Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Reliable Reasoning Path: Distilling Effective Guidance for LLM Reasoning with Knowledge Graphs

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims RRP's ranked reasoning paths give a 7B LLM state-of-the-art QA accuracy on WebQSP and CWQ.

desk verdict RRP is a genuinely strong empirical entry in KG-LLM reasoning with real SOTA numbers, but its headline bidirectional mechanism is underspecified to the point of being untestable from the text. read the letter →

arxiv 2506.10508 v1 pith:IJVFN7IG submitted 2025-06-12 cs.CL cs.AI

classification cs.CLcs.AI
keywords knowledgegraphreasoninglargelanguagemodelsmulti-hopquestionansweringpathsrelationembeddingbidirectionaldistributionlearningrethinkingmoduleknowledge-intensiveQA
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

RRP is a framework for teaching LLMs to answer knowledge-intensive multi-hop questions by giving them distilled reasoning paths mined from a knowledge graph. The paper's claim is that a reliable reasoning path—a sequence of entity–relation steps connecting the question's entity to the answer—is what LLMs are missing, and that both the path's meaning and its graph structure matter. RRP generates two kinds of paths: semantic paths produced by an LLM, and structural paths produced by relation embedding with bidirectional distribution learning. A rethinking module scores, filters, and orders those paths before they are given to the LLM. On WebQSP and CWQ, RRP reports Hits@1 of 90.0 and 64.5, surpassing the prior RoG method by 4.3 and 2.0 points with a 7B-parameter LLM.

What carries the argument

The load-bearing objects are reasoning paths $\gamma = e_0 \xrightarrow{r_1} e_1 \xrightarrow{r_2} \cdots \xrightarrow{r_n} e_n$, sequences of entities and relations in the KG that lead from the question entity to the answer. The structural generator is carried by relation embedding plus bidirectional distribution learning: it builds entity representations from incident relation vectors, walks the graph with an attention-weighted decoder conditioned on question instructions $\omega_i$, and trains with a loss whose Jensen–Shannon term $\sum_i D_{JS}(P_f^i, P_b^{n-i})$ forces the forward path distribution to agree with the backward one. The rethinking module combines a semantic score $S_1(q,\gamma_i)=\cos(v_q,v_\gamma^i)$ from the LLM's embeddings with a structural score $S_2(q,\gamma_i)$ from the path-entity embeddings, forming $S=\lambda_1 S_1 + \lambda_2 S_2$ to rank and filter candidate paths.

What would settle it

On the CWQ test set, check whether paths whose forward and backward walks agree are more likely to end at the ground-truth answer than paths where they disagree; also, remove the agreement term from the training loss and see whether the multi-hop Hits@1 stays high.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that organizing retrieved facts into a ranked, logically consistent reasoning path improves LLM reasoning more than simply supplying the facts. The paper formalizes a reliable path as $\gamma = e_0 \xrightarrow{r_1} e_1 \xrightarrow{r_2} \cdots \xrightarrow{r_n} e_n$ connecting the question entity to the answer entity, and trains a semantic generator to imitate such paths by minimizing KL divergence. A structural generator adds relation embeddings and a bidirectional distribution loss, enforcing that the forward entity distribution $P_f^i$ (from question to answer) and backward distribution $P_b^i$ (from answer to question) agree, so multi-hop paths are consistent. A rethinking module then scores each candidate path by semantic cosine similarity and structural cosine similarity, filters the weakest, and orders the rest by importance. With those ordered paths as context, RRP reports state-of-the-art Hits@1 of 90.0 on WebQSP and 64.5 on CWQ, and a plug-and-play study shows consistent gains when the same ranked paths are given to a range of different base LLMs without fine-tuning.

Load-bearing premise

The structural path generator assumes that making the path found when walking forward from the question's entity to the answer agree with the path found when walking backward from the answer to the question's entity will pick out reasoning paths that lead to correct answers, but the paper never defines what those target walks are or shows that agreement predicts a right answer.

