arxiv: 2604.14172 · v1 · submitted 2026-03-25 · 💻 cs.CL · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Tug-of-War within A Decade: Conflict Resolution in Vulnerability Analysis via Teacher-Guided Retrieval-Augmented Generations

Ziyin Zhou , Jianyi Zhang , Xu ji , Yilong Li , Jiameng Han , Zhangchi Zhao

Authors on Pith no claims yet

Pith reviewed 2026-05-15 00:40 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords vulnerability analysisknowledge conflictsretrieval-augmented generationteacher-guided optimizationCVE detectionLLM fine-tuningconflict resolutionpreference optimization

0 comments

The pith

A two-stage teacher-guided RAG framework resolves knowledge conflicts for LLMs analyzing updated CVEs.

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

The paper tackles how LLMs produce inconsistent or fabricated answers about cybersecurity vulnerabilities because over 30,000 of the 200,000+ known issues have been revised in the past decade. It introduces CRVA-TGRAG, which first segments documents into parent-child structures and combines semantic and keyword retrieval to pull the most current records, then applies teacher-guided preference optimization to fine-tune the model toward accurate generations. This combination is meant to keep LLM outputs aligned with the latest CVE data instead of drifting into hallucinations from stale training knowledge. A sympathetic reader would see this as a way to make LLM-assisted security analysis more reliable without constant full retraining. The experiments claim measurable gains in retrieval accuracy for recent vulnerabilities over baseline external knowledge bases.

Core claim

The paper claims that the CRVA-TGRAG framework, built from Parent Document Segmentation, an ensemble retrieval scheme using semantic similarity and inverted indexing, and teacher-guided preference optimization, mitigates knowledge conflicts and inconsistencies that arise when LLMs rely solely on internal knowledge for CVE detection and analysis.

What carries the argument

Teacher-guided preference optimization applied after ensemble retrieval to steer LLM generations toward consistent, up-to-date CVE facts.

If this is right

LLMs achieve higher accuracy when retrieving the most recent CVEs compared to external knowledge bases alone.
Knowledge conflicts and factually incorrect generations decrease in vulnerability analysis tasks.
Models maintain better knowledge consistency across frequent CVE updates without requiring full retraining.
Answers to security questions become more precise by combining improved retrieval with preference-tuned generation.

Where Pith is reading between the lines

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

The same retrieval-plus-preference pattern could apply to other domains whose facts change rapidly, such as medical treatment guidelines.
Focusing updates on retrieval and light preference tuning may lower the compute cost of keeping LLMs current compared with full retraining cycles.
The method might combine with other RAG techniques to handle conflicting sources beyond CVE records.

Load-bearing premise

The assumption that teacher-guided preference optimization and ensemble retrieval will reliably resolve conflicts and improve accuracy without introducing new biases or overfitting to the CVE dataset.

What would settle it

Run the framework on a held-out collection of CVEs updated after the fine-tuning data cutoff and check whether retrieval accuracy falls below standard RAG baselines or whether the rate of fabricated details rises.

Figures

Figures reproduced from arXiv: 2604.14172 by Jiameng Han, Jianyi Zhang, Xu ji, Yilong Li, Zhangchi Zhao, Ziyin Zhou.

**Figure 2.** Figure 2: The pipeline of our framework. The user queries GPT-4o-mini, but due to internal knowledge conflicts, the returned results are [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: CVE vulnerabilities that have changed from 2014 to 2024. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Three different values of the split point percentiles. The different colors denote the split results and the redline indicates the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

