    Zhang, Y

    X. Zhang, Y . Chen, S. Hu, Z. Xu, J. Chen, M. Hao, X. Han, Z. Thai, S. Wang, Z. Liu, and M. Sun. ∞Bench: Extending long context evaluation beyond 100K tokens. In L.-W. Ku, A. Martins, and V . Srikumar, editors,Proceedings of the 62nd Annual Meeting of the Association for Compu...

    2024

  40. [48]

    Y . Zhou, Y . Su, Y . Sun, S. Wang, T. Wang, R. He, Y . Zhang, S. Liang, X. Liu, Y . Ma, and Y . Fang. In-depth analysis of graph-based rag in a unified framework, 2025. 13 A Pseudo Code Algorithm 1The pseudo-code for retrieval stage. Require:q,G,T,k,h,l Ensure:Supplemental te...

    2025

Pith tools

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