Pith. sign in

REVIEW 4 major objections 7 minor 90 references

Multi-hop graph retrieval works better when you separate strongly query-aligned facts from weakly aligned but connected ones, and train for both.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 11:06 UTC pith:A7OYZGNZ

load-bearing objection Clean Graph-RAG systems paper with a real failure mode and matching train/infer design; gains look real on multi-hop path recovery, but the synthetic Freebase pretraining is coupled to the eval entity sets and the “consistent across four benchmarks” claim is oversold. the 4 major comments →

arxiv 2607.27136 v1 pith:A7OYZGNZ submitted 2026-07-29 cs.IR

KAMR: Grounding Generation via Knowledge-Aligned Multi-hop Retrieval

Xiaochen Wang , Yuan Zhong , Haoyu Wang , Ting Wang , Fenglong Ma This is my paper
classification cs.IR
keywords multi-hop retrievalknowledge graphsretrieval-augmented generationanchor tripletsconnected tripletscontrastive pretrainingpartial alignmentgraph RAG
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Answering hard questions over a knowledge graph often means chaining several facts together. Standard retrievers score each fact only by how similar its text looks to the question, so they miss facts that are needed for the chain but barely mentioned in the question. This paper argues that those facts come in two roles: anchor triplets the question tightly constrains, and connected triplets that attach to anchors through the graph even when their wording is weak. The authors build training data by masking parts of real graph triplets and asking a language model to invent matching questions, then train a retriever with two contrastive goals—one for partial pairs (anchors) and one for single elements (connected facts). At search time the system finds anchors globally, then walks the local graph neighborhood to pull in connected evidence. Across several multi-hop benchmarks and three different generators, that design improves both recovery of full reasoning paths and final answer quality versus many lexical, dense, and graph baselines.

Core claim

Graph-based multi-hop retrieval should treat anchor triplets (strongly constrained by the query) and connected triplets (weakly aligned but structurally linked) as different targets, trained with pair-level and element-level contrastive matching on synthetic partial alignments and retrieved by global anchor search followed by local expansion; doing so consistently raises multi-hop path recovery and downstream QA over independent global ranking.

What carries the argument

KAMR’s two-stage design: pair-level InfoNCE on partial triplets for anchors, element-level InfoNCE for connected facts, then inference that ranks anchors globally and iteratively expands to graph-neighbors scored by remaining elements.

Load-bearing premise

Questions that a language model invents from masked Freebase triplets are enough like real multi-hop questions that training on them teaches true query–fact alignment without leaking the test structure.

What would settle it

If you hold the same encoder and budget fixed and replace KAMR’s synthetic partial-alignment pretraining with ordinary text–triplet matching (or remove element-level training and local expansion), path recall on 3-hop PathQuestion and F1 on CWQ should stop beating strong dense baselines such as Dragon and Hybrid.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Fixed-budget multi-hop RAG should reserve part of the budget for structural expansion rather than ranking every triplet only by global similarity.
  • Benchmarks that give only final answers can still supervise multi-hop retrievers via masked-triplet query synthesis plus pair- and element-level contrastive losses.
  • Longer reasoning chains benefit from more expansion rounds, but over-expanding can starve anchor budget and hurt noisy multi-answer sets.
  • The same anchor-first mechanism remains usable on single-hop questions because anchors alone often recover the needed fact.

Where Pith is reading between the lines

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

  • Any RAG stack that already has a strong dense encoder could bolt on neighborhood expansion scored by element similarity without replacing the whole index.
  • If synthetic masked queries systematically under-cover certain relation phrasings, gains should shrink on domains far from Freebase-style graphs—worth a cross-domain stress test.
  • The anchor/connected split is a general recipe for other structured stores (tables, citation graphs) where only part of a record is named in the query.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes KAMR, a retriever for graph-based multi-hop RAG. It distinguishes "anchor triplets" (matching exactly two query-extracted elements, strongly query-aligned) from "connected triplets" (matching exactly one element but graph-attached to retrieved evidence), per Definitions 2–3. Because multi-hop benchmarks often lack query–triplet alignment supervision, the authors construct a synthetic pretraining corpus: Freebase triples (restricted to entities linked to CWQ and PathQuestion) are masked one element at a time, and LLaMA-3.1-8B generates natural-language queries from the partial triples, yielding 840,875 pairs (§3.3.1, Appendix C.2). KAMR is pretrained with two InfoNCE objectives — pair-level (La, Eq. 4) and element-level (Lc, Eq. 5) — and at inference retrieves top-M anchors globally, then expands iteratively within g-hop neighborhoods under a budget K (§3.4, Algorithm 1). Experiments on PathQuestion (PQ-2H/3H), CWQ, LC-QuAD, and WebQSP with three LLM backbones and fourteen baselines report large retrieval gains on PQ (Path Recall 85.78 vs 59.10 best baseline on PQ-3H), small QA gains on CWQ, near-ties on LC-QuAD, and a loss on WebQSP.

