Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Enhancing Android Malware Detection with Retrieval-Augmented Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Grounded LLM descriptions of Android app features reach 96.69% recall in malware classification, beating a fused-LLM baseline on accuracy and F1.

desk verdict Useful internal comparison of RAG-based vs fusion-based app descriptions, but the advertised accuracy gain over conventional feature-based detection is never tested. read the letter →

arxiv 2506.22750 v1 pith:OQBO64OK submitted 2025-06-28 cs.CR

Saraga S. , Anagha M. S. , Dincy R. Arikkat , Rafidha Rehiman K. A. , Serena Nicolazzo , Antonino Nocera , Vinod P This is my paper · ORCID
classification cs.CR
keywords Androidmalwaredetectionretrieval-augmentedgenerationstaticanalysisLLM-generateddescriptionsBERTclassificationcybersecurityNLPagenticRAGfunctionaldescription
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

This paper tries to establish that Android malware detection improves when raw static features are first converted into functional descriptions by a retrieval-augmented LLM pipeline, rather than classified directly. The pipeline extracts permissions, services, receivers, and intent actions from APKs without running them, grounds an LLM's writing in curated Android documentation through AgenticRAG, and feeds the resulting descriptions to cybersecurity-tuned BERT classifiers. On a dataset of 18,000 apps labeled by a multi-engine scanning API, the best combination reaches 92.89% accuracy, 96.69% recall, and an F1-score of 92.86% for AgenticRAG descriptions, beating an LLM-fusion description baseline. The paper asserts this improves over conventional feature-based methods; its reported experiments compare description-generation strategies and classifier choices.

What carries the argument

The load-bearing mechanism is AgenticRAG, named in the paper as an agentic retrieval-augmented generation pipeline. It combines a curated corpus of Android API documentation covering permissions, services, receivers, and intent actions; an ensemble retriever pairing dense vector search with sparse keyword search fused by reciprocal rank fusion; exact-then-fuzzy matching with a similarity threshold for unseen or misspelled feature names; a cache memory for reused lookups; and a fallback LLM query for unmatched features. The retrieved context is inserted into a structured prompt that forbids speculation and demands one cohesive paragraph describing what the app can do, and that paragraph is tokenized and classified by domain-adapted BERT models. The machinery turns sparse manifest features into dense semantic text that a transformer classifier can separate.

What would settle it

Train the same cybersecurity-tuned BERT classifier on raw static feature strings (permissions, services, receivers, and intents serialized as text) and on the AgenticRAG descriptions using identical train and test splits. If the raw-feature model matches or exceeds 92.89% accuracy or 96.69% recall, the central claim that grounded descriptions improve over conventional feature-based representations is refuted. A separate audit of the labels would also settle it: if a random sample of one-engine-flagged apps is mostly adware or greyware, the reported recall is measured against a noisy target.

Watch

Extended reading notes

Core claim

The paper's central claim is that retrieval-grounded LLM descriptions are a better text representation for Android malware classification than descriptions produced by fusing two independent LLMs. In the AgenticRAG pipeline, each APK's static features are normalized, matched against a documentation corpus by dense and sparse retrieval, and synthesized by an LLM under a strict prompt into a single factual paragraph. When these paragraphs are classified by CySecBERT, a BERT model adapted to cybersecurity text, the paper reports 92.89% accuracy, 96.69% recall, and a 92.86% F1-score, while the Gemini Fusion descriptions reach 91.36% accuracy, 90.50% recall, and a 91.25% F1-score. CySecBERT also beats SecBERT, another security-oriented BERT variant, on recall and F1 for both description types, although SecBERT's plain accuracy on AgenticRAG descriptions is slightly higher at 93.31%. The paper attributes the gain to contextual specificity, better handling of unusual permission combinations, semantic coherence, and malware-specific terminology in grounded descriptions.

Load-bearing premise

The whole experiment relies on labels from a multi-engine scanning API: an app flagged by at least one engine counts as malicious and an app flagged by none counts as benign, with no independent validation of those labels.

Editorial extensions

If this is right

  • Replacing raw static features with grounded functional descriptions could let small BERT-size models serve as fast, deployable malware classifiers that never execute an app.
  • If the reported recall of 96.69% holds, the approach would miss only about 3 in 100 malicious apps, which matters more in security settings than small accuracy differences.
  • Improving the retrieval corpus, fuzzy matching, or fallback generation should directly shift the classification ceiling, because description quality is the identified bottleneck.
  • The same grounded-description pipeline should transfer to other platforms or feature sets such as API calls and network permissions, which the paper lists as future work.

