Pith. sign in

REVIEW 6 major objections 5 minor 27 references

LLM Embedding-based Attribution (LEA): Quantifying Source Contributions to Generative Model's Response for Vulnerability Analysis

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read LEA, computed from layer-0 embedding ranks, quantifies how much of an LLM response comes from retrieved context versus internal knowledge, and separates valid, generic, and absent retrieval with over 95% accuracy on larger models.

desk verdict A cheap, high-performing retrieval-quality detector; the attribution claim is a leap the paper never validates. read the letter →

arxiv 2506.12100 v2 pith:LZKB3U66 submitted 2025-06-12 cs.CR cs.AI

classification cs.CRcs.AI
keywords LLMattributionretrieval-augmentedgenerationvulnerabilityanalysislinearindependencehiddenstatesCVEexplainability
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

Large language models answering questions about software vulnerabilities can draw on information retrieved at query time, on knowledge stored in their weights, or on the wording of the question itself. The paper proposes LEA (LLM Embedding-based Attribution), a metric that assigns each generated token to one of those three sources by checking, at layer 0 of the transformer, whether the token's embedding adds a new linearly independent direction to the question's embedding matrix. Applied to 500 critical and high-severity CVEs spanning 2016–2025, LEA produces a distribution that clearly separates no-retrieval, generic-retrieval, and valid-retrieval scenarios, with test accuracy above 95% for the larger models evaluated. The paper also finds that when the retrieved text is factually incorrect, LEA's distribution looks almost identical to the valid-retrieval case, which it reads as evidence that the models trust whatever context they receive rather than recognizing the specific CVE. If this is right, security analysts get a cheap, layer-wise audit signal for deciding when a RAG-grounded answer is trustworthy and when human verification is needed.

What carries the argument

The load-bearing object is layer-0 linear dependence. For each token of the generated response, the paper appends the token's layer-0 hidden state (embedding plus positional encoding) to the matrix formed by the question-token embeddings and asks whether the numerical rank of the matrix increases, giving a binary signal per token. This is computed twice, for the sequence without retrieved context and for the sequence with retrieved context, and the two bits are aggregated over response tokens into the three-component LEA distribution $A_{\mathrm{fnd}}$, $A_{\mathrm{rag}}$, and $A_q$. The mechanism rests on the observation that transformer hidden states exhibit high redundancy, so only a subset of tokens adds genuinely new linear directions; layer 0 is chosen because later layers increasingly treat tokens as independent and lose the inter-token dependency signal.

What would settle it

Build a response whose tokens are known by construction to come from one source—some tokens copied verbatim from the retrieved paragraph, some copied from the question, some produced without retrieval—and compute LEA's per-token $\delta$ values; if copied-retrieval tokens are not flagged as $A_{\mathrm{rag}}$, or question-copied tokens are, the signal is not measuring source. A second decisive check is to vary the rank threshold or numerical tolerance and show that the 95% separation between valid and generic retrieval survives the variation.

Watch

Extended reading notes

Core claim

The central claim is that the relative influence of the input query, the retrieved context, and the model's internal parameters on a generated response can be read off from the linear dependence structure of layer-0 token embeddings. For each generated token $i$, the paper defines two binary indicators $\delta_i^{M}$ and $\delta_i^{M'}$, equal to 1 exactly when appending that token's embedding to the question-token matrix increases the numerical rank of the matrix, computed with and without the retrieved context in the sequence. The three-component LEA distribution $(A_{\mathrm{fnd}}; A_{\mathrm{rag}}; A_q)$ aggregates those indicators over the response: the fraction of tokens that stay linearly independent, the fraction that become dependent only after the retrieved context is added, and the fraction that remain dependent. The paper's empirical claim is that a well-grounded response shows a balanced distribution with substantial $A_{\mathrm{rag}}$, a generic or retrieval-free response is dominated by $A_{\mathrm{fnd}}$, and an incorrectly retrieved response mirrors the valid case, exposing the model's inability to distinguish wrong context from right context.

Load-bearing premise

The whole method stands on the assumption that whether a token's embedding adds a new independent direction to the question's layer-0 matrix is a faithful indicator of where the token's meaning came from; the paper never validates this against token-level ground truth and never specifies the numerical tolerance used to decide whether the rank actually increases.