Large Language Models (LLMs) are essential for analyzing and addressing vulnerabilities in cybersecurity. However, among over 200,000 vulnerabilities were discovered in the past decade, more than 30,000 have been changed or updated. This necessitates frequent updates to the training datasets and internal knowledge bases of LLMs to maintain knowledge consistency. In this paper, we focus on the problem of knowledge discrepancy and conflict within CVE (Common Vulnerabilities and Exposures) detection and analysis. This problem hinders LLMs' ability to retrieve the latest knowledge from original training datasets, leading to knowledge conflicts, fabrications of factually incorrect results, and generation hallucinations. To address this problem, we propose an innovative two-stage framework called CRVA-TGRAG (Conflict Resolution in Vulnerability Analysis via Teacher-Guided Retrieval-Augmented Generation). First, to improve document retrieval accuracy during the retrieval stage, we utilize Parent Document Segmentation and an ensemble retrieval scheme based on semantic similarity and inverted indexing. Second, to enhance LLMs' capabilities based on the retrieval of CVE dataset in generation stage, we employ a teacher-guided preference optimization technique to fine-tune LLMs. Our framework not only enhances the quality of content retrieval through RAG but also leverages the advantages of preference fine-tuning in LLMs to answer questions more effectively and precisely. Experiments demonstrate our method achieves higher accuracy in retrieving the latest CVEs compared to external knowledge bases. In conclusion, our framework significantly mitigates potential knowledge conflicts and inconsistencies that may arise from relying solely on LLMs for knowledge retrieval.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper describes a practical two-stage RAG and preference tuning approach for resolving CVE knowledge conflicts in LLMs, but without experimental details the claims remain unverified.

read the letter

The main point is a two-stage framework called CRVA-TGRAG for keeping LLMs accurate on updated CVE records. The first stage splits documents at the parent level and runs ensemble retrieval that mixes semantic similarity with inverted indexing. The second stage uses a teacher model to generate preference pairs and fine-tunes the generator on them. That specific sequence applied to CVE conflict resolution is new, even though the individual pieces are familiar from other RAG and tuning papers. The motivation is solid: more than 30,000 CVEs changed in the last decade, so static models quickly produce outdated or conflicting answers. The description of why parent segmentation and ensemble retrieval should improve coverage is clear and practical. The soft spot is the missing evidence. The abstract states that experiments show higher accuracy, yet it contains no numbers, no baselines, no description of how conflicts were measured, and no ablation of the teacher step. Without those details it is impossible to tell whether the gains come from better document retrieval or from the preference optimization actually fixing inconsistencies. The worry that the teacher could share the same cutoff knowledge is also left open. This work is aimed at people building retrieval systems for fast-changing technical domains such as security. A reader looking for concrete pipeline ideas could borrow the segmentation and ensemble tricks. I would bring it to a reading group to discuss the retrieval choices. It is not ready to cite as is, but it deserves peer review once the authors add the quantitative results and targeted conflict tests.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes CRVA-TGRAG, a two-stage framework for resolving knowledge conflicts and hallucinations in LLMs applied to CVE vulnerability analysis. Stage one improves retrieval via parent document segmentation plus an ensemble of semantic similarity and inverted-index methods. Stage two applies teacher-guided preference optimization to fine-tune the LLM on retrieved CVE data. The authors assert that experiments demonstrate higher accuracy in retrieving the latest CVEs relative to external knowledge bases, thereby mitigating inconsistencies arising from LLM knowledge cutoffs.

Significance. If the empirical claims are substantiated with proper controls, the work would address a concrete operational problem: maintaining factual consistency for LLMs on a rapidly changing corpus of >200k CVEs where >30k entries have been revised. The combination of retrieval engineering and preference tuning is a plausible practical response, though its incremental value over standard RAG pipelines remains to be quantified.

major comments (2)

[Abstract] Abstract: the statement that 'experiments demonstrate our method achieves higher accuracy in retrieving the latest CVEs' supplies no numerical results, no metrics (accuracy, precision@K, conflict-resolution rate, hallucination rate), no baselines, and no description of how knowledge conflicts were operationalized (e.g., contradictory CVE pairs, post-cutoff updates, or targeted hallucination probes). Without these elements the central empirical claim is unsupported.
[Abstract] Abstract: the teacher-guided preference optimization step presupposes that the teacher model itself is free of the same knowledge-cutoff discrepancies that affect the base LLM; the manuscript provides neither a selection criterion for the teacher nor any verification that preference pairs do not simply reinforce existing errors.

minor comments (1)

[Title] The title 'Tug-of-War within A Decade' is not referenced or explained in the abstract or provided text; a brief clarification of its relation to the technical contribution would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and have revised the manuscript to strengthen the presentation of empirical results and methodological details.

read point-by-point responses

Referee: [Abstract] Abstract: the statement that 'experiments demonstrate our method achieves higher accuracy in retrieving the latest CVEs' supplies no numerical results, no metrics (accuracy, precision@K, conflict-resolution rate, hallucination rate), no baselines, and no description of how knowledge conflicts were operationalized (e.g., contradictory CVE pairs, post-cutoff updates, or targeted hallucination probes). Without these elements the central empirical claim is unsupported.