Editorial extensions

If this is right

  • Supplying LLMs with ordered, filtered reasoning paths rather than raw retrieved paths should reduce hallucinations on questions whose answers require combining several facts.
  • The framework can be attached to any LLM at inference time without fine-tuning, with reported Hits@1 gains exceeding 20 percent across several base models.
  • A 7B-parameter LLM with RRP surpasses much larger LLM-based and KG-augmented baselines on WebQSP and CWQ, suggesting structured guidance can partly substitute for model scale.
  • The tuned weights put much more weight on structural similarity for the multi-hop CWQ dataset, supporting the paper's argument that complex questions need graph-structural paths, not just semantically relevant triples.

Reading between the lines

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

  • A testable extension the authors leave implicit is to make the filtering threshold $\theta$ per question instead of global, since questions with many candidate paths may need stricter filtering than simple ones.
  • The bidirectional agreement idea could transfer to other graph-constrained generation tasks, such as multi-hop retrieval or faithful summarization, wherever a forward and a backward pass over the same structure are available.
  • The claimed value of the learned structural module could be probed by replacing it with a cheap graph-traversal heuristic that ranks paths by length or frequency, isolating how much of the gain comes from bidirectional distribution learning.
  • The plug-and-play results imply the distilled paths carry knowledge the base models do not reliably infer, so one natural use is improving smaller, edge-deployed models where fine-tuning is infeasible.
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 introduces RRP, a KG-enhanced LLM framework for knowledge-intensive QA. It generates two sets of candidate reasoning paths: semantic paths from a fine-tuned LLM (trained to maximize the likelihood of gold shortest paths between question and answer entities) and structural paths from an LSTM-based generator with relation embeddings and a bidirectional distribution consistency loss. A rethinking module scores each candidate path by semantic and structural cosine similarities, filters by a threshold θ, and feeds the ranked paths to an LLM for the final answer. On WebQSP and CWQ, RRP reports Hits@1 of 90.0 and 64.5 with LLaMA2-Chat-7B, surpassing RoG by 4.3 and 2.0 points, and shows consistent gains when the generated paths are used to prompt several frozen LLMs.

Significance. If the results are reproducible, the paper would make a useful contribution: it demonstrates a plug-and-play path-distillation method that improves multiple LLMs without fine-tuning them, and it reports held-out test numbers on two standard benchmarks against a broad baseline set. The evaluation is not internally circular: the training losses in Section IV are supervised by gold answers and paths, while the reported Hits@1/F1 are test-set numbers, and Table V shows the contribution of each module. The main weaknesses are that the structural path generator's objective is under-specified in Eq. (8) and the rethinking module's inputs are not fully defined, so the central claimed mechanism cannot currently be reproduced or isolated.