Significance. If the results hold, the paper makes a useful contribution to graph-based RAG: a concrete, formalized decomposition of multi-hop retrieval targets (Definitions 2–3), a supervision strategy that works without gold reasoning paths, and a simple two-stage inference procedure (Algorithm 1) with a fixed retrieval budget. The evaluation is unusually broad for this subarea — fourteen baselines across lexical, dense, structure-based, and graph-pretraining categories, three LLM backbones, gold-path retrieval metrics (Triplet/Path Recall) on PathQuestion, and component-level ablations in which removing the pair-level objective or the expansion module produces large degradations consistent with the design rationale. Controlling the text encoder (Dragon) across KAMR and baselines, fixing the seed, and releasing the construction prompt are good reproducibility practice. The PQ retrieval gains, if attributable to the method rather than to train/test entity coupling, are large and would matter for the Graph RAG community.

major comments (4)
  1. [§3.3.1 and Appendix C.2] The claimed leakage prevention is asserted, not demonstrated. The pretraining corpus is built from Freebase triples 'restricting them to entities linked to the CWQ and PathQuestion datasets' — i.e., precisely the entity universe of two of the four evaluation benchmarks — and the only safeguard offered is the sentence asserting the corpus is kept 'independent of retrieval evaluation and downstream question answering.' No overlap audit is provided: no check that gold-path triplets (or their one-element-masked variants) do not appear in the 840,875 synthetic pairs, no entity- or relation-disjoint split, no held-out control. Since synthetic queries are generated from masked triples in the same KG neighborhoods that test gold paths traverse, the retriever may be pretrained on near-paraphrases of test-path content. This is load-bearing for the central claim that the anchor/connected design, ra
  2. [§4.3, Tables 1–2] Two headline margins rest on a single fixed seed (42, §C.2) with no variance reported. On CWQ, KAMR(N=1) beats Dragon on ChatGPT F1 by 42.70 vs 42.57 (0.13 points) and on Qwen3-8B accuracy by 38.71 vs 37.25 — differences well within typical seed variance for LLM-based QA evaluation. The abstract and §4.3 claim KAMR 'consistently achieves the best performance across all three backbones'; this claim is only as strong as the stability of these sub-point margins. At minimum, report mean ± std over multiple seeds (or multiple generation runs for the API backbone) for Tables 1–2, and temper 'consistently best' language where margins are within noise.
  3. [Appendix D, Table 5] The characterization of the WebQSP results is not supported by the numbers. KAMR scores 55.59 accuracy against ColBERT's 66.95, Dragon's 63.71, and BGE's 59.87 — an 11.4-point gap to the best baseline — yet the text states KAMR 'still achieves comparable performance on single-hop question answering.' 'Comparable' here reads as claim-laundering a clear loss. Relatedly, on LC-QuAD ChatGPT accuracy, KAMR (57.18) ties Dragon and loses to Hybrid (58.36); the appendix text claims best/second-best only 'for Qwen3-8B and LLaMA2-7B,' which is technically accurate but the main-text abstract's blanket 'consistently improves ... across four benchmarks' overstates the record. The claims should be scoped to where they hold (multi-hop retrieval and PQ/CWQ QA), with WebQSP and the LC-QuAD ChatGPT column reported as honest counterpoints and discussed (e.g., why anchor-first retrieval underperforms late-i
  4. [§4.2, Table 1] The headline retrieval result compares KAMR — pretrained on 840,875 task-adjacent synthetic pairs — against off-the-shelf retrievers with no such tuning. The only pretrained baseline is SKP, whose architecture was not designed for this data, so the comparison does not isolate the contribution of the anchor/connected decomposition from the contribution of contrastive fine-tuning per se. A natural, inexpensive control is missing: fine-tune the same Dragon encoder with a single flat InfoNCE objective on the same corpus U (i.e., query vs. full or partial triplets, without the La/Lc split and without two-stage inference). Without it, the reader cannot tell how much of the Table 1 gain is 'we trained on in-distribution synthetic data' versus 'the anchor/connected design matters.' Note that the 'Without Lc' ablation (Table 3) only costs 0.76 points (75.86→75.10), so the element-level objective'
minor comments (7)
  1. [§4.2, footnote 3] The claim that nDCG-style metrics are 'impossible' because two triplet types are retrieved is too strong: ranked-list metrics can be computed over the merged ranked output before the Aq/Cq split, or per-type. Recall@K is fine, but the justification should be softened or removed.
  2. [Definition 3] The 'smallest set' recursive definition is informal; as stated it is a least-fixed-point construction and should be phrased as such (or as the closure of Aq under the one-overlap-plus-connectivity rule) for precision.
  3. [Table 1] Table 1 shows KAMR(N=2) underperforming N=1 on PQ-2H generation for Qwen3-8B (69.70 vs 73.53) and LLaMA2-7B on PQ-2H (68.69 vs 68.61 is a tie), but the text only discusses the N=2 degradation for CWQ. A sentence acknowledging the PQ-2H pattern would make the N analysis complete.
  4. [Figure 2(b)] The anchor/connected triplet lists in panel (b) contain triplet IDs (R2: Juliana Maria... Spouse...) that appear to be errors or leftovers relative to the gold path (Frederica → Spouse → Nationality), which is confusing in a figure whose purpose is to argue KAMR retrieves the right evidence. Please audit the figure content.
  5. [Table 3] The ablation uses only PQ-2H, ChatGPT-3.5, N=1, single seed. Given that Lc removal costs <1 point here, ablations on PQ-3H (where connected retrieval should matter most) would be more informative.
  6. [Reproducibility] The paper would benefit from an explicit statement on code/data availability for the synthetic corpus U and the retrieval pipeline; Appendix B/C give the prompt and hyperparameters, which is good, but a release statement is standard for reproducibility claims.
  7. [Appendix B.1] The prompt in Figure 3 includes 'Masked Answer: {answer}' as input while instructing the model not to output it — this is fine, but worth one sentence in §3.3.1 clarifying that the generated query is conditioned on the masked element's identity, since this affects how 'partial' the alignment supervision really is.

