Pith. sign in

REVIEW 3 major objections 5 minor 51 references

Multi-Hop Reasoning for Question Answering with Hyperbolic Representations

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Adding a single hyperbolic layer to a frozen T5 encoder-decoder yields consistent Exact Match gains over an identical Euclidean layer across four multi-hop QA datasets.

desk verdict A genuinely paired comparison of one hyperbolic layer versus a Euclidean control in soft-prompt multi-hop QA, but the control is underspecified and the single-run results don't yet isolate geometry. read the letter →

arxiv 2507.03612 v1 pith:EOE4L6JB submitted 2025-07-04 cs.CL cs.AI

Simon Welz , Lucie Flek , Akbar Karimi This is my paper · ORCID
classification cs.CLcs.AI
keywords multi-hopreasoningquestionansweringhyperbolicembeddingsPoincaréballdeltahyperbolicitycurvatureinitializationsoftpromptsT5
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

The paper tries to isolate whether the geometry of an added layer, rather than extra parameters, drives gains in multi-hop question answering. It inserts a single Poincaré-ball (hyperbolic) layer between a frozen T5 encoder and decoder, with soft prompts, and compares it against an identical Euclidean layer with the same number of trainable parameters. Across four datasets (2WikiMultiHopQA, MetaQA, MLPQ, PQ), the hyperbolic layer consistently improves Exact Match, with the largest gains on datasets whose knowledge graphs branch like trees. The paper also argues that initializing the layer's learnable curvature from the dataset's measured delta-hyperbolicity works better than random initialization, and that the benefit scales with the hierarchy of the data.

What carries the argument

The load-bearing object is the Poincaré ball model of hyperbolic space with a single trainable Poincaré linear layer positioned between the frozen T5 encoder and decoder. The encoder's Euclidean embeddings are pushed into the ball via the exponential map, transformed by a hyperbolic multinomial logistic regression layer, and pulled back to Euclidean space with the logarithmic map so the decoder can read them. The layer's curvature $c$ is learnable, and the paper initializes it from the relative delta-hyperbolicity of the data, $\delta_{\text{rel}}(X)$, via $c = (0.144/\delta_{\text{rel}}(X))^2$; the mechanism carries the argument because the only difference from the baseline is the geometry of this one layer, with the same parameter count.

What would settle it

Re-run the hopping-prompt experiments with a Euclidean layer of identical architecture whose learning rate and other hyperparameters are tuned on a validation split with multiple seeds; if the Euclidean layer matches or exceeds the hyperbolic gains on MetaQA and PQ, the claimed geometric advantage is an artifact of baseline quality.

Watch

Extended reading notes

Core claim

The central claim is that hyperbolic representations, added through one Poincaré linear layer with exponential and logarithmic mappings around a frozen T5 encoder-decoder, consistently outperform their Euclidean counterparts in a controlled multi-hop reasoning setting. With the same architecture and equal trainable parameters, the hyperbolic layer raises Exact Match on the hopping-prompt task from 44.36% to 46.93% on 2WikiHop, 22.92% to 28.33% on MetaQA, 81.03% to 82.60% on MLPQ, and 18.28% to 29.03% on PQ (development sets), with test-set gains following the same pattern. The paper attributes the advantage to hyperbolic space's exponential capacity to separate entities and relations along reasoning paths, reducing interference between rival paths, and it reports that setting the initial curvature from the dataset's relative delta-hyperbolicity outperforms random initialization. A distance analysis shows that for tree-like knowledge graphs, geodesic distances in the hyperbolic layer are larger than Euclidean distances, while the mostly linear MLPQ graph shows smaller gains, which the paper reads as evidence that the benefit is tied to hierarchy.

Load-bearing premise

The load-bearing premise is that the Euclidean comparison layer is a fair, equally-well-tuned counterpart, so that the reported improvements come from geometry rather than from a weaker baseline.