major comments (3)
  1. [§IV-B, Eq. (8)] The loss Lr2 in Eq. (8) is the training objective of the structural path generator, yet the target distributions P*_f and P*_b are never defined in the manuscript. Since P^i_f and P^i_b are the model distributions from Eq. (7), P*_f and P*_b must be provided independently (e.g., one-hot distributions at ea and eq, label-smoothed targets, or distributions over shortest paths), but no construction is given. The KL terms are the dominant part of Lr2, so without these definitions the structural module—the paper's main novelty—cannot be implemented or checked. Moreover, no experiment isolates the bidirectional consistency term: the ablations in Table V remove the entire structural module (WebQSP Hits@1 drops from 90.0 to 86.7 without rethinking), so the reported CWQ gain over RoG cannot be attributed to bidirectional agreement rather than to ordinary forward-only supervised path prediction. Please define P*_f and P*_b and add an ablation that removes or replaces the DJS term while keeping the rest of the structural module.
  2. [§IV-B, Eqs. (3)-(4)] The relation embeddings v_r are the core representation used in the match vectors m^i_<e,r,e'> and in the initialization of entity embeddings, but their origin and training are unspecified. It is not stated whether v_r are randomly initialized and learned with Lr2, or initialized from a pretrained embedding method; the paper only says GloVe is used for the question's word embeddings. The dimensions of v_r, W1, W2, and the LSTM hidden state are also omitted. This prevents reimplementation and leaves open the possibility that the structural gains depend on an unspecified initialization. Please state how v_r are obtained, whether they are trainable, and their dimensionality.
  3. [§IV-C, Eqs. (9)-(10)] The rethinking module uses embeddings whose extraction is not described. For S1, the text says 'we utilize the module described in Section IV-A to obtain the embedding v_q of the question and the embedding v^i_γ of the i-th reasoning path,' but Section IV-A defines a distribution P(γ|a,q,G) and a KL loss; it does not describe how a question or a path is mapped to a vector. For S2, Eq. (10) averages entity embeddings, but it is unclear whether ~v_q is the LSTM final hidden state or a learned question representation from the structural module, and how the path's entities are selected. Please specify the exact encoders used at inference time and whether the same trained weights are reused for scoring and generation.
minor comments (5)
  1. [Throughout] There are numerous typos that should be corrected: Section II heading 'Enhancing LLMs withot KG', Algorithm 2 line 2 'Gnerate' and line 7 'socre', Section IV-C 'modue', and Eq. (10) '˜v donates'.
  2. [Table V] The checkmarks in Table V are rendered as '!' symbols; please use a proper legend with check/cross marks and ensure the symbols are consistent in the PDF.
  3. [Figures 1 and 2] The figures contain garbled mathematical text (e.g., 'ℎ𝑎𝑎𝑎𝑎 𝑔𝑔𝑔𝑔𝑎𝑔𝑔𝑔𝑔𝑎𝑔𝑔𝑔𝑔' in Figure 1), likely from a PDF extraction issue; please replace them with clean vector graphics with correct math rendering.
  4. [§V-E, Table II vs. §V-D, Table I] Table II reports LLaMA2-Chat-7B + RRP at 86.79 Hits@1 on WebQSP, while Table I reports RRP at 90.0 with the same base model; please clarify whether Table I includes fine-tuning of the reasoning LLM while Table II uses a frozen model, and state which setting is used for the SOTA comparison.
  5. [§V-F, Figures 3-4] The hyperparameter sensitivity analysis is reported only on WebQSP, yet the optimal (λ1, λ2) shifts from (0.5, 0.5) on WebQSP to (0.1, 0.9) on CWQ; please add a sensitivity analysis for CWQ or explain why the optimal balance shifts so strongly toward structural information.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: RRP's losses are supervised by gold paths and answers, and all headline numbers are held-out test results; the undefined P* targets in Eq. (8) are a reproducibility gap, not a circular step.

full rationale

The derivation chain is not circular. The semantic path generator (Eq. 2) minimizes KL divergence between a uniform posterior over gold reasoning paths connecting the question entity to the answer entity and the LLM's path distribution; this is standard supervised maximum-likelihood training, with the gold answer/path as external supervision rather than a quantity derived from the model's own output. The structural path generator (Eq. 8) uses relation embeddings and a bidirectional distribution loss; while the target distributions P*_f and P*_b are never defined in the paper, the text gives no construction that identifies them with P^i_f/P^i_b or with the model's predictions, so no self-definitional reduction can be exhibited. The rethinking module reuses embeddings from the trained generators to compute cosine similarity scores (Eqs. 9-11), but that is parameter sharing within one trained system, not fitting a test statistic from the quantity it predicts. The reported Hits@1/F1 on WebQSP and CWQ are held-out test numbers, and the plug-and-play study (Table II) evaluates the same generated paths with frozen, non-fine-tuned LLMs against each LLM's unaugmented baseline, providing externally grounded evidence. The only self-citations ([5], [6]) are related-work references about GraphRAG and KG prompting; they are not used to justify RRP's central mechanism or to forbid alternatives. The main documented weakness, the undefined P* targets in Eq. (8), is a reproducibility/completeness issue and should be resolved by the authors, but it is not circularity because no equation in the paper equates the structural loss to its own input by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The framework's central claims rest on several modeling assumptions and per-dataset hyperparameters. No new physical or mathematical entities are postulated; the method introduces only modules, objectives, and scoring functions.