Reading between the lines

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

  • The abstract claims improvement over conventional feature-based methods, but the experiments compare only description-generation strategies; a reader should treat that broader claim as untested until a raw-static-feature baseline is added.
  • The one-engine-flag labeling rule likely classes some adware or greyware as malware, so re-labeling with a stricter multi-engine agreement threshold would test whether the ranking of description methods is stable under cleaner labels.
  • An ablation with plain LLM descriptions and no retrieval would isolate how much of the reported gain comes from RAG grounding versus prompt design or the LLM itself.
  • Because the descriptions are generated without executing apps, the pipeline could be chained with dynamic analysis to catch dormant or evasive malware, an extension the paper names as future work.
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 an Android malware detection pipeline that converts static APK features (permissions, services, receivers, intent actions) into natural-language functional descriptions using a custom AgenticRAG system built on Gemini 2.0 Flash Lite, and then classifies these descriptions with domain-adapted BERT models (CySecBERT and SecBERT). For comparison, the authors generate a second set of descriptions by fusing LLaMA2 and Mistral outputs with Gemini, and they report that AgenticRAG descriptions plus CySecBERT achieve 92.89% accuracy, 96.69% recall, and 92.86% F1-score, outperforming the Gemini fusion pipeline. The abstract and contributions claim this approach improves detection accuracy over conventional feature-based methods.

Significance. If properly supported, the idea of using retrieval-augmented, LLM-generated functional descriptions as an intermediate representation for transformer-based malware classification would be a useful contribution. The paper has concrete strengths: the description-generation architecture is described in detail, the prompts are provided, two description-generation strategies are compared under identical preprocessing and classification conditions, and two domain-specific BERT models are evaluated with confusion matrices. However, the advertised central claim—improvement over conventional feature-based malware detection—is not tested by any experiment in the manuscript, and the ground-truth labels rest on a single unvalidated VirusTotal threshold. As reported, the results only show that one LLM-generated text representation beats another LLM-generated text representation with the same BERT classifier family.

major comments (4)
  1. [§4.3, Tables 5–8] The central claim in the abstract and Section 1—that the method improves detection accuracy over conventional feature-based methods—is unsupported because no conventional feature-based baseline is evaluated. Tables 5–8 compare only two LLM description-generation pipelines and two BERT classifiers; there is no experiment training, for example, a Drebin-style classifier on permissions/intents/API calls with SVM or Random Forest, or any other standard static-feature baseline on the same corpus. Without such a baseline, the reported 92.89% accuracy and 96.69% recall do not demonstrate any gain over conventional features, and the paper's headline contribution remains untested.
  2. [§3.1] The VirusTotal labeling protocol uses a single threshold—one or more detecting engines means malicious, zero engines means benign—with no validation against a trusted malware or benign set. This is known to be noisy: many benign apps are flagged by one engine and many malware samples evade all engines. Because every reported metric depends on these labels, the accuracy, recall, and F1 values in Tables 5–8 are only as meaningful as this unvalidated ground truth. The authors should provide distribution of detection counts, or cross-validate with a curated dataset, or otherwise justify the threshold.
  3. [§4.2, §5] The paper reports single-run results without error bars, confidence intervals, or significance tests, yet the conclusion states that the method yields "statistically significant improvements in all the considered evaluation metrics." No statistical claim of this kind can be made from the presented data. The authors should run multiple trials with different seeds, report standard deviations, and apply appropriate significance tests, or soften the claim accordingly.
  4. [§4.3.3] The post hoc explanation attributing AgenticRAG's superiority to "enhanced contextual specificity," "improved edge case handling," "semantic coherence," and "malware-specific terminology" is not supported by controlled experiments. There is no ablation that removes the retrieval component while keeping the same Gemini generator, no variation of the prompt without retrieval, and no analysis isolating the agentic planning behavior. Without such ablations, the causal claim that retrieval-augmented generation (rather than prompt differences or model stochasticity) drives the observed improvement is unsupported.
minor comments (5)
  1. [§4.3.1] The text states "CySecBERT with a recall rate of 96.89%" but Table 5 and Table 7 report 96.69%; these numbers should be reconciled.
  2. [Figures 8 and 9] The caption for Figure 8 labels subfigure (b) as "Gemini Fusion Descriptions Classified using SecBert," which appears to be a copy-paste error; it should likely refer to Mistral fusion descriptions. The captions for Figure 8 and Figure 9 should be checked against their content.
  3. [References [20] and [21]] Reference [21] for SecBERT describes a privacy-preserving neural network inference system, which does not match the HuggingFace SecBERT model used in the paper; the citation should be corrected or clarified.
  4. [§3.3.1, Table 8] The paper uses inconsistent naming for the open-source models: "LLaMA2" in the text and "Llama" or "Llama2" in tables and captions. Please standardize model names throughout.
  5. [General] The manuscript does not mention any plan to release code, prompts, or the generated description dataset. Providing an artifact would substantially improve reproducibility, especially given the complexity of the AgenticRAG pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the pipeline is supervised on external VirusTotal labels and the compared descriptions are independent transforms of static features; the missing conventional baseline is an evidential gap, not circularity.