Circularity Check

0 steps flagged

Empirical systems paper: external benchmark metrics are not forced by the training construction; no load-bearing circular derivation.

full rationale

KAMR’s claimed chain is methodological and empirical, not a first-principles derivation that reduces outputs to inputs by construction. Anchor vs. connected triplets are definitional framing (Defs. 2–3) used to design two InfoNCE objectives (Eqs. 4–5) and a two-stage retrieval procedure (Alg. 1); those definitions do not make Triplet/Path Recall, Accuracy, or F1 true by construction. Pretraining uses LLM-generated partial query–triplet pairs from masked Freebase triples; evaluation uses held-out PathQuestion gold paths and QA labels on PathQuestion, CWQ, LC-QuAD, and WebQSP against fourteen external baselines and three LLM backbones. Success or failure on those metrics is independently falsifiable. Concerns about entity-restricted synthetic data overlapping evaluation KG neighborhoods are leakage/generalization risks, not circularity under the stated patterns (no fitted parameter renamed as prediction, no self-definitional metric, no uniqueness theorem imported from the authors that forces the result). Self-citations (e.g., Wang et al. related work) are peripheral, not load-bearing premises of the main claim. Therefore no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The central claim rests on operational definitions of anchor/connected triplets, the adequacy of LLM-synthesized partial query–triplet pairs as supervision, standard contrastive learning assumptions, and fixed retrieval-budget experimental choices. No new physical entities; invented constructs are methodological categories and the synthetic dataset.