free parameters (2)
  • lambda1, lambda2 = WebQSP: (0.5, 0.5); CWQ: (0.1, 0.9)
    Weights balancing semantic and structural similarity scores in Eq. (11); set differently per dataset with no documented validation protocol.
  • theta = 0.6 on WebQSP; not reported for CWQ
    Threshold for filtering low-importance reasoning paths; sensitivity analysis is only shown for WebQSP, so its selection on CWQ is unclear.
assumptions (4)
  • domain assumption Uniform prior over reliable reasoning paths connecting question entity to answer entity (Eq. 1).
    Used to define the KL objective for semantic path generation; assumes all shortest paths are equally reliable.
  • ad hoc to paper The subset of shortest paths Gamma* approximates the full set of reliable paths in the KL expectation (Eq. 2).
    Computational approximation; no bound on the error introduced by ignoring non-shortest paths.
  • ad hoc to paper Forward and backward entity distributions should be similar for correct reasoning (Eq. 8).
    Load-bearing for the structural module; the paper asserts this assumption without independent evidence, and P*_f and P*_b are undefined.
  • domain assumption Distributional hypothesis: question and correct reasoning path embeddings should be similar in LLM feature space (Section IV-C).
    Justifies semantic similarity score S1; imported from linguistics and not validated in this setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reliable Reasoning Path: Distilling Effective Guidance for LLM Reasoning with Knowledge Graphs." pith.science (2026). https://pith.science/paper/IJVFN7IG

@misc{pith2026250610508,
  author       = {Pith},
  title        = {Pith review of: Reliable Reasoning Path: Distilling Effective Guidance for LLM Reasoning with Knowledge Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJVFN7IG}},
  note         = {Machine review of arXiv:2506.10508}
}
read the original abstract

Large language models (LLMs) often struggle with knowledge-intensive tasks due to a lack of background knowledge and a tendency to hallucinate. To address these limitations, integrating knowledge graphs (KGs) with LLMs has been intensively studied. Existing KG-enhanced LLMs focus on supplementary factual knowledge, but still struggle with solving complex questions. We argue that refining the relationships among facts and organizing them into a logically consistent reasoning path is equally important as factual knowledge itself. Despite their potential, extracting reliable reasoning paths from KGs poses the following challenges: the complexity of graph structures and the existence of multiple generated paths, making it difficult to distinguish between useful and redundant ones. To tackle these challenges, we propose the RRP framework to mine the knowledge graph, which combines the semantic strengths of LLMs with structural information obtained through relation embedding and bidirectional distribution learning. Additionally, we introduce a rethinking module that evaluates and refines reasoning paths according to their significance. Experimental results on two public datasets show that RRP achieves state-of-the-art performance compared to existing baseline methods. Moreover, RRP can be easily integrated into various LLMs to enhance their reasoning abilities in a plug-and-play manner. By generating high-quality reasoning paths tailored to specific questions, RRP distills effective guidance for LLM reasoning.

Figures

Figures reproduced from arXiv: 2506.10508 by the authors.

Figure 1
Figure 1. LLMs frequently hallucinate when addressing complex questions. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of the proposed method. Given the question and KG, 1) we first utilize the powerful semantic capabilities of LLMs to generate [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance across different values of θ on WebQSP. 87.7 88.3 89.2 90.0 89.7 89.3 89.0 86 87 88 89 90 91 0.2 0.3 0.4 0.5 0.6 0.7 0.8 λ1 Hits@1 (%) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance across different values of λ1 on WebQSP. λ1 represents the weight of semantic information, λ2 = 1 – λ1 represents the weight of structural information, and together they satisfy λ1 + λ2 = 1. To thoroughly evaluate the sensitivity of our method to various hy…
Figure 6
Figure 6. Figure 6: Prompt template for large language model reasoning. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Enhancing Large Language Models with Reliable Knowledge Graphs

    cs.CL 2025-06 conditional novelty 2.0 of 10

    A thesis composed of four published papers proposes contrastive KG error detection, attribute-aware error-aware embedding, inductive graph completion, and KG prompting, but adds no new result beyond those papers.