full rationale

Walking the paper's derivation chain, each stage receives its input from an independent source. Labels are assigned via VirusTotal multi-engine detections (Section 3.1), static features are extracted from AndroZoo APKs with Androguard (Section 3.2), descriptions are generated from those static features by AgenticRAG or Gemini fusion (Section 3.3), and CySecBERT/SecBERT are fine-tuned on the descriptions to predict the external labels (Section 3.5). No parameter is fitted to a subset of the target metric and then reported as a prediction: the 70:10:20 train/validation/test split with class weights is standard supervised learning on an independent ground-truth source. No load-bearing step invokes a self-citation; references [20] and [21] are external model artifacts, and no uniqueness theorem or ansatz is imported from the authors' prior work. The abstract's claim of "improving detection accuracy over conventional feature-based methods" and the contribution claiming the system "outperforms state-of-the-art baselines" are not supported by the experiments in Tables 5-8, which compare only LLM-generated description pipelines; however, an absent conventional-feature baseline is a missing-experiment/correctness gap, not a circular reduction of the result to its inputs. The post hoc attributions in Section 4.3.3 are interpretive hypotheses, not derivations from the data. No quoted equation, fitted value, or definition in the paper is equivalent by construction to the claimed outcome, so the circularity score is 0.

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

The central pipeline introduces no new physical or formal entities. Its load-bearing assumptions are empirical: VirusTotal one-engine labels are correct, manifest-level static features capture malicious behavior, LLM summaries preserve features without hallucination, and the Android-documentation retrieval corpus is accurate. Vocabulary choices such as the fuzzy-match threshold, RRF weights, and class weights are unspecified or configurable and could affect the result.

free parameters (5)
  • VirusTotal malicious threshold = >=1 engine flags malicious
    Section 3.1: threshold heuristic defines the ground truth; no validation against a trusted malware label set.
  • Fuzzy matching similarity threshold = 65% (example)
    Section 3.3: configurable threshold for accepting fuzzy matches in retrieval; choice may affect recall and description content.
  • Train/validation/test split ratio = 70:10:20
    Section 4.1: single split, no cross-validation; results may depend on the particular split.
  • Class weights and weighted sampling parameters = not specified
    Section 4.1: class imbalance mitigation is mentioned, but the actual weights and sampling parameters are not disclosed.
  • RRF ensemble weights for FAISS and BM25 = not specified
    Section 3.3: configurable weights for ensemble retrieval are not reported.
assumptions (5)
  • domain assumption VirusTotal detection scores are a valid ground truth for Android malware.
    Section 3.1 uses a one-or-more-engine flag as malicious; no independent label verification is provided.
  • domain assumption Manifest-level static features (permissions, services, receivers, intents) are sufficient to characterize malicious behavior.
    Section 3.2 extracts only manifest-level features; no dynamic analysis or deeper code-level features are used.
  • domain assumption LLM-generated functional descriptions faithfully preserve the static features without hallucination.
    Section 3.3 uses RAG to ground output, but the generated text is not validated against the input features.
  • domain assumption The Android developer documentation corpus is accurate and complete for describing Android features.
    Section 3.3 builds a retrieval corpus from Android developer documentation; no audit of completeness or correctness is reported.
  • domain assumption Differences in classification performance are caused by the description generation strategy, not by prompt or pipeline artifacts.
    AgenticRAG and Gemini Fusion differ in prompts, models, and pipeline steps; these factors are not controlled in an ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Android Malware Detection with Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/OQBO64OK

@misc{pith2026250622750,
  author       = {Pith},
  title        = {Pith review of: Enhancing Android Malware Detection with Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQBO64OK}},
  note         = {Machine review of arXiv:2506.22750}
}
read the original abstract

The widespread use of Android applications has made them a prime target for cyberattacks, significantly increasing the risk of malware that threatens user privacy, security, and device functionality. Effective malware detection is thus critical, with static analysis, dynamic analysis, and Machine Learning being widely used approaches. In this work, we focus on a Machine Learning-based method utilizing static features. We first compiled a dataset of benign and malicious APKs and performed static analysis to extract features such as code structure, permissions, and manifest file content, without executing the apps. Instead of relying solely on raw static features, our system uses an LLM to generate high-level functional descriptions of APKs. To mitigate hallucinations, which are a known vulnerability of LLM, we integrated Retrieval-Augmented Generation (RAG), enabling the LLM to ground its output in relevant context. Using carefully designed prompts, we guide the LLM to produce coherent function summaries, which are then analyzed using a transformer-based model, improving detection accuracy over conventional feature-based methods for malware detection.