Editorial extensions

If this is right

  • A learned threshold on $A_{\mathrm{rag}}$ can serve as a triage rule: responses that fall below it are flagged as generic or ungrounded and can be sent for human review before patch priorities are set.
  • Valid retrieval becomes the expected, verifiable baseline: when the NVD description is supplied, LEA shows a balanced dependence on retrieved context and internal knowledge, giving analysts a reference distribution for a trustworthy answer.
  • Because the incorrect-retrieval distribution closely matches the valid-retrieval distribution, retrieval quality must be verified before the LLM call, not inferred from the answer's fluency.
  • The metric separates valid from generic or absent retrieval more sharply on larger models, suggesting that the attribution signal is clearer when the underlying model has more capacity.

Reading between the lines

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

  • Editorial inference: LEA's layer-0 rank test should transfer to non-cybersecurity RAG tasks such as open-domain QA, where template-based prompts can produce generated tokens whose true source is known; a positive result would turn LEA into a general RAG-fidelity audit tool.
  • Editorial inference: The near-identical distributions for valid and incorrect retrieval imply that LEA detects reliance, not correctness; catching wrong-but-confidently-used context needs an additional consistency check between the retrieved text and the generation.
  • Editorial inference: The paper reports accuracy at one threshold per model chosen from the ROC curve; reporting calibration curves and per-year breakdowns, especially for post-cutoff 2025 CVEs, would show how stable the 95% figure is in operational settings.
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

6 major / 5 minor

Summary. The paper proposes LLM Embedding-based Attribution (LEA), a metric intended to quantify how much of an LLM's generated response is attributable to the input query, retrieved context (RAG), or the model's internal knowledge. LEA is computed by measuring, at layer 0, whether appending each generated token's vector to a question matrix increases the numerical rank of that matrix, and then aggregating the resulting binary transition patterns into three distribution components. The authors evaluate LEA on 500 CVEs from 2016–2025 across three RAG scenarios (valid, generic, incorrect) and three LLMs, reporting distributional differences and classification accuracies above 95% for distinguishing valid retrieval from generic or no retrieval. They also report that incorrect retrieval produces LEA distributions similar to valid retrieval, which they present as a caution about blind reliance on RAG. The paper includes a dataset and code release.

Significance. If the proposed metric were well-defined and validated, LEA could be a practically useful tool for auditing RAG-based vulnerability analysis pipelines, and the curated dataset of 500 CVEs with three retrieval scenarios is a valuable resource for the security community. The empirical observation that LEA captures clear distributional differences between informative and non-informative retrieval is a credible and reproducible starting point. However, the central claim of the paper—that LEA quantifies the relative causal contribution of internal knowledge versus retrieved content—is not supported by the evidence presented. The aggregate distributional separation does not establish token-level attribution, and the metric itself is not specified precisely enough to be reproduced as written. The paper's main value at present is as an empirical demonstration that certain rank-based statistics differ across retrieval quality, not as a validated attribution method.