free parameters (3)
  • Retrieval budgets M and K (anchor count and total triplets) = M=25, K=50
    Chosen experimentally (default M=25, K=50); directly controls the anchor/connected tradeoff the authors say can hurt CWQ when N=2.
  • Expansion iterations N and hop radius g = N∈{1,2}; g=1-hop neighborhood
    Inference hyperparameters that change multi-hop coverage; results reported for N=1 and N=2 with 1-hop neighborhoods.
  • InfoNCE temperature T and optimization hyperparameters = lr=2e-5, batch=512, 5 epochs; T not numerically specified in main text
    Standard contrastive/training knobs (AdamW lr 2e-5, batch 512, 5 epochs) that shape the learned similarity space.
axioms (5)
  • domain assumption Multi-hop evidence can be operationalized as a small connected set of triplets with semantic relevance plus endpoint-entity connectivity (Definition 1).
    Connectivity is necessary but not sufficient; the paper treats this as the retrieval constraint throughout §3.1.
  • ad hoc to paper Anchor triplets are those matching exactly two query-extracted elements; connected triplets match exactly one and attach via graph connectivity (Definitions 2–3).
    This taxonomy drives both losses and inference; it is a design choice, not a theorem from prior IR.
  • domain assumption LLM-generated natural-language questions from masked triples are valid positive supervision for query–partial-triplet and query–element alignment.
    Section 3.3.1 and Appendix B.1; without this, La and Lc have no grounding.
  • domain assumption InfoNCE with in-batch negatives yields embeddings whose cosine similarity ranks useful anchors and connected elements at inference.
    Standard contrastive retrieval assumption (Eqs. 4–5), inherited from dense retrieval practice.
  • domain assumption Comparing retrievers under a fixed top-K triplet budget into frozen LLM generators isolates retrieval quality for Graph RAG.
    Stated scope in §2 and §4; excludes end-to-end KGQA systems from direct comparison by design.
invented entities (2)
  • Anchor triplets / connected triplets (as formal retrieval targets) no independent evidence
    purpose: Split multi-hop evidence into strongly constrained vs weakly aligned but structurally required facts.
    Core conceptual objects of KAMR; defined in §3.2 and used to justify dual losses and two-stage inference.
  • Partial alignment dataset U of LLM-generated (query, partial-triplet) pairs no independent evidence
    purpose: Provide pretraining supervision without gold multi-hop path labels.
    Constructed in §3.3.1 (~840k queries from Freebase-linked triples); exists only as this paper’s training resource unless released.

pith-pipeline@v1.2.0-grok45-kimik3 · 23405 in / 3762 out tokens · 83711 ms · 2026-07-30T11:06:45.080078+00:00 · methodology

0 comments
read the original abstract

Graph-based retrieval-augmented generation increasingly relies on multi-hop retrieval, where answering a query requires composing multiple connected knowledge-graph triplets. However, existing retrievers often rank triplets independently via global semantic matching. Moreover, many multi-hop benchmarks provide only final answers, which limits supervision for query--triplet alignment and causes structurally necessary but weakly aligned facts to be missed. To address these issues, we propose a knowledge-aligned multi-hop retriever, KAMR, which distinguishes anchor triplets that are strongly constrained by the query from connected triplets that are weakly aligned yet structurally linked to the anchors. To mitigate the lack of query-triplet alignment supervision, we build a partial alignment dataset by masking triplet elements and prompting an LLM to generate corresponding queries, and optimize two contrastive objectives for pair-level and element-level matching. At inference time, KAMR retrieves anchors globally and then expands locally to collect connected evidence. Across four benchmarks, three LLM backbones, and fourteen baselines, KAMR consistently improves multi-hop retrieval and downstream question answering performance.

Figures

Figures reproduced from arXiv: 2607.27136 by Fenglong Ma, Haoyu Wang, Ting Wang, Xiaochen Wang, Yuan Zhong.