Reference graph

Works this paper leans on

41 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Llama JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale et al. , “Llama JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023

  2. [2]

    Stanford alpaca: An instruction-following llama model,

    R. Taori, I. Gulrajani, T. Zhang, Y . Dubois, X. Li, C. Guestrin, P. Liang, and T. B. Hashimoto, “Stanford alpaca: An instruction-following llama model,” https://github.com/tatsu-lab/stanford alpaca, 2023

  3. [3]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei...

  4. [4]

    Palm: Scaling language modeling with pathways,

    A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra, A. Roberts, P. Barham, H. W. Chung, C. Sutton, S. Gehrmann, P. Schuh, K. Shi, S. Tsvyashchenko, J. Maynez, A. Rao, P. Barnes, Y . Tay, N. Shazeer, V . Prabhakaran, E. Reif, N. Du, B. Hutchinson, R. Pope, J. Bradbury, J. Austin, M. Isard, G. Gur-Ari, P. Yin, T. Duke, A. Levskaya, S. Ghemawat, S. Dev,...

  5. [5]

    A survey of graph retrieval- augmented generation for customized large language models,

    Q. Zhang, S. Chen, Y . Bei, Z. Yuan, H. Zhou, Z. Hong, J. Dong, H. Chen, Y . Chang, and X. Huang, “A survey of graph retrieval- augmented generation for customized large language models,” arXiv preprint arXiv:2501.13958, 2025

  6. [6]

    Knowgpt: Knowledge graph based prompting for large language models,

    Q. Zhang, J. Dong, H. Chen, D. Zha, Z. Yu, and X. Huang, “Knowgpt: Knowledge graph based prompting for large language models,” in Advances in Neural Information Processing Systems , A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, Eds., vol. 37. Curran Associates, Inc., 2024, pp. 6052–6080. [Online]. Available: https://pr...

  7. [7]

    Controlling conditional language models without catastrophic forgetting,

    T. Korbak, H. Elsahar, G. Kruszewski, and M. Dymetman, “Controlling conditional language models without catastrophic forgetting,” in Inter- national Conference on Machine Learning . PMLR, 2022, pp. 11 499– 11 528

  8. [8]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models.” in The Tenth International Conference on Learning Representations. OpenReview.net, 2022. [Online]. Available: http: //dblp.uni-trier.de/db/conf/iclr/iclr2022.html#HuSW ALWWC22

Show all 41 references
  1. [9]

    Does fine-tuning LLMs on new knowledge encourage hallucinations?

    Z. Gekhman, G. Yona, R. Aharoni, M. Eyal, A. Feder, R. Reichart, and J. Herzig, “Does fine-tuning LLMs on new knowledge encourage hallucinations?” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , Y . Al-Onaizan, M. Bansal, and Y .-N. ...

  2. [10]

    Towards a unified view of parameter- efficient transfer learning,

    J. He, C. Zhou, X. Ma, and et al., “Towards a unified view of parameter- efficient transfer learning,” in The Tenth International Conference on Learning Representations, 2022

  3. [11]

    Retrieval-augmented generation for large language models: A survey,

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, M. Wang, and H. Wang, “Retrieval-augmented generation for large language models: A survey,” 2024. [Online]. Available: https://arxiv.org/abs/2312.10997

  4. [12]

    RAPTOR: Recursive abstractive processing for tree-organized retrieval,

    P. Sarthi, S. Abdullah, A. Tuli, S. Khanna, A. Goldie, and C. D. Manning, “RAPTOR: Recursive abstractive processing for tree-organized retrieval,” in The Twelfth International Conference on Learning Representations , 2024. [Online]. Available: https: //openreview.net/forum?id=...

