Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation

T0 review · 4 major / 7 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 FaithfulRAG claims that faithful RAG should resolve, not suppress, the conflict between a model's stored knowledge and retrieved context.

desk verdict Solid RAG faithfulness paper with a real trade-off finding, but the fact-level alignment mechanism is under-validated. read the letter →

arxiv 2506.08938 v2 pith:7DNNKX67 submitted 2025-06-10 cs.CL

Qinggang Zhang , Zhishang Xiang , Yilin Xiao , Le Wang , Junhui Li , Xinrun Wang , Jinsong Su This is my paper
classification cs.CL
keywords retrieval-augmentedgenerationknowledgeconflictfactualfaithfulnessself-factminingcontextualalignmentself-thinkparametricRAG
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

Retrieval-augmented language models often fail when the retrieved context contradicts what the model already knows. The paper argues that existing faithfulness methods—prompting the model to obey context or modifying decoding—buy faithfulness by suppressing the model's parametric knowledge, which increases the chance that the model blindly adopts errors in the context. FaithfulRAG instead tries to keep both knowledge sources active and reconcile them explicitly: it externalizes the model's own understanding of the question as 'self-facts', aligns those facts against chunks of the retrieved context by embedding similarity, and then asks the model to think through the conflicts before answering. The reported result is state-of-the-art accuracy on FaithEval, RealtimeQA, MuSiQue, and SQuAD, with a simultaneous reduction in over-confidence errors and incorrect-match errors. A sympathetic reader would take the paper's central claim to be that diagnosing conflicts at the fact level is both possible and strictly better than suppressing the model's prior knowledge.

What carries the argument

The load-bearing mechanism is a three-stage pipeline. Self-Fact Mining externalizes the model's parametric understanding of the question: it first extracts high-level knowledge needs, generates a 'self-context' from them, then distills that into atomic self-facts $f_1,\ldots,f_m$. Contextual Knowledge Alignment chunks the retrieved context and scores each chunk against each self-fact with cosine similarity $\mathrm{Sim}(f_i,c_j)=\cos(f_i,c_j)$, keeping the top-$k$ aligned chunks. Self-Think then reasons over those aligned chunks together with the original context, first judging whether the aligned chunks suffice, then fusing in missing context to produce a final answer. The identity that carries the argument is the alignment score: it is the point where the model's own facts are matched to the retrieved text, and everything downstream depends on it.

What would settle it

Take a sample of conflicting examples with known gold conflict spans and measure whether the embedding-based top-$k$ alignment recovers those spans at rates above chance; if recall is at chance or if swapping the embedding model for a random chunk selector leaves accuracy unchanged, the alignment step is not carrying the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that faithfulness and accuracy need not trade off once conflicts are handled at the fact level rather than by global suppression. Concretely, the paper reports that prompting-based and decoding-based faithful methods reduce over-confidence errors (the model sticks to its parametric answer) by about 6 percent on average but raise incorrect-match errors (the model adopts a wrong contextual claim) by a similar amount; FaithfulRAG is claimed to reduce over-confidence errors by 6.8 percent and incorrect-match errors by 1.6 percent on average, and to outperform the strongest baselines by 8.5 to 9.3 points on FaithEval and SQuAD. The method also reports gains in non-conflict settings, where suppression-based baselines often lose accuracy. The authors frame this as evidence that an LLM can act as a critical reasoner that keeps its parametric knowledge and the retrieved context in view at once.

Load-bearing premise

The method's conflict detection assumes that cosine similarity between embeddings of the model's own facts and chunks of the retrieved context reliably surfaces the spans where the conflict actually lies; if the embedding model misses those spans, the self-think step starts from the wrong evidence and the whole reconciliation collapses.

Editorial extensions

If this is right

  • Faithfulness methods should stop treating the model's parametric knowledge as noise to be suppressed; keeping it active is what allows conflicting context to be evaluated rather than rubber-stamped.
  • Because the method is prompt-level and does not retrain, it can be applied to any instruction-following LLM, and the paper demonstrates gains on 3B, 7B, 8B, 13B, and 16B backbones.
  • Fact-level conflict modeling should produce balanced error reduction: the reported error analysis shows both over-confidence and incorrect-match errors falling together, unlike prompting or decoding baselines.
  • Non-conflict performance need not suffer: the paper reports accuracy gains on golden-context versions of MuSiQue and SQuAD, where suppression-based methods often drop.