Editorial extensions

If this is right

  • The reported gains mean that a single hyperbolic layer, not added parameters, drives the improvement, so the same recipe could be transferred to other frozen sequence-to-sequence models.
  • Initializing curvature from the dataset's measured delta-hyperbolicity gives a practical, cheap rule for setting the layer's key hyperparameter.
  • The out-degree analysis suggests that hyperbolic layers will help most on multi-hop benchmarks whose knowledge graphs branch, and help least on linear chains.
  • The negligible inference-time overhead reported in the paper means the hyperbolic layer is a lightweight addition for deployment.

Reading between the lines

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

  • A testable extension: applying the same frozen-backbone plus Poincaré-layer recipe to decoder-only LLMs should show comparable Exact Match lifts on multi-hop benchmarks, if the geometric benefit is model-agnostic.
  • The curvature-initialization heuristic could generalize: one could estimate the delta-hyperbolicity of a task's embedding geometry and use it to set the curvature of other manifold layers, such as spherical or product-space layers.
  • The paper leaves unspecified whether the Euclidean layer was hyperparameter-tuned; an editorially added experiment tuning the Euclidean baseline would pin down the true size of the geometric advantage.
  • The distance diagnostic suggests a cheap predictor for new datasets: compute the fraction of reasoning paths where hyperbolic geodesic distance exceeds Euclidean distance, and use it to forecast whether a hyperbolic layer will help.
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 / 5 minor

Summary. The paper proposes a controlled comparison of hyperbolic versus Euclidean representations for multi-hop question answering. The method adds a single Poincaré (hyperbolic) layer to a frozen T5 encoder-decoder with soft prompts, using an exponential map to enter the ball and a logarithmic map to return to Euclidean space for the decoder. The curvature of the Poincaré layer is initialized from the dataset's relative δ-hyperbolicity and then learned. Experiments on four 2-hop QA datasets (2WikiMultiHopQA, MetaQA, MLPQ, PQ) report Exact Match improvements for the hyperbolic layer over a Euclidean layer of matched parameter count across the hopping stage (Table 3), an ablation of curvature initialization (Section 6.2), a computational overhead comparison, and an analysis of embedding distances and out-degree distributions.

Significance. If the central comparison is valid, the paper would provide a relatively clean demonstration that a single hyperbolic layer can improve multi-hop reasoning in a large language model without adding parameters, and that data-driven curvature initialization helps. The use of a frozen backbone and matched parameter counts across several datasets is a sensible research design, and the curvature-initialization result is a useful practical finding. However, the main result is currently supported only by point estimates from single runs, and the Euclidean control is not sufficiently specified to justify the attribution of the gains to hyperbolic geometry. The paper's value therefore depends on strengthening the experimental control and statistical evidence.

