Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Improving Factuality for Dialogue Response Generation via Graph-Based Knowledge Augmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that resolving dialogue coreferences before selecting knowledge triples improves response factuality, with its graph-aware method beating the previous best by 3.47% and 3.12% on two benchmarks under a new fact score.

desk verdict Sensible graph-augmented dialogue frameworks whose headline gains rest on a circularly validated metric; the evidence is weaker than the abstract implies. read the letter →

arxiv 2506.12496 v2 pith:G6QRGPSG submitted 2025-06-14 cs.CL cs.HC

Xiangyan Chen , Yujian Gan , Yimeng Gu , Matthew Purver This is my paper
classification cs.CLcs.HC
keywords dialogueresponsegenerationknowledgegraphaugmentationfactualityevaluationfactscorecoreferenceresolutionhallucinationmitigationneuralnetworksretrieval-augmented
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 is trying to establish that dialogue systems hallucinate less when the generation pipeline first rewrites the conversation to resolve pronouns, then selects knowledge triples relevant to the reformulated dialogue, and only then feeds that knowledge to the language model either as plain text or through a graph-neural encoder. If this is right, the main fixable source of factual error in multi-turn dialogue is not missing knowledge but unresolved reference: an utterance about 'her' is really an utterance about Diane Lane, and retrieval works only after that link is made explicit. The paper reports that both proposed frameworks raise a new Dialogue Fact Score over BM25, KAPING, and the strongest prior graph method, with the supervised graph-aware variant exceeding that prior method by 3.47% on OpendialKG and 3.12% on HybriDialogue. It also claims that its Dialogue Fact Score, which verifies atomic facts against both dialogue history and available evidence and adds a 'not enough information' verdict, agrees with human annotators at Cohen's kappa 0.65-0.67. The reason to care is simple: if the metric is sound, then a concrete resolve-then-select-then-augment recipe exists for cutting unsupported claims in open-domain dialogue.

What carries the argument

The load-bearing mechanism is a two-stage front end: reasoning-guided dialogue reformulation, a chain-of-thought pass that replaces each pronoun with its explicit entity, followed by dialogue sense knowledge selection, which scores triples from an encyclopedia-derived knowledge graph against the reformulated dialogue and keeps the top-N as a dialogue sense graph. That sense graph then feeds one of two generators: TG-DRG, which textualises the triples into prompts, or GA-DRG, which encodes node structure with a graph-transformer layer and a graph projector before generating. The paper's evaluation mechanism is the Dialogue Fact Score, a fine-grained precision metric that splits responses into atomic facts, verifies each against dialogue history and retrieved passages, and returns 'true', 'false', or 'not enough information'. The reformulation does the semantic work, the sense graph limits prompt noise, and the fact score converts generation into a measurable factuality signal.

What would settle it

A decisive check is a system-level human evaluation: raters see full responses from GA-DRG and the strongest baseline, without seeing which triples each system retrieved, and rank them for factual support; the human system ordering is then compared with the Dialogue Fact Score ordering. A complementary check is to recompute the score against a second, independent encyclopedia that was sealed from the generator: if the gap shrinks sharply, the gain is driven by knowledge shared between generation and evaluation rather than by the method itself.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that conditioning knowledge selection on a coreference-resolved dialogue is what makes graph knowledge augmentation pay off for response generation. The authors propose TG-DRG as an unsupervised pipeline that serializes the selected triples into a textualised graph, and GA-DRG as a supervised variant that encodes the selected subgraph with a graph transformer and injects a pooled graph representation into the LLM. They report fact-score gains from both, and that GA-DRG surpasses the strongest prior graph-based method by 3.47% and 3.12% on the two benchmarks. The accompanying Dialogue Fact Score reworks atomic-fact verification for dialogue by including conversation history and by allowing 'not enough information' when evidence is silent, on the grounds that forcing true/false on unevidenced claims mispenalizes dialogue systems. The paper frames the result as evidence that existing QA-style graph retrieval overlooks the coreference structure that multi-turn dialogue constantly exhibits, and that repairing this structure is the mechanism behind the improved factuality.