Reading between the lines

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

  • A direct test of the core mechanism is to measure whether the cosine-similarity top-$k$ chunks actually contain the sentences that contradict the model's own facts; if recall of conflict-bearing chunks is no better than chance, the reported gains would have to come from the self-think prompt rather than the alignment step.
  • The same self-fact, alignment, and self-think recipe could be ported to multimodal RAG, where conflicts between text and image or audio evidence are common; the paper itself notes this as a limitation.
  • One could ablate more finely by replacing the embedding model with a deliberately poor retriever; the paper's own small embedding-model comparison suggests the results are fairly stable, which indicates the alignment step is not brittle within the tested range.
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 / 7 minor

Summary. The paper studies knowledge conflicts in retrieval-augmented generation (RAG), where retrieved context contradicts the LLM's parametric knowledge. It argues that existing context-faithful methods improve faithfulness by suppressing parametric knowledge, which increases 'incorrect-match' errors. The proposed FaithfulRAG framework externalizes the model's parametric knowledge into self-facts, aligns these self-facts with retrieved context chunks using cosine similarity, and applies a Self-Think module that reasons over the aligned chunks and the original context before generating an answer. Experiments on the Counterfactual subset of FaithEval, RealtimeQA, MuSiQue, and SQuAD with three backbone LLMs report accuracy improvements over several prompting- and decoding-based baselines, along with an error analysis showing simultaneous reduction of over-confidence and incorrect-match errors.

Significance. If the mechanism is as claimed, the paper offers a useful alternative to suppression-based faithfulness methods: instead of weakening parametric knowledge, it attempts to localize and reason about conflicts at the fact level. The empirical scope is a strength: four datasets, three backbone models, additional model-size experiments, MR-based faithfulness evaluation, and an informative ablation study. The authors also release code and use deterministic decoding, which aids reproducibility. However, the central mechanism — that cosine-similarity-based chunk selection identifies the actual conflict-bearing spans — is not validated, and some headline claims are overbroad given the reported numbers. These issues are fixable with additional analysis, but they currently prevent the paper's central contribution from being fully supported.

major comments (4)
  1. [Section 4.2, Eq. (6), Algorithm 1] The paper's core claim is that FaithfulRAG resolves knowledge conflicts at the fact level. That claim rests on the Contextual Knowledge Alignment module selecting chunks Caligned that contain the conflicting facts. However, the paper never verifies that the top-k chunks selected by cosine similarity are actually the conflict-bearing spans. Equation (6) measures topical similarity between self-facts and context chunks, not contradiction or conflict. Table 8 only shows that final accuracy is insensitive to the choice of embedding model, and the 'w/o whole Module' ablation in Table 3 shows that prepending self-aligned context matters, but it does not show that the selected chunks are the conflicting ones. The authors should report, for example, recall@k of the selected chunks against the gold or substituted spans in KRE, or a human-annotated evaluation on FaithEval, and should compare against random chunk selection. Without this evidence, the improvement could come from reordering or emphasizing any topically related context rather than from fact-level conflict localization.
  2. [Section 5.2, Obs. 1, Table 1] The observation that 'FaithfulRAG consistently outperforms baseline models on four benchmark datasets' is contradicted by the reported numbers. On RealtimeQA with Qwen2.5-7B, KRE achieves 86.7 while FaithfulRAG achieves 84.1. The sentence in Obs. 1 stating that FaithfulRAG 'dominates RealtimeQA with scores of 79.9% and 84.1%' is therefore inaccurate for at least one backbone. The authors should either qualify the claim (e.g., by reporting average rank across backbones or noting per-backbone exceptions) or explain why the RealtimeQA result is not considered a counterexample. This is a load-bearing claim because the abstract and introduction describe the method as outperforming state-of-the-art methods across the board.
  3. [Section 3, Figure 2, Section 5.3] The causal claim that existing faithful methods 'forcibly suppress the model's parametric knowledge' is inferred from error distributions but is not directly measured. The observed increase in Case 2 (incorrect-match) errors is consistent with that explanation, but it is also consistent with other mechanisms, such as instruction-following biases or decoding-level shifts that do not explicitly modulate parametric knowledge. Since this causal story motivates the entire framework, the authors should provide direct evidence, for instance by measuring the model's confidence in or probability assigned to the parametric answer under each method, or by performing an intervention where the context is removed and comparing how each method behaves. Without such evidence, the 'suppression' narrative remains an assumption rather than a measured effect.
  4. [Section 5.2, Table 1, Section C.1] The paper reports no measures of variance or statistical significance, and the RealtimeQA test set appears to contain only six questions. Some of the accuracy differences that support the state-of-the-art claim are small (e.g., 77.0 vs. 78.0 on one RealtimeQA row), and the FaithEval results are only for the Counterfactual subset even though the tables and abstract label the dataset as 'FaithEval'. The authors should report the subset size, state explicitly in the main text that only the Counterfactual subset is used, and ideally provide paired significance tests or bootstrap confidence intervals across questions. This would strengthen the reliability of the empirical comparisons, especially where margins are small.
minor comments (7)
  1. [Abstract and Section 5.1] The abstract and Table 1 refer to 'FaithEval' without noting that only the Counterfactual subset is used; this should be clarified at the first mention in the main text.
  2. [Section C.1] The RealtimeQA-22 test set is described as selecting six questions; this very small test size should be reported near Table 1, because it affects how the RealtimeQA numbers should be interpreted.
  3. [Section 4.2 and Algorithm 1] It is not clear whether top-k chunks are selected per self-fact or globally across all self-facts, nor how ties are broken; the selection procedure should be specified precisely.
  4. [Section 5.4] The heading 'Alation Study' contains a typo and should read 'Ablation Study'.
  5. [Section 2] The phrase 'Knowldge Retrieval' contains a typo and should read 'Knowledge Retrieval'.
  6. [Appendix B.2] Case 3 is defined only as 'all other scenarios beyond these two cases'; providing one or two concrete examples would make the error taxonomy clearer.
  7. [Section 5.2, Obs. 2] The claim that KRE and CAD exhibit 'severe instability' is based on a small number of backbones; the refusal cases marked with an asterisk should be discussed separately from accuracy numbers, since a refusal is not the same as an incorrect answer.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FaithfulRAG is an empirical pipeline whose components are neither fitted to nor defined by the target benchmark results, and no load-bearing self-citation chain is present.