Figure 1
Figure 1. Figure 1: An example demonstrating the necessity of differ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative study. triplets; this ablation is denoted “Full Anchor Triplets” in [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Prompt for Pretraining Dataset Construction. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

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

90 extracted references · 1 canonical work pages

  1. [1]

    Attention is All you Need , url =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , url =

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    G-retriever: Retrieval-augmented generation for textual graph understanding and question answering , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    arXiv preprint arXiv:2212.00959 , year=

    Unikgqa: Unified retrieval and reasoning for solving multi-hop question answering over knowledge graph , author=. arXiv preprint arXiv:2212.00959 , year=

    Pith/arXiv arXiv

  4. [4]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Graf: Graph retrieval augmented by facts for romanian legal multi-choice question answering , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  5. [5]

    Proceedings of the ACM Web Conference 2026 , pages=

    GPR: Empowering Generation with Graph-Pretrained Retriever , author=. Proceedings of the ACM Web Conference 2026 , pages=

    2026

  6. [6]

    arXiv preprint arXiv:2606.26458 , year=

    MKG-RAG-Bench: Benchmarking Retrieval in Multimodal Knowledge Graph-Augmented Generation , author=. arXiv preprint arXiv:2606.26458 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2505.17214 , year=

    MEDMKG: Benchmarking Medical Knowledge Exploitation with Multimodal Knowledge Graph , author=. arXiv preprint arXiv:2505.17214 , year=

    Pith/arXiv arXiv

  8. [8]

    Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    ReaRev: Adaptive reasoning for question answering over knowledge graphs , author=. Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    2022

  9. [9]

    Proceedings of the 13th international conference on web search and data mining , pages=

    Stepwise reasoning for multi-relation question answering over knowledge graph with weak supervision , author=. Proceedings of the 13th international conference on web search and data mining , pages=

  10. [10]

    Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    Open domain question answering using early fusion of knowledge bases and text , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    2018

  11. [11]

    Proceedings of the 27th International Conference on Computational Linguistics , pages=

    Modeling semantics with gated graph neural networks for knowledge base question answering , author=. Proceedings of the 27th International Conference on Computational Linguistics , pages=

  12. [12]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    Transfernet: An effective and transparent framework for multi-hop question answering over relation graph , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  13. [13]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    Improving multi-hop question answering over knowledge graphs using knowledge base embeddings , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  14. [14]

    Pullnet: Open domain question answering with iterative retrieval on knowledge bases and text , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  15. [15]

    Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Subgraph retrieval enhanced model for multi-hop knowledge base question answering , author=. Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  16. [16]

    Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Beamqa: Multi-hop knowledge graph question answering with sequence-to-sequence prediction and beam search , author=. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  17. [17]

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

    Breadth first reasoning graph for multi-hop question answering , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  18. [18]

    Proceedings of the 14th ACM international conference on web search and data mining , pages=

    Improving multi-hop knowledge base question answering by learning intermediate supervision signals , author=. Proceedings of the 14th ACM international conference on web search and data mining , pages=

  19. [19]

    Proceedings of COLING 2016, the 26th international conference on computational linguistics: technical papers , pages=

    Constraint-based question answering with knowledge graph , author=. Proceedings of COLING 2016, the 26th international conference on computational linguistics: technical papers , pages=

    2016

  20. [20]

    International semantic web conference , pages=

    Lc-quad: A corpus for complex question answering over knowledge graphs , author=. International semantic web conference , pages=. 2017 , organization=

    2017

  21. [21]

    language , volume=

    9th challenge on question answering over linked data (QALD-9) , author=. language , volume=

  22. [22]

    , author=

    Dense Passage Retrieval for Open-Domain Question Answering. , author=. EMNLP (1) , pages=

  23. [23]

    Journal of machine learning research , volume=

    Visualizing data using t-SNE , author=. Journal of machine learning research , volume=

  24. [24]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2402.03216 , year=

    Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation , author=. arXiv preprint arXiv:2402.03216 , year=

    Pith/arXiv arXiv

  26. [26]

    Boost your search with the crispy mixedbread rerank models , author=

  27. [27]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  28. [28]

    arXiv preprint arXiv:2501.00309 , year=

    Retrieval-augmented generation with graphs (graphrag) , author=. arXiv preprint arXiv:2501.00309 , year=

    Pith/arXiv arXiv

  29. [29]

    Foundations and Trends

    The probabilistic relevance framework: BM25 and beyond , author=. Foundations and Trends. 2009 , publisher=

    2009

  30. [30]

    Information processing & management , volume=

    Term-weighting approaches in automatic text retrieval , author=. Information processing & management , volume=. 1988 , publisher=

    1988

  31. [31]