Load-bearing premise

The whole comparison rests on the new Dialogue Fact Score being a valid measure of factual consistency that ranks systems the way humans would; if that score systematically favors responses that reuse the same knowledge the generator was given, the reported 3.47% and 3.12% margins are partly an artifact.

Editorial extensions

If this is right

  • TG-DRG is training-free, so if its claimed effect holds, any off-the-shelf LLM can improve factual consistency by resolving the dialogue, selecting triples, and generating from the textualised graph without weight updates.
  • The consistent fact-score improvement of TG-DRG-R over TG-DRG-NR across base LLMs makes dialogue reformulation a portable component, not a model-specific trick.
  • GA-DRG's margin over the strongest prior graph method indicates that encoding subgraph connectivity, rather than only reading triples as text, contributes to factual accuracy; the ablation removing graph representation supports this reading.
  • If the Dialogue Fact Score is accepted, BLEU, ROUGE, and entity F1 should no longer be treated as factuality proxies in open-domain dialogue, since the paper shows cases where their ordering contradicts the fact score.
  • The 'not enough information' category creates a concrete behavioral incentive: future dialogue systems can score higher by declining to fabricate when evidence is missing.

Reading between the lines

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

  • Editorial inference: the reported margins may be inflated by shared knowledge between generation and evaluation, because the same encyclopedia-derived passages are used both to select triples during generation and to verify atomic facts during scoring; a fair test would score outputs against a sealed, independent evidence set.
  • Editorial inference: the resolve-then-retrieve pattern should transfer to other reference-heavy generation tasks, such as long-form question answering or biography writing, where unresolved mentions routinely block retrieval before generation.
  • Editorial inference: the ablations remove one component at a time but do not cross reformulation and knowledge selection; an experiment that turns both off and on independently would show whether the gains are additive or one component carries the effect.
  • Editorial inference: the metric rewards staying inside available evidence, so a companion measure of informativeness or query coverage is needed to distinguish genuine factuality gains from evasive, evidence-safe responses.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes two graph knowledge-augmented frameworks for dialogue response generation: TG-DRG (unsupervised, using textualised graphs) and GA-DRG (supervised, using a GNN-based graph encoder). Both are built around a chain-of-thought-driven dialogue reformulation module and a dialogue sense knowledge selection module. The paper also introduces a Dialogue Fact Score and a Not Enough Information Proportion (NEIP) metric for evaluating factuality in dialogue, and reports experiments on OpendialKG and HybriDialogue comparing against BM25, KAPING, G-Retriever, and several standalone LLMs. The authors claim that their frameworks, especially GA-DRG, noticeably improve factual accuracy over the baselines, including the state-of-the-art G-Retriever, based on the new metric.

Significance. If the reported factuality improvements are reliable, the proposed frameworks offer a practical combination of coreference resolution, knowledge selection, and graph-aware generation for grounding dialogue responses, and the Dialogue Fact Score addresses a real gap by incorporating dialogue context and handling unverifiable claims. The paper provides detailed prompts, data collection procedures, and an ablation study that supports the contribution of each proposed component, which aids reproducibility. However, the central claim currently rests on a metric whose validation is partially circular and on single-run experiments with no variance reporting, so the significance is contingent on additional evidence that the metric reflects human factuality judgments and that the improvements are stable.