Figures

Figures reproduced from arXiv: 2506.22750 by the authors.

Figure 1
Figure 1. Architecture of the proposed Android Malware Classification System incorporating APK based Feature [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the AgenticRAG System descriptions, this approach facilitates more effective utilization of transformer-based models for security classification tasks. The foundation of the AgenticRAG system rests upon a comprehensive Data Corpus that was meticulously constructed through systematic extraction using custom web crawlers and consolidation of authoritative information from the official Android Developer… view at source ↗
Figure 3
Figure 3. Architecture of Gemini Fusion 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Architecture of Llama Fusion 4.3.3 Fusion Models Performance Comparison As anticipated above, we conducted a comprehensive comparative analysis to identify the most effective large language model (LLM) for fusion tasks. Our evaluation systematically compared three dist…
Figure 5
Figure 5. Figure 5: Architecture of Mistral Fusion 13 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Confusion Matrix for (a) AgenticRAG Descriptions Classified using CySecBert, (b)AgenticRAG Descriptions [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Confusion Matrix for (a) Gemini Fusion Descriptions Classified using CySecBert, (b) Gemini Fusion [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Confusion Matrix for (a) Mistral Fusion Descriptions Classified using CySecBert, (b) Gemini Fusion [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Confusion Matrix for (a)Llama Fusion Descriptions Classified using CySecBert, (b)Gemini Fusion [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. TraceRAG: A LLM-Based Framework for Explainable Android Malware Detection and Behavior Analysis

    cs.SE 2025-09 conditional novelty 4.0 of 10

    An LLM-based RAG framework that retrieves method-level Java code snippets to explain and detect malicious behavior in Android apps.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    Flowdroid: Precise context, flow, field, object-sensitive and lifecycle-aware taint analysis for android apps

    Steven Arzt, Siegfried Rasthofer, Christian Fritz, Eric Bodden, Alexandre Bartel, Jacques Klein, Yves Le Traon, Damien Octeau, and Patrick McDaniel. Flowdroid: Precise context, flow, field, object-sensitive and lifecycle-aware taint analysis for android apps. ACM sigplan notices, 49(6):259–269, 2014

    work page 2014

  2. [2]

    Android permissions demystified

    Adrienne Porter Felt, Erika Chin, Steve Hanna, Dawn Song, and David Wagner. Android permissions demystified. In Proceedings of the 18th ACM conference on Computer and communications security , pages 627–638, 2011

    work page 2011

  3. [3]

    Dissecting android malware: Characterization and evolution

    Yajin Zhou and Xuxian Jiang. Dissecting android malware: Characterization and evolution. In 2012 IEEE symposium on security and privacy , pages 95–109. IEEE, 2012

    2012

  4. [4]

    Drebin: Effective and explainable detection of android malware in your pocket

    Daniel Arp, Michael Spreitzenbarth, Malte Hubner, Hugo Gascon, Konrad Rieck, and CERT Siemens. Drebin: Effective and explainable detection of android malware in your pocket. In Ndss, volume 14(1), pages 23–26, 2014. 16 Enhancing Android Malware Detection with Retrieval-Augmented Generation

    work page 2014

  5. [5]

    Copperdroid: Automatic reconstruction of android malware behaviors

    Kimberly Tam, Aristide Fattori, Salahuddin Khan, and Lorenzo Cavallaro. Copperdroid: Automatic reconstruction of android malware behaviors. In NDSS Symposium 2015, pages 1–15, 2015

    work page 2015

  6. [6]

    Droidecho: an in-depth dissection of malicious behaviors in android applications

    Guozhu Meng, Ruitao Feng, Guangdong Bai, Kai Chen, and Yang Liu. Droidecho: an in-depth dissection of malicious behaviors in android applications. Cybersecurity, 1:1–17, 2018

    work page 2018

  7. [7]

    Yes, machine learning can be more secure! a case study on android malware detection