  5. [13]

    Lightrag: Simple and fast retrieval-augmented generation,

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

  6. [14]

    From local to global: A graph rag approach to query-focused summarization,

    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. [Online]. Available: https://arxiv.org/abs/2404.16130

  7. [15]

    Meta-learning with memory-augmented neural networks,

    A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap, “Meta-learning with memory-augmented neural networks,” in Interna- tional conference on machine learning . PMLR, 2016, pp. 1842–1850

  8. [16]

    Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph,

    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,” in The Twelfth International Conference on Learning Representations , 2024. [Online]. Available: http...

  9. [17]

    Reasoning on graphs: Faithful and interpretable large language model reasoning,

    L. LUO, Y .-F. Li, R. Haf, and S. Pan, “Reasoning on graphs: Faithful and interpretable large language model reasoning,” in The Twelfth International Conference on Learning Representations , 2024. [Online]. Available: https://openreview.net/forum?id=ZGNWW7xZ6Q

  10. [18]

    Kg-cot: Chain-of-thought prompting of large language models over knowledge graphs for knowledge-aware question answering,

    R. Zhao, F. Zhao, L. Wang, X. Wang, and G. Xu, “Kg-cot: Chain-of-thought prompting of large language models over knowledge graphs for knowledge-aware question answering,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K....

  11. [19]

    UniKGQA: Unified retrieval and reasoning for solving multi-hop question answering over knowledge graph,

    J. Jiang, K. Zhou, X. Zhao, and J.-R. Wen, “UniKGQA: Unified retrieval and reasoning for solving multi-hop question answering over knowledge graph,” in The Eleventh International Conference on Learning Representations , 2023. [Online]. Available: https: //openreview.net/forum?...

  12. [20]

    Self-distillation bridges distribution gap in language model fine-tuning,

    Z. Yang, Q. Liu, T. Pang, H. Wang, H. Feng, M. Zhu, and W. Chen, “Self-distillation bridges distribution gap in language model fine-tuning,” arXiv preprint arXiv:2402.13669 , 2024

  13. [21]

    Does fine-tuning llms on new knowledge encourage hallucinations?

    Z. Gekhman, G. Yona, R. Aharoni, M. Eyal, A. Feder, R. Reichart, and J. Herzig, “Does fine-tuning llms on new knowledge encourage hallucinations?” arXiv preprint arXiv:2405.05904 , 2024

  14. [22]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschel et al. , “Retrieval- augmented generation for knowledge-intensive nlp tasks,” Advances in Neural Information Processing Systems , vol. 33, pp. 9459–9474, 2020

  15. [23]

    Re- trieving and reading: A comprehensive survey on open-domain question answering,

    F. Zhu, W. Lei, C. Wang, J. Zheng, S. Poria, and T.-S. Chua, “Re- trieving and reading: A comprehensive survey on open-domain question answering,” arXiv preprint arXiv:2101.00774 , 2021

  16. [24]

    Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation,

    Y . Sun, S. Wang, S. Feng, S. Ding, C. Pang, J. Shang, J. Liu, X. Chen, Y . Zhao, Y . Lu et al. , “Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation,” arXiv preprint arXiv:2107.02137, 2021

  17. [25]

    K-bert: Enabling language representation with knowledge graph,

    W. Liu, P. Zhou, Z. Zhao, Z. Wang, Q. Ju, H. Deng, and P. Wang, “K-bert: Enabling language representation with knowledge graph,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 34, 2020, pp. 2901–2908

  18. [26]

    Knowledge-driven cot: Exploring faithful reasoning in llms for knowledge-intensive question answering,

    K. Wang, F. Duan, S. Wang, P. Li, Y . Xian, C. Yin, W. Rong, and Z. Xiong, “Knowledge-driven cot: Exploring faithful reasoning in llms for knowledge-intensive question answering,” arXiv preprint arXiv:2308.13259, 2023

  19. [27]

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

    M. Li, S. Miao, and P. Li, “Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation,” in International Conference on Learning Representations , 2025

  20. [28]

    Glove: Global vectors for word representation,

    J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for word representation,” in Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , 2014, pp. 1532–1543

  21. [29]

    The value of semantic parse labeling for knowledge base question answering,

    W.-t. Yih, M. Richardson, C. Meek, M.-W. Chang, and J. Suh, “The value of semantic parse labeling for knowledge base question answering,” in Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , K. Erk and N. A. Smit...