    2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021 , pages=

    SimCSE: Simple Contrastive Learning of Sentence Embeddings , author=. 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021 , pages=. 2021 , organization=

    2021

  32. [32]

    Findings of the association for computational linguistics: ACL 2022 , pages=

    Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models , author=. Findings of the association for computational linguistics: ACL 2022 , pages=

    2022

  33. [33]

    Unsupervised Dense Information Retrieval with Contrastive Learning , author=

  34. [34]

    2023 , howpublished=

    BCEmbedding: Bilingual and Crosslingual Embedding for RAG , author=. 2023 , howpublished=

    2023

  35. [35]

    Proceedings of the 27th International Conference on Computational Linguistics , pages=

    An Interpretable Reasoning Network for Multi-Relation Question Answering , author=. Proceedings of the 27th International Conference on Computational Linguistics , pages=

  36. [36]

    arXiv preprint arXiv:2302.07452 , year=

    How to train your dragon: Diverse augmentation towards generalizable dense retrieval , author=. arXiv preprint arXiv:2302.07452 , year=

    Pith/arXiv arXiv

  37. [37]

    Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction , author=. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2022

  38. [38]

    Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    SPLADE: Sparse lexical and expansion model for first stage ranking , author=. Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  39. [39]

    Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023) , pages=

    Jina Embeddings: A Novel Set of High-Performance Sentence Embedding Models , author=. Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023) , pages=

    2023

  40. [40]

    arXiv preprint arXiv:2405.20139 , year=

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

    Pith/arXiv arXiv

  41. [41]

    2018 , publisher=

    Improving language understanding by generative pre-training , author=. 2018 , publisher=

    2018

  42. [42]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  43. [43]

    ArXiv , year=

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter , author=. ArXiv , year=

  44. [44]

    arXiv preprint arXiv:2501.13958 , year=

    A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models , author=. arXiv preprint arXiv:2501.13958 , year=

    arXiv

  45. [46]

    Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    The value of semantic parse labeling for knowledge base question answering , author=. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  46. [47]

    arXiv preprint arXiv:1803.06643 , year=

    The web as a knowledge-base for answering complex questions , author=. arXiv preprint arXiv:1803.06643 , year=

    Pith/arXiv arXiv

  47. [48]

    arXiv preprint arXiv:2305.18742 , year=

    Graph reasoning for question answering with triplet retrieval , author=. arXiv preprint arXiv:2305.18742 , year=

    Pith/arXiv arXiv

  48. [49]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Facenet: A unified embedding for face recognition and clustering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  49. [50]

    Proceedings of the 2008 ACM SIGMOD international conference on Management of data , pages=

    Freebase: a collaboratively created graph database for structuring human knowledge , author=. Proceedings of the 2008 ACM SIGMOD international conference on Management of data , pages=

    2008

  50. [51]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    Pith/arXiv arXiv

  51. [52]

    arXiv preprint arXiv:2307.07697 , year=

    Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph , author=. arXiv preprint arXiv:2307.07697 , year=

    Pith/arXiv arXiv

  52. [53]

    arXiv preprint arXiv:2310.01061 , year=

    Reasoning on graphs: Faithful and interpretable large language model reasoning , author=. arXiv preprint arXiv:2310.01061 , year=

    Pith/arXiv arXiv

  53. [54]

    International conference on machine learning , pages=

    Retrieval augmented language model pre-training , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  54. [55]

    arXiv preprint arXiv:1807.03748 , year=

    Representation learning with contrastive predictive coding , author=. arXiv preprint arXiv:1807.03748 , year=

    Pith/arXiv arXiv

  55. [56]

    arXiv preprint arXiv:1911.03868 , year=

    Knowledge guided text retrieval and reading for open domain question answering , author=. arXiv preprint arXiv:1911.03868 , year=

    Pith/arXiv arXiv 1911

  56. [57]

    Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , pages=

    Bridging the kb-text gap: Leveraging structured knowledge-aware pre-training for kbqa , author=. Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , pages=

  57. [58]

    arXiv preprint arXiv:2307.09288 , year=

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

    Pith/arXiv arXiv

  58. [59]

    arXiv preprint arXiv:2405.16506 , year=

    Grag: Graph retrieval-augmented generation , author=. arXiv preprint arXiv:2405.16506 , year=