major comments (4)
  1. [§3.4, Appendix E, Table 1] The validation of the Dialogue Fact Score is not independent of the metric itself. Appendix E states that the human annotators were given the Dialogue Fact Score prompt (Table 8) as their instruction, so the reported agreement in Table 1 measures whether LLaMA-3.3-70B follows the authors' rubric, not whether the rubric captures what humans regard as factual. No human system-level factuality ranking or comparison against existing factuality metrics (e.g., FactScore) is provided. Since the primary claim in Table 2 is based on this metric, the central conclusion is currently supported only by a self-validated evaluation instrument.
  2. [§4.5] The experimental setup states that "all experiments were run a single time," and Table 2 reports no standard deviations, confidence intervals, or significance tests. The headline differences, such as 87.73 vs. 84.79 on OpendialKG and 73.00 vs. 70.29 on HybriDialogue, could plausibly fall within run-to-run noise for LLM fine-tuning and prompting. The authors should provide multiple runs or at least significance testing before claiming that their methods "noticeably improve" over G-Retriever.
  3. [Abstract and Introduction] The abstract claims improvements of 3.47% on OpendialKG and 3.12% on HybriDialogue in terms of dialogue fact score. From Table 2, GA-DRG vs. G-Retriever on OpendialKG is (87.73 - 84.79)/84.79 = 3.47%, which matches, but on HybriDialogue it is (73.00 - 70.29)/70.29 = 3.85%, not 3.12%. The 3.12% figure appears in the Introduction as well and is not reproducible from the reported data. This inconsistency should be corrected or reconciled.
  4. [§3.2, §3.4, Appendix H Case 3] The evaluation metric is aligned with the same knowledge source used by the generation frameworks: the knowledge selection module retrieves Wikidata triples, and the Dialogue Fact Score verifies atomic facts against Wikipedia passages retrieved from entities in the response. Appendix H Case 3 shows TG-DRG-NR generating a verbatim triple, "Montevideo, population, 309331," which would be scored as verifiable by the metric despite being an ungrammatical and non-fluent response. This illustrates that responses echoing the supplied knowledge can inflate the Dialogue Fact Score without constituting good dialogue, so the reported gains may partly reflect measurement bias rather than a genuine improvement in dialogue factuality.
minor comments (5)
  1. [§4.2, Table 4] The human evaluation covers coherence, fluency, and informativeness but does not include a factuality judgment, so it does not directly test the paper's main claim. Adding a human factuality assessment, even on a subsample, would strengthen the evaluation.
  2. [§3.3, §4.5] The base LLM used for GA-DRG response generation is not explicitly stated. Section 4.5 says G-retriever is implemented with Llama2-7B and LoRA and that "we adopt the same model under identical settings to ensure fairness," but it is unclear whether GA-DRG also uses Llama2-7B. Please clarify the generation backbone for GA-DRG.
  3. [Appendix F] The entity linker "RefinED" (also spelled "REfinED") is used for the BM25 and KAPING baselines, but no citation is provided. Please add the appropriate reference.
  4. [Table 2 caption] The caption states "The bold number is the best result within each category," but this is ambiguous for NEIP, where lower is better. Please clarify whether bold indicates the best value (e.g., lowest NEIP) or simply all results from the proposed methods.
  5. [Appendix B] "HybriDial follows the MIT license" should be "HybriDialogue follows the MIT license" for consistency with the dataset name.

Circularity Check

1 steps flagged · score 5.0 of 10

The primary outcome measure (Dialogue Fact Score) is validated by comparing LLaMA against annotators instructed with the metric's own prompt; the reported factuality gains are therefore measured on a self-validated rubric.

  1. self definitional [Section 3.4 (Dialogue Fact Score), Table 1, Appendix E, Table 8]
    "Three annotators were involved in the assessment of the dialogue fact score task. We used the dialogue fact score prompt, described in Table 8, as an instruction. Two annotators independently assessed the dialogue fact score for 100 samples from the HybriDial and OpendialKG datasets, respectively. ... Finally, we report the final agreement and Cohen’s Kappa score with the evaluation model, as shown in Table 1."