major comments (3)
  1. [Section 5 (Experimental Setup) and Table 3] The Euclidean baseline layer is never described. Section 5 states only that both layers have the same number of trainable parameters and the same learning rate, but the architecture of the Euclidean layer is not given. The hyperbolic layer defined in Section 4.2 includes an exponential mapping into the ball, a normalization step y = w / (1 + sqrt(1 + c||w||)), and the specific parameterization of Eq. (1). A plain linear layer would lack these non-linear transformations, so the comparison in Table 3 conflates geometry with boundedness/non-linearity and normalization. To isolate the effect of hyperbolic geometry, the authors should either specify the Euclidean layer explicitly and show it is a true c=0 limit of the Poincaré layer, or add a control that applies an equivalent squash/normalization in Euclidean space. Without this, the consistent EM improvements cannot be attributed to hyperbolic space.
  2. [Section 6.1 (Hyperbolic vs. Euclidean Layer)] All comparisons in Tables 3, 5, and 6 are based on a single run per condition; no error bars, standard deviations, or significance tests are reported. The paper claims that the hyperbolic layer 'consistently outperforms' the Euclidean layer across datasets, but with point estimates alone this could reflect optimization noise or a single favorable seed. The authors should run multiple seeds (at least three) and report mean and variance, or provide a paired significance test across datasets, to support the consistency claim.
  3. [Section 6.5 (Dataset Difficulty) and Figure 6] The claim that hyperbolic representations are 'significantly more advantageous' for datasets with more hierarchical structure is inferred from only four datasets using out-degree distribution as the proxy for hierarchy. No statistical relationship is quantified, and the small observed improvement on MLPQ could be due to its high baseline or other factors. This claim should be framed as a hypothesis or supported with a more rigorous analysis (e.g., a correlation across a larger set of datasets or a controlled intervention on graph structure).
minor comments (5)
  1. [Section 5 (Experimental Setup)] The learning rate for the additional Euclidean layer is not specified. The sentence 'for the additional hyperbolic layer, we use the same learning rate of 0.001' implies the Euclidean layer may have been tuned differently; please clarify the learning rate and any tuning performed for the Euclidean baseline.
  2. [Section 6.4 (Embedding Distances) and Table 8] The description of what is being compared in Table 8 is unclear. The text says it reports the percentage of cases where the geodesic distance in hyperbolic space is larger than the Euclidean distance 'for the Euclidean embeddings,' which is ambiguous. Please clarify which embeddings are used and what distances are being contrasted.
  3. [References] References Zhou et al. (2018a) and Zhou et al. (2018b) have identical titles, authors, and page numbers; one cites 'Santa Fe, New Mexico, USA' and the other does not. This appears to be a duplicate and should be corrected.
  4. [Section 6.2 (Curvature Ablation)] Figure 3 is referenced but the discussion does not provide numerical values for the ablation in the text; please ensure the figure is legible and that axis labels and units are defined, or include the numeric results in the caption or body.
  5. [General] The paper provides no code or instructions for reproducing the experiments beyond citing the open-source implementation of the hyperbolic operations. Given that the Euclidean baseline is not described, releasing code for the full pipeline would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the hyperbolic-vs-Euclidean comparison and curvature ablation are empirical and not derived from their own inputs.

full rationale

The paper's central claims are empirical comparisons, not derivations that reduce to their own inputs. The main result—that a Poincaré layer outperforms a Euclidean layer on multi-hop QA (Tables 3 and 5)—rests on running both models and reporting Exact Match scores; no equation defines one result in terms of the other, and the Euclidean baseline is an external control rather than a fitted quantity renamed as a prediction. The curvature initialization from δ-hyperbolicity (Eq. 2) is taken from Khrulkov et al. (2020), an external source, and the ablation in Figure 3 empirically tests this initialization against random values rather than forcing the outcome by construction. The paper contains no load-bearing self-citation: the Poincaré layer implementation is attributed to Ryohei et al. (2021) and van Spengler et al. (2023), the PaTH prompting framework to Misra et al. (2023), and the evaluation code to Ho et al. (2020), none of which are the current authors' own work. The distance analysis in Table 8 is a descriptive post-hoc observation, not a derivation of the performance gains. The fact that Section 5 does not fully specify the Euclidean layer's architecture is a legitimate experimental-control concern about whether the comparison is truly matched, but that is a validity threat, not circular reasoning. Therefore no circularity is identified.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central empirical claim relies on standard geometry and a heuristic formula from prior work. The only data-fitted parameter is the curvature. No new theoretical entities are introduced.

free parameters (1)
  • curvature c = initialized via c=(0.144/delta_rel)^2, then learned; final values not reported
    The paper's ablations show performance depends strongly on the initialization of this learnable curvature, and it is set based on a statistic of each dataset, so it is a data-fitted hyperparameter.
assumptions (4)
  • standard math Poincaré ball model and exponential/logarithmic mappings correctly represent hyperbolic geometry
    Used in Section 3.2; standard geometric background.
  • domain assumption Formula c=(0.144/delta_rel)^2 from Khrulkov et al. maps delta-hyperbolicity to a useful curvature
    Adopted in Section 3.3 from prior work; its validity for these text datasets is not re-established.
  • domain assumption Delta-hyperbolicity estimated from 1500 sampled points, repeated 5 times, approximates the dataset's structure
    Sampling details in Section 5; no convergence analysis.
  • ad hoc to paper Out-degree distribution is a meaningful measure of a knowledge graph's hierarchical structure
    Introduced in Section 5.1 and used in Section 6.5 to explain dataset difficulty and hyperbolic benefit; not validated against other hierarchy measures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Hop Reasoning for Question Answering with Hyperbolic Representations." pith.science (2026). https://pith.science/paper/EOE4L6JB