Authors: We agree that the abstract should explicitly report quantitative results, metrics, and operational details to support the central claim. The full manuscript (Section 4) contains these elements, including accuracy improvements over baselines such as standard RAG and direct LLM generation, precision@K scores, and conflict operationalization via post-cutoff CVE updates as ground truth. We have revised the abstract to include key numerical findings (e.g., accuracy gains and hallucination rate reductions) and a brief description of the evaluation protocol. revision: yes
Referee: [Abstract] Abstract: the teacher-guided preference optimization step presupposes that the teacher model itself is free of the same knowledge-cutoff discrepancies that affect the base LLM; the manuscript provides neither a selection criterion for the teacher nor any verification that preference pairs do not simply reinforce existing errors.

Authors: This concern is valid. The original manuscript did not sufficiently detail teacher selection or verification. In the revision we specify that the teacher is a model with a later knowledge cutoff than the base LLM and the CVE update dates under study. We have added a new subsection describing the selection criterion and an explicit verification procedure in which generated preference pairs are cross-checked against an external up-to-date CVE database before use in fine-tuning, thereby reducing the risk of reinforcing cutoff-related errors. revision: yes

Circularity Check

0 steps flagged

No circularity: procedural pipeline with independent experimental support

full rationale

The paper presents CRVA-TGRAG as a two-stage procedural framework (parent document segmentation plus ensemble retrieval, followed by teacher-guided preference optimization) without any equations, derivations, fitted parameters, or self-referential definitions. No step reduces a claimed result to its own inputs by construction, and no load-bearing self-citation or uniqueness theorem is invoked. Experiments are described as demonstrating higher CVE retrieval accuracy, providing external falsifiability outside any internal fit. This matches the default expectation of a self-contained description.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review reveals no mathematical derivations, free parameters, axioms, or invented entities; the proposal rests on standard RAG and LLM fine-tuning techniques.

pith-pipeline@v0.9.0 · 5595 in / 1119 out tokens · 36290 ms · 2026-05-15T00:40:46.951486+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

two-stage framework called CRVA-TGRAG ... Parent Document Segmentation and an ensemble retrieval scheme ... teacher-guided preference optimization technique to fine-tune LLMs
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Experiments demonstrate our method achieves higher accuracy in retrieving the latest CVEs

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 5 internal anchors

[1]

GPT-4 Technical Report

[Achiamet al., 2023 ] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

work page internal anchor Pith review Pith/arXiv arXiv 2023
[2]

DeepSeek LLM: Scaling Open-Source Language Models with Longtermism

[Biet al., 2024 ] Xiao Bi, Deli Chen, Guanting Chen, Shan- huang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scal- ing open-source language models with longtermism.arXiv preprint arXiv:2401.02954,

work page internal anchor Pith review Pith/arXiv arXiv 2024
[3]

Language models are few-shot learners.Advances in neural information processing sys- tems, 33:1877–1901,

[Brownet al., 2020 ] Tom Brown, Benjamin Mann, Nick Ry- der, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing sys- tems, 33:1877–1901,

work page 2020
[4]

Say what you mean! large language models speak too positively about negative commonsense knowledge.arXiv preprint arXiv:2305.05976,

[Chenet al., 2023 ] Jiangjie Chen, Wei Shi, Ziquan Fu, Si- jie Cheng, Lei Li, and Yanghua Xiao. Say what you mean! large language models speak too positively about negative commonsense knowledge.arXiv preprint arXiv:2305.05976,

work page arXiv 2023
[5]

Time-aware language models as temporal knowledge bases.Transactions of the Associa- tion for Computational Linguistics, 10:257–273,

[Dhingraet al., 2022 ] Bhuwan Dhingra, Jeremy R Cole, Ju- lian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W Cohen. Time-aware language models as temporal knowledge bases.Transactions of the Associa- tion for Computational Linguistics, 10:257–273,

work page 2022
[6]