full rationale

FaithfulRAG is an empirical pipeline rather than a formal derivation, and I find no step in which a claimed output is equivalent by construction to an input. The three modules—self-fact mining (Eqs. 2-4), contextual knowledge alignment (Eqs. 5-6), and self-think (Eqs. 7-8)—are compositions of prompting, embedding similarity, and fixed top-k selection; chunk size (20) and k (5) are fixed hyperparameters, not fitted to the test sets. The central claims are evaluated by ACC and MR on FaithEval, RealtimeQA, MuSiQue, and SQuAD against external baselines, so the headline result is not a renamed or refitted input. The preliminary error analysis in Section 3 and Figure 2 defines Case 1 and Case 2 by comparing parametric and contextual predictions, but this is an evaluation taxonomy, not a fitted quantity later renamed as a prediction. Citations to the authors' own prior work (e.g., KnowGPT and the retrieval-augmented generation survey) appear as background context and are not load-bearing for the paper's claimed novelty or effectiveness. The skeptic's concern—that cosine similarity may not select the truly conflict-bearing spans—is a genuine correctness and robustness risk, but it is not circularity: selecting top-k chunks neither guarantees nor presupposes the final answer, and the Self-Think module retains the original context as a fallback. No uniqueness theorem, self-citation chain, or definitional identification forces the method's choice. The paper is therefore self-contained with respect to its empirical claims, and the appropriate circularity score is 0.

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

The central claim rests on the availability and reliability of the model's parametric knowledge, semantic similarity for alignment, and the effectiveness of the multi-step prompting; no new entities or fitted scientific parameters are introduced.

free parameters (3)
  • chunk_size = 20
    Fixed-size chunking of the original context; chosen by hand and not swept in the paper.
  • top_k = 5
    Number of context chunks selected per self-fact; set by default and not ablated.
  • embedding_model = all-MiniLM-L6-v2
    Default embedding model for similarity matching; alternatives in Appendix B.5 give comparable results.
assumptions (3)
  • domain assumption Self-facts extracted from the LLM's own parametric knowledge are reliable anchors for identifying conflicting context.
    The entire alignment module depends on this. If self-facts are wrong or incomplete, alignment will select irrelevant chunks (Section 4.1, Equation 4).
  • domain assumption Cosine similarity between sentence embeddings of facts and context chunks measures factual alignment.
    Used in Equation 6 to select Caligned; no validation that high cosine similarity corresponds to conflicting facts.
  • domain assumption The LLM can accurately externalize its knowledge through the three-stage self-fact mining prompts.
    The self-context generation and fact extraction steps assumed to produce fine-grained facts without hallucination (Section 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/7DNNKX67