@misc{pith2026250703612,
  author       = {Pith},
  title        = {Pith review of: Multi-Hop Reasoning for Question Answering with Hyperbolic Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EOE4L6JB}},
  note         = {Machine review of arXiv:2507.03612}
}
read the original abstract

Hyperbolic representations are effective in modeling knowledge graph data which is prevalently used to facilitate multi-hop reasoning. However, a rigorous and detailed comparison of the two spaces for this task is lacking. In this paper, through a simple integration of hyperbolic representations with an encoder-decoder model, we perform a controlled and comprehensive set of experiments to compare the capacity of hyperbolic space versus Euclidean space in multi-hop reasoning. Our results show that the former consistently outperforms the latter across a diverse set of datasets. In addition, through an ablation study, we show that a learnable curvature initialized with the delta hyperbolicity of the utilized data yields superior results to random initializations. Furthermore, our findings suggest that hyperbolic representations can be significantly more advantageous when the datasets exhibit a more hierarchical structure.

Figures

Figures reproduced from arXiv: 2507.03612 by the authors.

Figure 1
Figure 1. A knowledge graph with entities as the nodes [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Model architecture for our approach. After the T5 encoder, we project the resulting Euclidean embeddings [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Average inference time per batch (=8) on [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Embeddings for 15 input samples from the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: On the left: EM score on PQ test set for our [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 38 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Ivana Balažević, Carl Allen, and Timothy Hospedales. 2019. https://arxiv.org/abs/1905.09791 Multi-relational poincar\'e graph embeddings . Preprint, arXiv:1905.09791

    work page Pith review arXiv 2019

  4. [4]

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

    2008

  5. [5]

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. https://proceedings.neurips.cc/paper_files/paper/2013/file/1cecc7a77928ca8133fa24680a88d2f9-Paper.pdf Translating embeddings for modeling multi-relational data . In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc

    work page 2013

  6. [6]

    Ines Chami, Adva Wolf, Da-Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher R \'e . 2020. Low-dimensional hyperbolic knowledge graph embeddings. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6901--6914

    work page 2020

  7. [7]

    Ines Chami, Zhitao Ying, Christopher R\' e , and Jure Leskovec. 2019. https://proceedings.neurips.cc/paper_files/paper/2019/file/0415740eaa4d9decbc8da001d3fd805f-Paper.pdf Hyperbolic graph convolutional neural networks . In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc

    arXiv 2019

  8. [8]

    Nan Chen, Xiangdong Su, and Feilong Bao. 2024. https://aclanthology.org/2024.lrec-main.744 Hyperbolic representations for prompt learning . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 8487--8492, Torino, Italia. ELRA and ICCL

    work page 2024

Show all 51 references
  1. [9]

    Bhuwan Dhingra, Christopher Shallue, Mohammad Norouzi, Andrew Dai, and George Dahl. 2018. Embedding text in hyperbolic spaces. In Proceedings of the Twelfth Workshop on Graph-Based Methods for Natural Language Processing (TextGraphs-12), pages 59--69

    2018

  2. [10]

    Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. 2022. Hyperbolic vision transformers: Combining improvements in metric learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7399--7409. IEEE

    2022

  3. [11]

    Yanlin Feng, Xinyue Chen, Bill Yuchen Lin, Peifeng Wang, Jun Yan, and Xiang Ren. 2020. Scalable multi-hop relational reasoning for knowledge-aware question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1295--1309

    2020

  4. [12]

    Octavian-Eugen Ganea, Gary B \'e cigneul, and Thomas Hofmann. 2018. Hyperbolic neural networks. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 5350--5360

    2018

  5. [13]

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multihop qa dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6609--6625, Barcelona, Spain (Onli...