Pairing security ad- visories with vulnerable functions using open-source llms

[Dunlapet al., 2024 ] Trevor Dunlap, John Speed Meyers, Bradley Reaves, and William Enck. Pairing security ad- visories with vulnerable functions using open-source llms. InInternational Conference on Detection of Intrusions and Malware, and Vulnerability Assessment, pages 350–

work page 2024
[7]

Proverag: Provenance-driven vulnerability analysis with au- tomated retrieval-augmented llms.arXiv preprint arXiv:2410.17406,

[Fayyaziet al., 2024 ] Reza Fayyazi, Stella Hoyos Trueba, Michael Zuzak, and Shanchieh Jay Yang. Proverag: Provenance-driven vulnerability analysis with au- tomated retrieval-augmented llms.arXiv preprint arXiv:2410.17406,

work page arXiv 2024
[8]

Reducing the dimensionality of data with neural networks.science, 313(5786):504–507,

[Hinton and Salakhutdinov, 2006] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks.science, 313(5786):504–507,

work page 2006
[9]

Do large language models know about facts?arXiv preprint arXiv:2310.05177,

[Huet al., 2023 ] Xuming Hu, Junzhe Chen, Xiaochuan Li, Yufei Guo, Lijie Wen, Philip S Yu, and Zhijiang Guo. Do large language models know about facts?arXiv preprint arXiv:2310.05177,

work page arXiv 2023
[10]

Degpt: Optimizing decompiler output with llm

[Huet al., 2024a ] Peiwei Hu, Ruigang Liang, and Kai Chen. Degpt: Optimizing decompiler output with llm. InProceedings 2024 Network and Distributed Sys- tem Security Symposium (2024). https://api. semantic- scholar.org/CorpusID, volume 267622140,

work page 2024
[11]

Evaluating ro- bustness of generative search engine on adversarial factual questions.arXiv preprint arXiv:2403.12077,

[Huet al., 2024b ] Xuming Hu, Xiaochuan Li, Junzhe Chen, Yinghui Li, Yangning Li, Xiaoguang Li, Yasheng Wang, Qun Liu, Lijie Wen, Philip S Yu, et al. Evaluating ro- bustness of generative search engine on adversarial factual questions.arXiv preprint arXiv:2403.12077,

work page arXiv
[12]

Towards continual knowledge learning of language models.arXiv preprint arXiv:2110.03215,

[Janget al., 2021 ] Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stan- ley Jungkyu Choi, and Minjoon Seo. Towards continual knowledge learning of language models.arXiv preprint arXiv:2110.03215,

work page arXiv 2021
[13]

Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity.arXiv preprint arXiv:2403.14403,

[Jeonget al., 2024 ] Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong C Park. Adaptive- rag: Learning to adapt retrieval-augmented large lan- guage models through question complexity.arXiv preprint arXiv:2403.14403,

work page arXiv 2024
[14]

Mistral 7B

[Jianget al., 2023 ] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825,

work page internal anchor Pith review Pith/arXiv arXiv 2023
[15]

Tug-of-war between knowledge: Exploring and resolving knowledge conflicts in retrieval-augmented lan- guage models.arXiv preprint arXiv:2402.14409,

[Jinet al., 2024a ] Zhuoran Jin, Pengfei Cao, Yubo Chen, Kang Liu, Xiaojian Jiang, Jiexin Xu, Qiuxia Li, and Jun Zhao. Tug-of-war between knowledge: Exploring and resolving knowledge conflicts in retrieval-augmented lan- guage models.arXiv preprint arXiv:2402.14409,

work page arXiv
[16]

Cutting off the head ends the conflict: A mechanism for interpreting and mitigating knowledge conflicts in language models.arXiv preprint arXiv:2402.18154,

[Jinet al., 2024b ] Zhuoran Jin, Pengfei Cao, Hongbang Yuan, Yubo Chen, Jiexin Xu, Huaijun Li, Xiaojian Jiang, Kang Liu, and Jun Zhao. Cutting off the head ends the conflict: A mechanism for interpreting and mitigating knowledge conflicts in language models.arXiv preprint arXiv:2402.18154,

work page arXiv
[17]

Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Infor- mation Processing Systems, 33:9459–9474,