major comments (6)
  1. [Eq. (10)–(11), Sec. 4] The definition of the LEA components is not reproducible as written. Equation (10) defines A(M,M′) as a sum over tokens of (1/h) δ^M_i ∧ δ^M′_i, but h is never defined, and Equation (11) then sets LEA = (A(1,1); A(1,0); A(0,0)), which is not a valid substitution because A takes matrices M,M′, not booleans. In addition, δ^M_i is defined as whether appending a token vector to the “question matrix Q” increases rank, but Q is never defined, nor is the vector representation to be appended. The paper must specify Q, the vector, h, and the exact rank test before the metric can be evaluated.
  2. [Table 2, Secs. 3–4] The rank computation is under-specified: Table 2 reports “rank xθy 48%” and similar values, but matrix rank is an integer, so the normalization is unclear. With floating-point embeddings, the exact rank of a token matrix is generically full, so “increases rank” is meaningful only relative to a numerical tolerance (e.g., an SVD singular-value cutoff). No tolerance is given, which makes the central δ_i values undefined.
  3. [Sec. 4, “should not occur”] The paper defines three LEA components (Afnd, Arag, Aq) and states that the (0,1) transition—dependent without RAG, independent with RAG—“should not occur as it indicates inconsistency.” If such transitions occur, the three components do not sum to 1, and the reported distributions are not a proper partition of tokens. The paper does not report the frequency of (0,1) transitions or explain how they are handled; this needs to be resolved for the metric to be meaningful.
  4. [Sec. 6.2, Table 4] The post-hoc filtering of tokens (stop-word removal plus discarding tokens with ∆p ≤ 0) is not justified beyond an appeal to semantic value. Since the threshold filter uses final-layer probabilities that are correlated with the very signal LEA measures, it may systematically alter the LEA distribution (e.g., by removing tokens that are invariant to RAG, inflating Arag). The paper should report unfiltered results or justify the filter with a sensitivity analysis.
  5. [Secs. 3, 6.2–6.3] The load-bearing premise that a token's rank-increase status at layer 0 tracks the token's causal source is never validated. The experimental evidence shows only that aggregate Arag distributions differ across valid, generic, and non-retrieval conditions, and that Arag drops when the response mismatches the retrieval. These are response-level associations, not token-level attributions; the cited LoRA/pruning results concern low-rank weight updates and redundant directions, not the causal origin of generated tokens. The authors should validate δ_i against token-level ground truth, for example via intervention experiments that remove or corrupt the retrieved context and check which tokens change.
  6. [Sec. 6.4, Figs. 8–9] The paper demonstrates that LEA cannot distinguish valid from incorrect retrieval, since the distributions closely overlap. Because incorrect retrieval is the most dangerous failure in the motivating vulnerability-analysis scenario, the abstract's claim that LEA “offers security analysts with a metric to audit RAG-enhanced workflows” is broader than what the metric delivers. The claims should be re-scoped to detecting non-informative versus informative retrieval, or the method should be extended to flag incorrect retrieval.
minor comments (5)
  1. [Eq. (8)] The row space of X ∈ F^{m×n} should be a subspace of F^n, not F^m as written.
  2. [Sec. 5.2] The citation list appears to have a mismatch: Mistral-Small-24B is cited as [18], but [12] is DeepSeek-R1-Distill-Llama-8B; check the references.
  3. [Sec. 6.5] The sentence “This distinction is 100% for the CVEs in 2025” is hard to reconcile with the 92–97% test accuracies in Table 5; clarify whether this refers to separation of distributions rather than classification accuracy.
  4. [Fig. 9] The captions and text could state more plainly that the overlapping distributions are the key finding, since the figure currently reads as a failed comparison.
  5. [Eq. (10)] If the factor 1/h is a typo or stands for a normalization constant, it should be defined; as written it is unexplained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LEA is a proposed definitional metric, and its empirical evaluation is grounded in externally labeled retrieval scenarios with a held-out test split.

full rationale

The paper defines LEA as a metric rather than deriving a prediction from fitted inputs: Arag is defined in Eq. 11 as the fraction of response tokens that are rank-independent without retrieval and rank-dependent with retrieval. The central claim that this fraction quantifies retrieved-content contribution is an interpretive assumption, not a derived result, and no equation in the paper reduces a predicted quantity to a fitted parameter. The empirical evaluation is externally grounded: the 'valid/generic/incorrect' labels are assigned by dataset construction (NVD description, RedHat generic text, suffix-matched wrong CVE), and Table 5 reports held-out test accuracy from an 80/20 split with a threshold chosen from the ROC curve. There are no self-citations to the authors' prior work, and the cited LoRA/pruning results are external, so no load-bearing self-citation chain exists. The lack of validation that a layer-0 rank increase tracks token-level causal source, and the undefined quantities Q, h, and the numerical rank tolerance, are correctness and reproducibility concerns, not circularity under the specified patterns. Therefore no significant circularity is found.

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

The metric rests on three unproven premises: that linear dependence of embeddings indicates source influence, that layer-0 rank changes isolate RAG effects, and that the (0,1) transition can be ignored. The classification threshold and the effective rank tolerance are fitted or unspecified parameters. No new physical entities are introduced.