    2020

  6. [14]

    Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. 2022. Understanding and improving zero-shot multi-hop reasoning in generative question answering. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1765--1775

    2022

  7. [15]

    Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. 2020. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418--6428

    2020

  8. [16]

    Prodromos Kolyvakis, Alexandros Kalousis, and Dimitris Kiritsis. 2019. Hyperkg: Hyperbolic knowledge graph embeddings for knowledge base completion. arXiv preprint arXiv:1908.04895

    2019 arXiv

  9. [17]

    Ni Lao, Tom Mitchell, and William W. Cohen. 2011. https://aclanthology.org/D11-1049 Random walk inference and learning in a large scale knowledge base . In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 529--539, Edinburgh, Scotla...

    2011

  10. [18]

    Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2018. https://doi.org/10.18653/v1/D18-1362 Multi-hop knowledge graph reasoning with reward shaping . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3243--3253, Brussels, Belgi...

    2018 doi

  11. [19]

    Jiaxu Liu, Xinping Yi, and Xiaowei Huang. 2024. https://doi.org/10.1109/TAI.2024.3440223 Deephgcn: Toward deeper hyperbolic graph convolutional networks . IEEE Transactions on Artificial Intelligence, 5(12):6172--6185

    2024

  12. [20]

    Peter J Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. 2018. Generating wikipedia by summarizing long sequences. In International Conference on Learning Representations

    2018

  13. [21]

    Qi Liu, Maximilian Nickel, and Douwe Kiela. 2019. Hyperbolic graph neural networks. Curran Associates Inc., Red Hook, NY, USA

    2019

  14. [22]

    Ang Ma, Yanhua Yu, Chuan Shi, Shuai Zhen, Liang Pang, and Tat-Seng Chua. 2024. Pmhr: Path-based multi-hop reasoning incorporating rule-enhanced reinforcement learning and kg embeddings. Electronics, 13(23):4847

    2024

  15. [23]

    Kanishka Misra, Cicero Nogueira dos Santos, and Siamak Shakeri. 2023. https://doi.org/10.18653/v1/2023.findings-acl.62 Triggering multi-hop reasoning for question answering in language models using soft prompts and random walks . In Findings of the Association for Computationa...

    2023 doi

  16. [24]

    Sebastien Montella, Lina M Rojas Barahona, and Johannes Heinecke. 2021. Hyperbolic temporal knowledge graph embeddings with relational and time curvatures. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3296--3308

    2021

  17. [25]

    Chau Duc Minh Nguyen, Tim French, Wei Liu, and Michael Stewart. 2023. Cyle: Cylinder embeddings for multi-hop reasoning over knowledge graphs. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1736--1751

    2023

  18. [26]

    Maximillian Nickel and Douwe Kiela. 2017. Poincar \'e embeddings for learning hierarchical representations. Advances in neural information processing systems, 30

    2017

  19. [27]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1--67

    2020

  20. [28]

    Shimizu Ryohei, Mukuta Yusuke, and Harada Tatsuya. 2021. Hyperbolic neural networks++. In Proceedings of International Conference on Learning Representations (ICLR)

    2021

  21. [29]

    Ramit Sawhney, Shrey Pandit, Vishwa Shah, Megh Thakkar, and Shafiq Joty. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.114 A da PT : A set of guidelines for hyperbolic multimodal multilingual NLP . In Findings of the Association for Computational Linguistics: NAACL 202...

    2024 doi

  22. [30]

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3--7, 2018, proc...

    2018

  23. [31]

    Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. In International Conference on Machine Learning, pages 4596--4604. PMLR

    2018

  24. [32]

    Yucheng Shi, Qiaoyu Tan, Xuansheng Wu, Shaochen Zhong, Kaixiong Zhou, and Ninghao Liu. 2024 a . https://doi.org/10.1145/3627673.3679722 Retrieval-enhanced knowledge editing in language models for multi-hop question answering . In Proceedings of the 33rd ACM International Confe...

    2024

  25. [33]

    Zhengliang Shi, Shuo Zhang, Weiwei Sun, Shen Gao, Pengjie Ren, Zhumin Chen, and Zhaochun Ren. 2024 b . https://doi.org/10.18653/v1/2024.acl-long.397 Generate-then-ground in retrieval-augmented generation for multi-hop question answering . In Proceedings of the 62nd Annual Meet...