    Pith/arXiv arXiv

  59. [60]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  60. [61]

    Journal of Machine Learning Research , volume=

    Scaling instruction-finetuned language models , author=. Journal of Machine Learning Research , volume=

  61. [62]

    arXiv preprint arXiv:2410.20724 , year=

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

    Pith/arXiv arXiv

  62. [63]

    Stanford Center for Research on Foundation Models

    Alpaca: A strong, replicable instruction-following model , author=. Stanford Center for Research on Foundation Models. https://crfm. stanford. edu/2023/03/13/alpaca. html , volume=

    2023

  63. [64]

    IEEE transactions on neural networks , volume=

    The graph neural network model , author=. IEEE transactions on neural networks , volume=. 2008 , publisher=

    2008

  64. [65]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  65. [66]

    arXiv preprint arXiv:2312.10997 , volume=

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

    Pith/arXiv arXiv

  66. [67]

    Prompt-Guided Retrieval Augmentation for Non-Knowledge-Intensive Tasks

    Guo, Zhicheng and Cheng, Sijie and Wang, Yile and Li, Peng and Liu, Yang. Prompt-Guided Retrieval Augmentation for Non-Knowledge-Intensive Tasks. Findings of the Association for Computational Linguistics. 2023. doi:10.18653/v1/2023.findings-acl.693

    work page doi:10.18653/v1/2023.findings-acl.693 2023

  67. [68]

    arXiv preprint arXiv:2309.01219 , year=

    Siren's song in the AI ocean: a survey on hallucination in large language models , author=. arXiv preprint arXiv:2309.01219 , year=

    Pith/arXiv arXiv

  68. [69]

    arXiv preprint arXiv:2401.01313 , volume=

    A comprehensive survey of hallucination mitigation techniques in large language models , author=. arXiv preprint arXiv:2401.01313 , volume=

    Pith/arXiv arXiv

  69. [70]

    Query Rewriting in Retrieval-Augmented Large Language Models

    Ma, Xinbei and Gong, Yeyun and He, Pengcheng and Zhao, Hai and Duan, Nan. Query Rewriting in Retrieval-Augmented Large Language Models. Proceedings of the Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.322

    doi:10.18653/v1/2023.emnlp-main.322 2023

  70. [71]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  71. [72]

    Solving Elaborate Problems with Large Language Models:

    Besta, Maciej and Blach, Nils and Kubicek, Ales and Gerstenberger, Robert and Podstawski, Michal and Gianinazzi, Lukas and Gajda, Joanna and Lehmann, Tomasz and Niewiadomski, Hubert and Nyczyk, Piotr and Hoefler, Torsten , booktitle =. Solving Elaborate Problems with Large Language Models:

  72. [73]

    Journal of documentation , volume=

    A statistical interpretation of term specificity and its application in retrieval , author=. Journal of documentation , volume=. 1972 , publisher=

    1972

  73. [74]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =

  74. [75]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , 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 (ACL) , year =

  75. [76]

    International Conference on Learning Representations (ICLR) , year =

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection , author =. International Conference on Learning Representations (ICLR) , year =

  76. [77]

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

    Hykge: A hypothesis knowledge graph enhanced rag framework for accurate and reliable medical llms responses , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  77. [78]

    arXiv preprint arXiv:2501.08686 , year=

    Knowledge graph-based retrieval-augmented generation for schema matching , author=. arXiv preprint arXiv:2501.08686 , year=

    Pith/arXiv arXiv

  78. [79]

    arXiv preprint arXiv:2310.11220 , year=

    Kg-gpt: A general framework for reasoning on knowledge graphs using large language models , author=. arXiv preprint arXiv:2310.11220 , year=

    Pith/arXiv arXiv

  79. [80]

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

    Mvp-tuning: Multi-view knowledge retrieval with prompt tuning for commonsense reasoning , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  80. [81]

    Proceedings of the 1st Workshop on Natural Language Reasoning and Structured Explanations (NLRSE) , pages=

    Knowledge-Augmented Language Model Prompting for Zero-Shot Knowledge Graph Question Answering , author=. Proceedings of the 1st Workshop on Natural Language Reasoning and Structured Explanations (NLRSE) , pages=

Showing first 80 references.