free parameters (3)
  • Arag classification threshold = 12.5% (Gemma-3-27B), 13.3% (Mistral-Small-24B), 12.3% (LLaMA-3.2-3B)
    Chosen as the point on the ROC curve closest to the perfect classifier, using the 80/20 train/test split (Table 5).
  • effective rank tolerance = not specified
    The paper reports layer-0 ranks around 48% in Table 2, which is inconsistent with standard numerical rank for distinct token embeddings, implying an unspecified effective-rank or tolerance criterion is used.
  • probability-delta filter cutoff = 0
    Tokens with delta-p = p(x theta y) - p(x y) <= 0 are filtered out before computing LEA distributions; this cutoff is fixed but arbitrary and affects the reported numbers.
assumptions (3)
  • domain assumption Linear dependence of token embeddings is a valid proxy for the source of token generation.
    Section 3 motivates this using LoRA and pruning literature, but no ground-truth validation is provided.
  • domain assumption Adding retrieved context changes the span of the question matrix in a way that isolates RAG influence.
    This is the basis for defining Arag = A(1,0) in Eq. 11.
  • ad hoc to paper The (0,1) transition (dependent without RAG, independent with RAG) should not occur and is ignored.
    Stated in Section 4 after Eq. 11; no proof is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM Embedding-based Attribution (LEA): Quantifying Source Contributions to Generative Model's Response for Vulnerability Analysis." pith.science (2026). https://pith.science/paper/LZKB3U66

@misc{pith2026250612100,
  author       = {Pith},
  title        = {Pith review of: LLM Embedding-based Attribution (LEA): Quantifying Source Contributions to Generative Model's Response for Vulnerability Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LZKB3U66}},
  note         = {Machine review of arXiv:2506.12100}
}
read the original abstract

Large Language Models (LLMs) are increasingly used for cybersecurity threat analysis, but their deployment in security-sensitive environments raises trust and safety concerns. With over 21,000 vulnerabilities disclosed in 2025, manual analysis is infeasible, making scalable and verifiable AI support critical. When querying LLMs, dealing with emerging vulnerabilities is challenging as they have a training cut-off date. While Retrieval-Augmented Generation (RAG) can inject up-to-date context to alleviate the cut-off date limitation, it remains unclear how much LLMs rely on retrieved evidence versus the model's internal knowledge, and whether the retrieved information is meaningful or even correct. This uncertainty could mislead security analysts, mis-prioritize patches, and increase security risks. Therefore, this work proposes LLM Embedding-based Attribution (LEA) to analyze the generated responses for vulnerability exploitation analysis. More specifically, LEA quantifies the relative contribution of internal knowledge vs. retrieved content in the generated responses. We evaluate LEA on 500 critical vulnerabilities disclosed between 2016 and 2025, across three RAG settings -- valid, generic, and incorrect -- using three state-of-the-art LLMs. Our results demonstrate LEA's ability to detect clear distinctions between non-retrieval, generic-retrieval, and valid-retrieval scenarios with over 95% accuracy on larger models. Finally, we demonstrate the limitations posed by incorrect retrieval of vulnerability information and raise a cautionary note to the cybersecurity community regarding the blind reliance on LLMs and RAG for vulnerability analysis. LEA offers security analysts with a metric to audit RAG-enhanced workflows, improving the transparent and trustworthy deployment of AI in cybersecurity threat analysis.

Figures

Figures reproduced from arXiv: 2506.12100 by the authors.

Figure 1
Figure 1. Probability shifts in the RAG-generated response of a CVE before and after including the retrieved content. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. LEA’s distribution for a vulnerability. visibility into how token interactions change with depth, and how the presence of RAG-derived context modulates these interactions. This leads us to consider how the concept can be applied in the analysis of vulnerabilities to detect which token generations in the responses were from the retrieved content. In this setting, rank-based analysis can help deter￾mine whether the mo… view at source ↗
Figure 3
Figure 3. The step-by-step of getting the hidden state progression and probability differences. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The process to compute LEA with the dependent vectors with (top) and without (bottom) retrieved context. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of LEA distribution of non-retrieval vs. generic retrieval vs. valid retrieval. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of filtered LEA distribution of the valid response with different RAG context. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Comparison of filtered LEA distribution of the valid ground-truth RAG with different response generations. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: LEA distribution for the incorrect retrieval of a CVE in the query [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Comparison of LEA’s A rag distribution of valid-retrieval-response vs. incorrect-retrieval-response over the years. reliance, which allows for informed decision-making when interpreting responses about unseen or complex vulnerabili￾ties. This analysis provides a refere…
Figure 10
Figure 10. Figure 10: Comparison of LEA’s A rag distribution of non-retrieval vs. generic retrieval vs. valid retrieval of all the curated CVEs [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 18 canonical work pages