    2024 doi

  26. [34]

    Rotate: Knowledge graph embedding by relational rotation in complex space

    Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations

  27. [35]

    Yiming Tan, Yongrui Chen, Guilin Qi, Weizhuo Li, and Meng Wang. 2023. https://doi.org/10.1016/j.bdr.2023.100381 Mlpq: A dataset for path question answering over multilingual knowledge graphs . Big Data Research, 32:100381

    2023

  28. [36]

    Max van Spengler, Erwin Berkhout, and Pascal Mettes. 2023. Poincar \'e resnet. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5396--5405. IEEE Computer Society

    2023

  29. [37]

    Petar Veli c kovi \'c , Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li \`o , and Yoshua Bengio. 2018. Graph attention networks. In International Conference on Learning Representations

    2018

  30. [38]

    Denny Vrande c i \'c and Markus Kr \"o tzsch. 2014. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10):78--85

    2014

  31. [39]

    Guojia Wan, Shirui Pan, Chen Gong, Chuan Zhou, and Gholamreza Haffari. 2021. Reasoning like human: hierarchical reinforcement learning for knowledge graph reasoning. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial In...

    2021

  32. [40]

    Bin Wang, Fuyong Xu, Peiyu Liu, and Zhenfang Zhu. 2024. Hypermr: Hyperbolic hypergraph multi-hop reasoning for knowledge-based visual question answering. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (...

    2024

  33. [41]

    Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. In Proceedings of the AAAI conference on artificial intelligence, volume 28

    2014

  34. [42]

    Guanchen Xiao, Jinzhi Liao, Zhen Tan, Yiqi Yu, and Bin Ge. 2022. Hyperbolic directed hypergraph-based reasoning for multi-hop kbqa. Mathematics, 10(20):3905

    2022

  35. [43]

    Weiwen Xu, Yang Deng, Huihui Zhang, Deng Cai, and Wai Lam. 2021. Exploiting reasoning chains for multi-hop science question answering. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1143--1156

    2021

  36. [44]

    Yi Xu, Dongsheng Wang, Bo Chen, Ruiying Lu, Zhibin Duan, Mingyuan Zhou, et al. 2022. Hyperminer: Topic taxonomy mining with hyperbolic embedding. Advances in Neural Information Processing Systems, 35:31557--31570

    2022

  37. [45]

    Bishan Yang, Scott Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Embedding entities and relations for learning and inference in knowledge bases. In Proceedings of the International Conference on Learning Representations (ICLR) 2015

    2015

  38. [46]

    Jiahao Zhang, Haiyang Zhang, Dongmei Zhang, Liu Yong, and Shen Huang. 2024 a . End-to-end beam retrieval for multi-hop question answering. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techn...

    2024

  39. [47]

    Xiaoming Zhang, Ming Wang, Xiaocui Yang, Daling Wang, Shi Feng, and Yifei Zhang. 2024 b . Hierarchical retrieval-augmented generation model with rethink for multi-hop question answering. arXiv e-prints, pages arXiv--2408

    2024

  40. [48]

    Mantong Zhou, Minlie Huang, and Xiaoyan Zhu. 2018 a . An interpretable reasoning network for multi-relation question answering. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2010--2022

    2018

  41. [49]

    Mantong Zhou, Minlie Huang, and Xiaoyan Zhu. 2018 b . https://aclanthology.org/C18-1171 An interpretable reasoning network for multi-relation question answering . In Proceedings of the 27th International Conference on Computational Linguistics, pages 2010--2022, Santa Fe, New ...

    2018

  42. [50]

    Xingchen Zhou, Peng Wang, Qiqing Luo, and Zhe Pan. 2021. Multi-hop knowledge graph reasoning based on hyperbolic knowledge graph embedding and reinforcement learning. In Proceedings of the 10th International Joint Conference on Knowledge Graphs, pages 1--9

    2021

  43. [51]

    Anjie Zhu, Deqiang Ouyang, Shuang Liang, and Jie Shao. 2022. Step by step:: A hierarchical framework for multi-hop knowledge graph reasoning with reinforcement learning

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.