    The claimed validation ('our dialogue fact score achieves substantial model-human agreement', Section 3.4) compares LLaMA-3.3-70B's application of the Table 8 prompt with humans' application of the same prompt. The human 'ground truth' is therefore generated by the metric's own definition, not by an independent judgment of factuality. Since Table 2's Fact* columns (the 'primary metrics', Table 2 caption) are the sole basis for the headline 3.47%/3.12% improvements, the central empirical claim rests on a self-defined yardstick whose only validation is self-application. This does not reduce the generation equations to the metric, but it makes the reported improvement evidence partially circular.

full rationale

The generation pipeline itself is not circular: TG-DRG and GA-DRG combine CoT reformulation, triple selection, and graph-conditioned generation, and none of their training objectives (e.g., Eq. 11 cross-entropy loss) is fitted to the Dialogue Fact Score. However, the paper's central factual claim is measured entirely by the newly proposed Dialogue Fact Score (Table 2, 'Fact*' columns), and the only validation of that metric is the model-human agreement in Table 1. Appendix E reveals that the human annotators were given the Dialogue Fact Score prompt (Table 8) as their instruction, so the 'ground truth' and the machine score are two applications of the same rubric. The reported 'substantial agreement' therefore evidences instruction-following consistency, not that the rubric matches an independent human assessment of factuality. The paper also does not compare the metric against existing factuality metrics on generated outputs, and Table 4's human evaluation covers coherence, fluency, and informativeness but not factuality, so no external human system-level ordering supports the headline gain. There is also a knowledge-alignment concern: the generation frameworks retrieve Wikidata triples and the metric verifies against Wikipedia passages retrieved from response entities; Case 3 shows a verbatim triple ('Montevideo, population, 309331') being counted as a generated response, illustrating that the metric can reward echoing the supplied knowledge. This alignment is a measurement-bias risk rather than a formal equation-level reduction, so I do not count it as a separate circular step. The numeric mismatch between the Abstract's 3.12% HybriDialogue improvement and Table 2's 2.71 absolute points (3.85% relative) and the single-run experiments are correctness/robustness concerns, not circularity. Overall: one load-bearing self-definitional validation makes the evidence for the central claim partially circular, scoring 5.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a large set of unstated hyperparameters, the correctness of external knowledge (Wikidata, Wikipedia), and the quality of GPT-4o generated training data. The evaluation metric itself is new and shares the knowledge source with generation.

free parameters (5)
  • Number of selected knowledge triples N
    Equation (3) selects top-N triples, but N is never stated in the experiments, so it is a hand-chosen value that affects generation.
  • Graph node embedding dimension d
    Equation (5) uses dimension d, not reported.
  • LoRA rank and alpha for GA-DRG = rank=8, alpha=16
    Section 4.5: GNN settings use rank 8, alpha 16; chosen by hand, not ablated.
  • LoRA rank and alpha for reformulation and knowledge selection = rank=32, alpha=32
    Section 4.5: LoRA rank and alpha are consistently set to 32; chosen by hand, not ablated.
  • Text embedding layer choice
    Equation (9) uses the TextEmb of an LLM, but the specific model is not named.
assumptions (3)
  • domain assumption Wikidata provides correct and current factual triples for knowledge grounding
    Section B: triples are re-extracted from Wikidata and treated as ground truth for factuality.
  • domain assumption Wikipedia passages retrieved by entity matching are sufficient evidence for fact verification
    Section 3.4: entities are turned into Wikipedia titles to retrieve supporting passages; the completeness and correctness of Wikipedia are assumed.
  • domain assumption GPT-4o generated reformulation and relevance labels are of sufficient quality for fine-tuning
    Section 3.2 and Appendix C: fine-tuning data for dialogue reformulation and knowledge selection is generated by GPT-4o, with only 100 samples manually checked.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Factuality for Dialogue Response Generation via Graph-Based Knowledge Augmentation." pith.science (2026). https://pith.science/paper/G6QRGPSG

@misc{pith2026250612496,
  author       = {Pith},
  title        = {Pith review of: Improving Factuality for Dialogue Response Generation via Graph-Based Knowledge Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G6QRGPSG}},
  note         = {Machine review of arXiv:2506.12496}
}
read the original abstract