[Lewiset al., 2020 ] Patrick Lewis, Ethan Perez, Aleksan- dra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K ¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Infor- mation Processing Systems, 33:9459–9474,

work page 2020
[18]

Rouge: A package for automatic evaluation of summaries

[Lin, 2004] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81,

work page 2004
[19]

Adversarial search engine opti- mization for large language models.arXiv preprint arXiv:2406.18382,

[Nestaaset al., 2024 ] Fredrik Nestaas, Edoardo Debenedetti, and Florian Tram `er. Adversarial search engine opti- mization for large language models.arXiv preprint arXiv:2406.18382,

work page arXiv 2024
[20]

Bleu: a method for au- tomatic evaluation of machine translation

[Papineniet al., 2002 ] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for au- tomatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Compu- tational Linguistics, pages 311–318,

work page 2002
[21]

arXiv preprint arXiv:1909.01066 , year=

[Petroniet al., 2019 ] Fabio Petroni, Tim Rockt ¨aschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases?arXiv preprint arXiv:1909.01066,

work page arXiv 2019
[22]

Rag-fusion: a new take on retrieval-augmented generation.arXiv preprint arXiv:2402.03367,

[Rackauckas, 2024] Zackary Rackauckas. Rag-fusion: a new take on retrieval-augmented generation.arXiv preprint arXiv:2402.03367,

work page arXiv 2024
[23]

Language models are unsupervised multitask learners

[Radfordet al., 2019 ] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9,

work page 2019
[24]

Direct preference optimization: Your lan- guage model is secretly a reward model.Advances in Neu- ral Information Processing Systems, 36,

[Rafailovet al., 2024 ] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your lan- guage model is secretly a reward model.Advances in Neu- ral Information Processing Systems, 36,

work page 2024
[25]

How Much Knowledge Can You Pack Into the Parameters of a Language Model?

[Robertset al., 2020 ] Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model?arXiv preprint arXiv:2002.08910,

work page internal anchor Pith review arXiv 2020
[26]

The probabilistic relevance framework: Bm25 and beyond.Foundations and Trends® in Information Re- trieval, 3(4):333–389,

[Robertsonet al., 2009 ] Stephen Robertson, Hugo Zaragoza, et al. The probabilistic relevance framework: Bm25 and beyond.Foundations and Trends® in Information Re- trieval, 3(4):333–389,

work page 2009
[27]

Llama 2: Open Foundation and Fine-Tuned Chat Models

[Touvronet al., 2023 ] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

work page internal anchor Pith review Pith/arXiv arXiv 2023
[28]

Using LLMs to Automate Threat Intelligence Analysis Workflows in Security Operation Centers, July 2024

[Tsenget al., 2024 ] PeiYu Tseng, ZihDwo Yeh, Xushu Dai, and Peng Liu. Using llms to automate threat intelligence analysis workflows in security operation centers.arXiv preprint arXiv:2407.13093,

work page arXiv 2024
[29]

Resolving knowledge conflicts in large lan- guage models.arXiv preprint arXiv:2310.00935,

[Wanget al., 2023 ] Yike Wang, Shangbin Feng, Heng Wang, Weijia Shi, Vidhisha Balachandran, Tianxing He, and Yu- lia Tsvetkov. Resolving knowledge conflicts in large lan- guage models.arXiv preprint arXiv:2310.00935,

work page arXiv 2023
[30]

Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowl- edge conflicts.arXiv preprint arXiv:2305.13300,

[Xieet al., 2023 ] Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, and Yu Su. Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowl- edge conflicts.arXiv preprint arXiv:2305.13300,

work page arXiv 2023
[31]

Knowledge conflicts for llms: A survey.arXiv:2403.08319, 2024

[Xuet al., 2024 ] Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu. Knowledge conflicts for llms: A survey.arXiv preprint arXiv:2403.08319,

work page arXiv 2024
[32]

A hybrid rag sys- tem with comprehensive enhancement on complex reason- ing.arXiv preprint arXiv:2408.05141, 2024

[Yuanet al., 2024 ] Ye Yuan, Chengwu Liu, Jingyang Yuan, Gongbo Sun, Siqi Li, and Ming Zhang. A hybrid rag sys- tem with comprehensive enhancement on complex reason- ing.arXiv preprint arXiv:2408.05141, 2024

work page arXiv 2024