  1. [1]

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. The Twelfth International Conference on Learning Represen- tations, 2023

  2. [2]

    Improving Language Models by Retrieving from Trillions of Tokens

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving Language Models by Retrieving from Trillions of Tokens. International Conference on Machine Learning , pages 2206–2240, 2022

  3. [3]

    Evaluation of ChatGPT Model for Vulnerabil- ity Detection

    Anton Cheshkov, Pavel Zadorozhny, and Rodion Levichev. Evaluation of ChatGPT Model for Vulnerabil- ity Detection. arXiv preprint arXiv:2304.07232, 2023

  4. [4]

    ChatNVD: Advancing Cybersecurity Vulnerability Assessment with Large Language Models

    Shivansh Chopra, Hussain Ahmad, Diksha Goel, and Claudia Szabo. ChatNVD: Advancing Cybersecurity Vulnerability Assessment With Large Language Models. arXiv preprint arXiv:2412.04756, 2024

  5. [5]

    PentestGPT: Evaluating and Harnessing Large Language Models for Automated Pen- etration Testing

    Gelei Deng, Yi Liu, Víctor Mayoral-Vilches, Peng Liu, Yuekang Li, Yuan Xu, Tianwei Zhang, Yang Liu, Martin Pinzger, and Stefan Rass. PentestGPT: Evaluating and Harnessing Large Language Models for Automated Pen- etration Testing. In 33rd USENIX Security Symposium (USENIX Security 24). USENIX Association, 2024

  6. [6]

    A Sliding Layer Merging Method for Effi- cient Depth-Wise Pruning in LLMs

    Xuan Ding, Yao Zhu, Yunjian Zhang, and Chuanlong Xie. A Sliding Layer Merging Method for Effi- cient Depth-Wise Pruning in LLMs. arXiv preprint arXiv:2502.19159, 2025

  7. [7]

    Vul-RAG: Enhancing LLM- based Vulnerability Detection via Knowledge-level RAG

    Xueying Du, Geng Zheng, Kaixin Wang, Jiayi Feng, Wentai Deng, Mingwei Liu, Bihuan Chen, Xin Peng, Tao Ma, and Yiling Lou. Vul-RAG: Enhancing LLM- based Vulnerability Detection via Knowledge-level RAG. arXiv preprint arXiv:2406.11147, 2024

  8. [8]

    Introducing Gemma 3: The Most Ca- pable Model You Can Run on a Single GPU or TPU

    Google. Introducing Gemma 3: The Most Ca- pable Model You Can Run on a Single GPU or TPU. https://blog.google/technology/ developers/gemma-3/, 2025. Accessed: 2025- 04

Show all 27 references
  1. [9]

    From ChatGPT to ThreatGPT: Impact of Generative AI in Cybersecurity and Privacy

    Maanak Gupta, CharanKumar Akiri, Kshitiz Aryal, Eli Parker, and Lopamudra Praharaj. From ChatGPT to ThreatGPT: Impact of Generative AI in Cybersecurity and Privacy. IEEE Access, 2023

  2. [10]

    LoRA: Low-Rank Adaptation of Large Language Models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-Rank Adaptation of Large Language Models. 2021

  3. [11]

    A Survey on Hallucination in Large Language Models: Princi- ples, Taxonomy, Challenges, and Open Questions.arXiv preprint arXiv:2311.05232, 2023

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Wei- hua Peng, Xiaocheng Feng, Bing Qin, et al. A Survey on Hallucination in Large Language Models: Princi- ples, Taxonomy, Challenges, and Open Questions.arXiv preprint arXiv:2311.0...