Large Language Models (LLMs) succeed in many natural language processing tasks. However, their tendency to hallucinate - generate plausible but inconsistent or factually incorrect text - can cause significant problems in certain tasks, including response generation in dialogue. To mitigate this issue, we propose two novel graph knowledge-augmented frameworks, Dialogue Response Generation via Textualised Graphs (TG-DRG) and Graph-Aware Dialogue Response Generation (GA-DRG), which combine reasoning-guided dialogue reformulation, dialogue sense knowledge selection, and graph-enhanced response generation to improve the factuality of dialogue responses. To evaluate the factuality of generated responses, we propose a dialogue fact score that addresses the limitations of existing fact-score methods in dialogue settings, providing a more reliable assessment of factual consistency. We evaluate our methods using different baselines on the OpendialKG and HybriDialogue datasets. Our methods noticeably improve factuality compared to other graph knowledge-augmentation baselines, including the state-of-the-art G-retriever, achieving improvements of 3.47% on OpendialKG and 3.12% on HybriDialogue in terms of dialogue fact score. The code will be released on GitHub.

Figures

Figures reproduced from arXiv: 2506.12496 by the authors.

Figure 1
Figure 1. An example illustrates how our frameworks TG [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The workflow of our proposed frameworks, TG-DRG and GA-DRG, begins with the dialogue. A CoT process first [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 11 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  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]

    Agrawal, G.; Kumarage, T.; Alghami, Z.; and Liu, H. 2023. Can knowledge graphs reduce hallucinations in llms?: A survey. arXiv preprint arXiv:2311.07914

    arXiv 2023

  4. [4]

    Asai, A.; Wu, Z.; Wang, Y.; Sil, A.; and Hajishirzi, H. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations

    work page 2023

  5. [5]

    F.; and Saffari, A

    Baek, J.; Aji, A. F.; and Saffari, A. 2023. Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. arXiv preprint arXiv:2306.04136

    arXiv 2023

  6. [6]

    Bast, H.; B \"a urle, F.; Buchhold, B.; and Hau mann, E. 2014. Easy access to the freebase dataset. In Proceedings of the 23rd international conference on World Wide Web, 95--98

    work page 2014

  7. [7]

    W.; Hou, L.; Longpre, S.; Zoph, B.; Tay, Y.; Fedus, W.; Li, Y.; Wang, X.; Dehghani, M.; Brahma, S.; et al

    Chung, H. W.; Hou, L.; Longpre, S.; Zoph, B.; Tay, Y.; Fedus, W.; Li, Y.; Wang, X.; Dehghani, M.; Brahma, S.; et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416

    arXiv 2022

  8. [8]

    Du, Z.; Qian, Y.; Liu, X.; Ding, M.; Qiu, J.; Yang, Z.; and Tang, J. 2021. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360

    arXiv 2021

Show all 33 references
  1. [9]

    Farquhar, S.; Kossen, J.; Kuhn, L.; and Gal, Y. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017): 625--630

    2024

  2. [10]

    He, X.; Tian, Y.; Sun, Y.; Chawla, N.; Laurent, T.; LeCun, Y.; Bresson, X.; and Hooi, B. 2024. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. Advances in Neural Information Processing Systems, 37: 132876--132907

    2024

  3. [11]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  4. [12]

    Huang, L.; Yu, W.; Ma, W.; Zhong, W.; Feng, Z.; Wang, H.; Chen, Q.; Peng, W.; Feng, X.; Qin, B.; et al. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232

    2023 arXiv

  5. [13]

    L.; Bahl, L

    Jelinek, F.; Mercer, R. L.; Bahl, L. R.; and Baker, J. K. 1977. Perplexity—a measure of the difficulty of speech recognition tasks. The Journal of the Acoustical Society of America, 62(S1): S63--S63

    1977

  6. [14]

    Kipf, T. 2016. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907

    2016 arXiv

  7. [15]

    R.; and Wang, L

    Li, Y.; Zhao, J.; Lyu, M. R.; and Wang, L. 2022. Eliciting knowledge from large pre-trained models for unsupervised knowledge-grounded conversation. arXiv preprint arXiv:2211.01587