    Ambra Demontis, Marco Melis, Battista Biggio, Davide Maiorca, Daniel Arp, Konrad Rieck, Igino Corona, Giorgio Giacinto, and Fabio Roli. Yes, machine learning can be more secure! a case study on android malware detection. IEEE transactions on dependable and secure computing , 16(4):711–724, 2017

    work page 2017

  8. [8]

    Agentic information retrieval, 2025

    Weinan Zhang, Junwei Liao, Ning Li, Kounianhua Du, and Jianghao Lin. Agentic information retrieval, 2025

    2025

Show all 21 references
  1. [9]

    Droidsieve: Fast and accurate classification of obfuscated android malware

    Guillermo Suarez-Tangil, Santanu Kumar Dash, Mansour Ahmadi, Johannes Kinder, Giorgio Giacinto, and Lorenzo Cavallaro. Droidsieve: Fast and accurate classification of obfuscated android malware. In Proceedings of the seventh ACM on conference on data and application security a...

    2017

  2. [10]

    Mamadroid: Detecting android malware by building markov chains of behavioral models (extended version)

    Lucky Onwuzurike, Enrico Mariconti, Panagiotis Andriotis, Emiliano De Cristofaro, Gordon Ross, and Gianluca Stringhini. Mamadroid: Detecting android malware by building markov chains of behavioral models (extended version). ACM Transactions on Privacy and Security (TOPS) , 22(...

    2019

  3. [11]

    Droidchameleon: evaluating android anti-malware against transformation attacks

    Vaibhav Rastogi, Yan Chen, and Xuxian Jiang. Droidchameleon: evaluating android anti-malware against transformation attacks. In Proceedings of the 8th ACM SIGSAC symposium on Information, computer and communications security, pages 329–334, 2013

    2013

  4. [12]

    Evolution, detection and analysis of malware for smart devices

    Guillermo Suarez-Tangil, Juan E Tapiador, Pedro Peris-Lopez, and Arturo Ribagorda. Evolution, detection and analysis of malware for smart devices. IEEE communications surveys & tutorials , 16(2):961–987, 2013

    2013

  5. [13]

    In 21st USENIX security symposium (USENIX security 12) , pages 569–584, 2012

    Lok Kwong Yan and Heng Yin.{DroidScope}: Seamlessly reconstructing the {OS} and dalvik semantic views for dynamic android malware analysis. In 21st USENIX security symposium (USENIX security 12) , pages 569–584, 2012

    2012

  6. [14]

    Effective android malware detection with a hybrid model based on deep autoencoder and convolutional neural network

    Wei Wang, Mengxue Zhao, and Jigang Wang. Effective android malware detection with a hybrid model based on deep autoencoder and convolutional neural network. Journal of Ambient Intelligence and Humanized Computing , 10(8):3035–3043, 2019

    2019

  7. [15]

    Hindroid: An intelligent android malware detection system based on structured heterogeneous information network

    Shifu Hou, Yanfang Ye, Yangqiu Song, and Melih Abdulhayoglu. Hindroid: An intelligent android malware detection system based on structured heterogeneous information network. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining ,...

    2017

  8. [16]

    Summary of chatgpt-related research and perspective towards the future of large language models

    Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, et al. Summary of chatgpt-related research and perspective towards the future of large language models. Meta-radiology, 1(2):100017, 2023

    2023

  9. [17]

    Apppoet: Large language model based android malware detection via multi-view prompt engineering

    Wenxiang Zhao, Juntao Wu, and Zhaoyi Meng. Apppoet: Large language model based android malware detection via multi-view prompt engineering. Expert Systems with Applications , 262:125546, 2025

    2025

  10. [18]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595–46623, 2023

    2023

  11. [19]

    The research trends of corpus-assisted stance research (2004–2023): a systematic literature review

    Ya Sun, Kaiwen Yang, and Gong Chen. The research trends of corpus-assisted stance research (2004–2023): a systematic literature review. Current Psychology, pages 1–16, 2024

    2004

  12. [20]

    Cysecbert: A domain-adapted language model for the cybersecurity domain

    Markus Bayer, Philipp Kuehn, Ramin Shanehsaz, and Christian Reuter. Cysecbert: A domain-adapted language model for the cybersecurity domain. ACM Transactions on Privacy and Security , 27(2):1–20, 2024

    2024

  13. [21]

    Secbert: Privacy-preserving pre-training based neural network inference system

    Hai Huang and Yongjian Wang. Secbert: Privacy-preserving pre-training based neural network inference system. Neural Networks, 172:106135, 2024. 17

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.