  22. [30]

    The web as a knowledge-base for answering complex questions,

    A. Talmor and J. Berant, “The web as a knowledge-base for answering complex questions,” in Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , M. Walker, H. Ji,...

  23. [31]

    Key-value memory networks for directly reading documents,

    A. Miller, A. Fisch, J. Dodge, A.-H. Karimi, A. Bordes, and J. Weston, “Key-value memory networks for directly reading documents,” in Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, J. Su, K. Duh, and X. Carreras, Eds. Austin, Texas: Ass...

  24. [32]

    Improving multi-hop question answering over knowledge graphs using knowledge base embeddings,

    A. Saxena, A. Tripathi, and P. Talukdar, “Improving multi-hop question answering over knowledge graphs using knowledge base embeddings,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , D. Jurafsky, J. Chai, N. Schluter, and J. Tetre...

  25. [33]

    Improving multi-hop knowledge base question answering by learning intermediate supervision signals,

    G. He, Y . Lan, J. Jiang, W. X. Zhao, and J.-R. Wen, “Improving multi-hop knowledge base question answering by learning intermediate supervision signals,” in Proceedings of the 14th ACM International Conference on Web Search and Data Mining , ser. WSDM ’21. New York, NY , USA:...

  26. [34]

    TransferNet: An effective and transparent framework for multi-hop question answering over relation graph,

    J. Shi, S. Cao, L. Hou, J. Li, and H. Zhang, “TransferNet: An effective and transparent framework for multi-hop question answering over relation graph,” in Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , M.-F. Moens, X. Huang, L. Specia...

  27. [35]

    Sequence-to-sequence knowledge graph completion and question answering,

    A. Saxena, A. Kochsiek, and R. Gemulla, “Sequence-to-sequence knowledge graph completion and question answering,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , S. Muresan, P. Nakov, and A. Villavicencio, Ed...

  28. [36]

    Open domain question answering using early fusion of knowledge bases and text,

    H. Sun, B. Dhingra, M. Zaheer, K. Mazaitis, R. Salakhutdinov, and W. Cohen, “Open domain question answering using early fusion of knowledge bases and text,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , E. Riloff, D. Chiang, J. Hoc...

  29. [37]

    PullNet: Open domain question answering with iterative retrieval on knowledge bases and text,

    H. Sun, T. Bedrax-Weiss, and W. Cohen, “PullNet: Open domain question answering with iterative retrieval on knowledge bases and text,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natura...

  30. [38]

    Subgraph retrieval enhanced model for multi-hop knowledge base question answering,

    J. Zhang, X. Zhang, J. Yu, J. Tang, J. Tang, C. Li, and H. Chen, “Subgraph retrieval enhanced model for multi-hop knowledge base question answering,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , S. Muresan...

  31. [39]

    Scaling instruction-finetuned language models,

    H. W. Chung, L. Hou, S. Longpre, B. Zoph, Y . Tay, W. Fedus, Y . Li, X. Wang, M. Dehghani, S. Brahma, A. Webson, S. S. Gu, Z. Dai, M. Suzgun, X. Chen, A. Chowdhery, A. Castro-Ros, M. Pellat, K. Robinson, D. Valter, S. Narang, G. Mishra, A. Yu, V . Zhao, Y . Huang, A. Dai, H. Y...

  32. [40]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” in Proceedings of the 36th International Conference on Neural Information Processing Systems , ser. NIPS ’22...

  33. [1901]

    Available: https://proceedings.neurips.cc/paper files/ paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf

Pith tools

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