@misc{pith2026250608938,
  author       = {Pith},
  title        = {Pith review of: FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7DNNKX67}},
  note         = {Machine review of arXiv:2506.08938}
}
read the original abstract

Large language models (LLMs) augmented with retrieval systems have demonstrated significant potential in handling knowledge-intensive tasks. However, these models often struggle with unfaithfulness issues, generating outputs that either ignore the retrieved context or inconsistently blend it with the LLM`s parametric knowledge. This issue is particularly severe in cases of knowledge conflict, where the retrieved context conflicts with the model`s parametric knowledge. While existing faithful RAG approaches enforce strict context adherence through well-designed prompts or modified decoding strategies, our analysis reveals a critical limitation: they achieve faithfulness by forcibly suppressing the model`s parametric knowledge, which undermines the model`s internal knowledge structure and increases the risk of misinterpreting the context. To this end, this paper proposes FaithfulRAG, a novel framework that resolves knowledge conflicts by explicitly modeling discrepancies between the model`s parametric knowledge and retrieved context. Specifically, FaithfulRAG identifies conflicting knowledge at the fact level and designs a self-thinking process, allowing LLMs to reason about and integrate conflicting facts before generating responses. Extensive experiments demonstrate that our method outperforms state-of-the-art methods. The code is available at https://github.com/DeepLearnXMU/Faithful-RAG

Figures

Figures reproduced from arXiv: 2506.08938 by the authors.

Figure 1
Figure 1. The running example of knowledge conflict [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The cases of errors and their distribution on [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overall pipeline of our FaithfulRAG framework. FaithfulRAG first designs a self-fact mining module [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The error distribution on MuSiQue and SQuAD datasets with Llama3.1-8b-instruct as the backbone LLM. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The error distribution on MuSiQue and SQuAD datasets with Llama3.1-8b-instruct as the backbone LLM. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Prompts for self-knowledge extraction, self-context generation, self-fact extraction, and the self-think. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Where Facts Go Missing: A Layerwise Taxonomy and Per-Layer Attribution of Information Omission in Air-Gapped LLMAgent Pipelines

    cs.MA 2026-07 conditional novelty 6.0 of 10

    In a controlled 75,476-trial stress test, about 73% of omitted-fact failures in LLM agent pipelines are traced to deterministic middleware (redaction, pagination, truncation) rather than model behavior.

  2. 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

45 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    AI Anthropic. 2024. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card

    2024

  2. [2]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. https://arxiv.org/abs/2310.11511 Self-rag: Learning to retrieve, generate, and critique through self-reflection . Preprint, arXiv:2310.11511

    arXiv 2023

  3. [3]

    Baolong Bi, Shaohan Huang, Yiwei Wang, Tianchi Yang, Zihan Zhang, Haizhen Huang, Lingrui Mei, Junfeng Fang, Zehao Li, Furu Wei, et al. 2024 a . Context-dpo: Aligning language models for context-faithfulness. arXiv preprint arXiv:2412.15280

    arXiv 2024

  4. [4]

    Baolong Bi, Shenghua Liu, Yiwei Wang, Lingrui Mei, Junfeng Fang, Hongcheng Gao, Shiyu Ni, and Xueqi Cheng. 2024 b . Is factuality enhancement a free lunch for llms? better factuality can lead to worse context-faithfulness. arXiv preprint arXiv:2404.00216

    work page Pith review arXiv 2024

  5. [5]

    Zhiwei Cao, Qian Cao, Yu Lu, Ningxin Peng, Luyang Huang, Shanbo Cheng, and Jinsong Su. 2024. https://arxiv.org/abs/2406.02376 Retaining key information under high compression ratios: Query-guided compressor for llms . Preprint, arXiv:2406.02376

    work page Pith review arXiv 2024

  6. [6]