    2022 arXiv

  8. [16]

    Lin, C.-Y. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 74--81

    2004

  9. [17]

    W.; Iyyer, M.; Zettlemoyer, L.; and Hajishirzi, H

    Min, S.; Krishna, K.; Lyu, X.; Lewis, M.; Yih, W.-t.; Koh, P. W.; Iyyer, M.; Zettlemoyer, L.; and Hajishirzi, H. 2023. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation. arXiv preprint arXiv:2305.14251

    2023 arXiv

  10. [18]

    Moon, S.; Shah, P.; Kumar, A.; and Subba, R. 2019. Opendialkg: Explainable conversational reasoning with attention-based walks over knowledge graphs. In Proceedings of the 57th annual meeting of the association for computational linguistics, 845--854

    2019

  11. [19]

    Nakamura, K.; Levy, S.; Tuan, Y.-L.; Chen, W.; and Wang, W. Y. 2022. HybriDialogue: An information-seeking dialogue dataset grounded on tabular and textual data. arXiv preprint arXiv:2204.13243

    2022 arXiv

  12. [20]

    Nan, F.; Nallapati, R.; Wang, Z.; Santos, C. N. d.; Zhu, H.; Zhang, D.; McKeown, K.; and Xiang, B. 2021. Entity-level factual consistency of abstractive text summarization. arXiv preprint arXiv:2102.09130

    2021 arXiv

  13. [21]

    Ni, X.; Dai, H.; Ren, Z.; and Li, P. 2023. Multi-source multi-type knowledge exploration and exploitation for dialogue generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 12522--12537

    2023

  14. [22]

    Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; Sutskever, I.; et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8): 9

    2019

  15. [23]

    Reiter, E. 2018. A structured review of the validity of BLEU. Computational Linguistics, 44(3): 393--401

    2018

  16. [24]

    Robertson, S.; Zaragoza, H.; et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval , 3(4): 333--389

    2009

  17. [25]

    Sen, P.; Mavadia, S.; and Saffari, A. 2023. Knowledge graph-augmented language models for complex question answering. In Proceedings of the 1st Workshop on Natural Language Reasoning and Structured Explanations (NLRSE), 1--8

    2023

  18. [26]

    Shi, Y.; Huang, Z.; Feng, S.; Zhong, H.; Wang, W.; and Sun, Y. 2020. Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509

    2020 arXiv

  19. [27]

    Touvron, H.; Lavril, T.; Izacard, G.; Martinet, X.; Lachaux, M.-A.; Lacroix, T.; Rozi \`e re, B.; Goyal, N.; Hambro, E.; Azhar, F.; et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  20. [28]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  21. [29]

    Wu, Y.; Hu, N.; Qi, G.; Bi, S.; Ren, J.; Xie, A.; and Song, W. 2023. Retrieve-rewrite-answer: A kg-to-text enhanced llms framework for knowledge graph question answering. arXiv preprint arXiv:2309.11206

    2023 arXiv

  22. [30]

    Yan, S.-Q.; Gu, J.-C.; Zhu, Y.; and Ling, Z.-H. 2024. Corrective retrieval augmented generation

    2024

  23. [31]

    Yu, J.; Zhang, X.; Xu, Y.; Lei, X.; Guan, X.; Zhang, J.; Hou, L.; Li, J.; and Tang, J. 2022. XDAI: A tuning-free framework for exploiting pre-trained language models in knowledge grounded dialogue generation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discov...

    2022

  24. [32]

    Zeng, A.; Liu, X.; Du, Z.; Wang, Z.; Lai, H.; Ding, M.; Yang, Z.; Xu, Y.; Zheng, W.; Xia, X.; et al. 2022. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414

    2022 arXiv

  25. [33]

    Zhang, Y.; Li, Y.; Cui, L.; Cai, D.; Liu, L.; Fu, T.; Huang, X.; Zhao, E.; Zhang, Y.; Chen, Y.; et al. 2023. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219

    2023 arXiv

Pith tools

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