  4. [12]

    DeepSeek-R1-Distill-Llama-8B

    HuggingFace. DeepSeek-R1-Distill-Llama-8B. https://huggingface.co/deepseek-ai/ DeepSeek-R1-Distill-Llama-8B , 2025. Accessed: 2025-03

  5. [13]

    Adaptive-RAG: Learn- ing to Adapt Retrieval-Augmented Large Language Models through Question Complexity

    Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong C Park. Adaptive-RAG: Learn- ing to Adapt Retrieval-Augmented Large Language Models through Question Complexity. arXiv preprint arXiv:2403.14403, 2024

  6. [14]

    Understand- ing the Effectiveness of Large Language Models in Detecting Security Vulnerabilities

    Avishree Khare, Saikat Dutta, Ziyang Li, Alaia Solko- Breslin, Rajeev Alur, and Mayur Naik. Understand- ing the Effectiveness of Large Language Models in Detecting Security Vulnerabilities. arXiv preprint arXiv:2311.16169, 2023

  7. [15]

    SuRe: Summarizing Retrievals using An- swer Candidates for Open-domain QA of LLMs

    Jaehyung Kim, Jaehyun Nam, Sangwoo Mo, Jongjin Park, Sang-Woo Lee, Minjoon Seo, Jung-Woo Ha, and Jinwoo Shin. SuRe: Summarizing Retrievals using An- swer Candidates for Open-domain QA of LLMs. arXiv preprint arXiv:2404.13081, 2024

  8. [16]

    LLM- Pruner: On the Structural Pruning of Large Language Models

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. LLM- Pruner: On the Structural Pruning of Large Language Models. Advances in neural information processing systems, 36:21702–21720, 2023

  9. [17]

    Llama 3.2: Revolutionizing edge ai and vi- sion with open, customizable models

    AI Meta. Llama 3.2: Revolutionizing edge ai and vi- sion with open, customizable models. Meta AI Blog. Retrieved December, 20, 2024. Accessed: 2025-01

  10. [18]

    Mistral Small 3

    Mistral AI. Mistral Small 3. https://mistral. ai/news/mistral-small-3, 2025. Accessed: 2025-03. 14

  11. [19]

    LOCALINTEL: Generating Organizational Threat Intelligence from Global and Local Cyber Knowl- edge

    Shaswata Mitra, Subash Neupane, Trisha Chakraborty, Sudip Mittal, Aritran Piplai, Manas Gaur, and Shahram Rahimi. LOCALINTEL: Generating Organizational Threat Intelligence from Global and Local Cyber Knowl- edge. arXiv preprint arXiv:2401.10036, 2024

  12. [20]

    CVE - Common Vulnerabilities and Expo- sures

    MITRE. CVE - Common Vulnerabilities and Expo- sures. https://cve.mitre.org/, 2024. Ac- cessed: 2024-11

  13. [21]

    National Vulnerability Database (NVD)

    NIST. National Vulnerability Database (NVD). https://nvd.nist.gov/, 2024. Accessed: 2024- 11

  14. [22]

    Empirical Validation of Automated Vulnerability Curation and Characterization

    Ahmet Okutan, Peter Mell, Mehdi Mirakhorli, Igor Khokhlov, Joanna CS Santos, Danielle Gonzalez, and Steven Simmons. Empirical Validation of Automated Vulnerability Curation and Characterization. IEEE Transactions on Software Engineering , 49(5):3241– 3260, 2023

  15. [23]

    AGIR: Automating Cyber Threat Intelligence Reporting with Natural Language Gener- ation

    Filippo Perrina, Francesco Marchiori, Mauro Conti, and Nino Vincenzo Verde. AGIR: Automating Cyber Threat Intelligence Reporting with Natural Language Gener- ation. In 2023 IEEE International Conference on Big Data (BigData), pages 3053–3062. IEEE, 2023

  16. [24]

    The Troubling Emergence of Hallucination in Large Language Models–An Extensive Definition, Quantification, and Prescriptive Remedia- tions

    Vipula Rawte, Swagata Chakraborty, Agnibh Pathak, Anubhav Sarkar, SM Tonmoy, Aman Chadha, Amit P Sheth, and Amitava Das. The Troubling Emergence of Hallucination in Large Language Models–An Extensive Definition, Quantification, and Prescriptive Remedia- tions. arXiv preprint a...

  17. [25]

    What is a CVE? https://www.redhat

    RedHat. What is a CVE? https://www.redhat. com/en/topics/security/what-is-cve,

  18. [26]

    Attention Is All You Need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. Ad- vances in Neural Information Processing Systems, 30, 2017

  19. [27]

    Hallucination Mitigation for Retrieval-Augmented Large Language Models: A Review

    Wan Zhang and Jing Zhang. Hallucination Mitigation for Retrieval-Augmented Large Language Models: A Review. Mathematics, 13(5):856, 2025. 15

Pith tools

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