    Shengyuan Chen, Qinggang Zhang, Junnan Dong, Wen Hua, Qing Li, and Xiao Huang. 2024. Entity alignment with noisy annotations from large language models. arXiv preprint arXiv:2405.16806

    work page Pith review arXiv 2024

  7. [7]

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, and Tat-Seng Chua. 2025. Alphaedit: Null-space constrained knowledge editing for language models. ICLR

    work page 2025

  8. [8]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

    arXiv 2023

Show all 45 references
  1. [9]

    Zijin Hong, Zheng Yuan, Hao Chen, Qinggang Zhang, Feiran Huang, and Xiao Huang. 2024. Knowledge-to-sql: Enhancing sql generation with data expert llm. arXiv preprint arXiv:2402.11517

    2024 arXiv

  2. [10]

    Cheng-Yu Hsieh, Yung-Sung Chuang, Chun-Liang Li, Zifeng Wang, Long T Le, Abhishek Kumar, James Glass, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna, et al. 2024. Found in the middle: Calibrating positional attention bias improves long context utilization. arXiv preprint arXiv:...

    2024 arXiv

  3. [11]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232

    2023 arXiv

  4. [12]

    Houcheng Jiang, Junfeng Fang, Ningyu Zhang, Guojun Ma, Mingyang Wan, Xiang Wang, Xiangnan He, and Tat-seng Chua. 2025. Anyedit: Edit any knowledge encoded in language models. ICML

    2025

  5. [13]

    Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A Smith, Yejin Choi, Kentaro Inui, et al. 2024. Realtime qa: what's the answer right now? Advances in Neural Information Processing Systems, 36

    2024

  6. [14]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Princip...

    2023

  7. [15]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/6b493230205f78...

    2020

  8. [16]

    Daliang Li, Ankit Singh Rawat, Manzil Zaheer, Xin Wang, Michal Lukasik, Andreas Veit, Felix Yu, and Sanjiv Kumar. 2023. https://doi.org/10.18653/v1/2023.findings-acl.112 Large language models with controllable working memory . In Findings of the Association for Computational L...

    2023 doi

  9. [17]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024 a . Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    2024 arXiv

  10. [18]

    Bingshuai Liu, Chenyang Lyu, Zijun Min, Zhanyu Wang, Jinsong Su, and Longyue Wang. 2024 b . https://arxiv.org/abs/2312.01714 Retrieval-augmented multi-modal chain-of-thoughts reasoning for large language models . Preprint, arXiv:2312.01714

    2024 arXiv

  11. [19]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024 c . https://doi.org/10.1162/tacl_a_00638 Lost in the middle: How language models use long contexts . Transactions of the Association for Computational Linguistics,...

    2024 doi

  12. [20]

    Zihan Liu, Wei Ping, Rajarshi Roy, Peng Xu, Chankyu Lee, Mohammad Shoeybi, and Bryan Catanzaro. 2024 d . Chatqa: Surpassing gpt-4 on conversational qa and rag. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  13. [21]

    Shayne Longpre, Kartik Perisetla, Anthony Chen, Nikhil Ramesh, Chris DuBois, and Sameer Singh. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.565 Entity-based knowledge conflicts in question answering . In Proceedings of the 2021 Conference on Empirical Methods in Natural L...

    2021 doi

  14. [22]

    Yifei Ming, Senthil Purushwalkam, Shrey Pandit, Zixuan Ke, Xuan-Phi Nguyen, Caiming Xiong, and Shafiq Joty. 2025. https://openreview.net/forum?id=UeVx6L59fg Faitheval: Can your language model stay faithful to context, even if ''the moon is made of marshmallows'' . In The Thirt...

    2025

  15. [23]

    Xuan-Phi Nguyen, Shrey Pandit, Senthil Purushwalkam, Austin Xu, Hailin Chen, Yifei Ming, Zixuan Ke, Silvio Savarese, Caiming Xong, and Shafiq Joty. 2024. Sfr-rag: Towards contextually faithful llms. arXiv preprint arXiv:2409.09916

    2024 arXiv

  16. [24]

    OpenAI. 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . Preprint, arXiv:2303.08774

    2023 arXiv

  17. [25]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. https://arxiv.org/abs/1606.05250 Squad: 100,000+ questions for machine comprehension of text . Preprint, arXiv:1606.05250

    2016 arXiv

  18. [26]

    Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen-tau Yih. 2023. Trusting your evidence: Hallucinate less with context-aware decoding. arXiv preprint arXiv:2305.14739

    2023 arXiv

  19. [27]

    Hexiang Tan, Fei Sun, Wanli Yang, Yuanzhuo Wang, Qi Cao, and Xueqi Cheng. 2024. Blinded by generated contexts: How language models merge generated and retrieved contexts when knowledge conflicts? In Proceedings of the 62nd Annual Meeting of the Association for Computational Li...

    2024

  20. [28]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. https://arxiv.org/abs/2108.00573 Musique: Multihop questions via single-hop question composition . Preprint, arXiv:2108.00573

    2022 arXiv

  21. [29]

    Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. 2024. Knowledge editing for large language models: A survey. ACM Computing Surveys, 57(3):1--37

    2024

  22. [30]

    Kevin Wu, Eric Wu, and James Zou. 2024. How faithful are rag models? quantifying the tug-of-war between rag and llms' internal prior. arXiv e-prints, pages arXiv--2404

    2024

  23. [31]

    Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, and Yu Su. 2024. https://openreview.net/forum?id=auKAUJZMO6 Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts . In The Twelfth International Conference on Learning Representations

    2024

  24. [32]

    Peng Xu, Wei Ping, Xianchao Wu, Chejian Xu, Zihan Liu, Mohammad Shoeybi, and Bryan Catanzaro. 2024 a . Chatqa 2: Bridging the gap to proprietary llms in long context and rag capabilities. arXiv preprint arXiv:2407.14482

    2024 arXiv

  25. [33]

    Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.486 Knowledge conflicts for LLM s: A survey . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p...

    2024 doi

  26. [34]

    Jiahao Ying, Yixin Cao, Kai Xiong, Long Cui, Yidong He, and Yongbin Liu. 2024. https://doi.org/10.18653/v1/2024.acl-long.232 Intuitive or dependent? investigating LLM s' behavior style to conflicting prompts . In Proceedings of the 62nd Annual Meeting of the Association for Co...

    2024 doi

  27. [35]

    Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. https://openreview.net/forum?id=fB0hRu9GZUS Generate rather than retrieve: Large language models are strong context generators . In The Eleventh Inter...

    2023

  28. [36]

    Xiaowei Yuan, Zhao Yang, Yequan Wang, Shengping Liu, Jun Zhao, and Kang Liu. 2024. https://doi.org/10.18653/v1/2024.findings-acl.234 Discerning and resolving knowledge conflicts through adaptive decoding with contextual information-entropy constraint . In Findings of the Assoc...

    2024 doi

  29. [37]

    Zheng Yuan, Hao Chen, Zijin Hong, Qinggang Zhang, Feiran Huang, and Xiao Huang. 2025. Knapsack optimization-based schema linking for llm-based text-to-sql generation. arXiv preprint arXiv:2502.12911

    2025 arXiv

  30. [38]

    Qinggang Zhang, Shengyuan Chen, Yuanchen Bei, Zheng Yuan, Huachi Zhou, Zijin Hong, Junnan Dong, Hao Chen, Yi Chang, and Xiao Huang. 2025. A survey of graph retrieval-augmented generation for customized large language models. arXiv preprint arXiv:2501.13958

    2025

  31. [39]

    Qinggang Zhang, Junnan Dong, Hao Chen, Daochen Zha, Zailiang Yu, and Xiao Huang. 2024. Knowgpt: Knowledge graph based prompting for large language models. Advances in Neural Information Processing Systems, 37:6052--6080

    2024

  32. [40]

    Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. 2023. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792

    2023

  33. [41]

    Chuang Zhou, Jiahe Du, Huachi Zhou, Hao Chen, Feiran Huang, and Xiao Huang. 2025. Text-attributed graph learning with coupled augmentations. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10865--10876

    2025

  34. [42]

    Huachi Zhou, Shuang Zhou, Hao Chen, Ninghao Liu, Fan Yang, and Xiao Huang. 2024. Enhancing explainable rating prediction through annotated macro concepts. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages ...

    2024

  35. [43]

    Wenxuan Zhou, Sheng Zhang, Hoifung Poon, and Muhao Chen. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.968 Context-faithful prompting for large language models . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14544--14556, Singapore. As...

    2023 doi

  36. [44]

    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...

  37. [45]

    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 gl...

Pith tools

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