Pith. sign in

REVIEW 4 major objections 7 minor 38 references

LLM-BSCVM: An LLM-Based Blockchain Smart Contract Vulnerability Management Framework

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 This paper claims that a six-agent, retrieval-augmented LLM framework can manage smart-contract vulnerabilities end to end—detecting them at over 91% accuracy and F1 while cutting the false positive rate to 5.1%—and that the full…

desk verdict Solid systems integration work undermined by a visible RAG leakage problem: the retrieval context in the paper's own figures contains ground-truth labels drawn from the same dataset used for evaluation, so the headline 91% numbers need a leakage-controlled rerun before they can be trusted. read the letter →

arxiv 2505.17416 v1 pith:3LFYQOHZ submitted 2025-05-23 cs.CR

Yanli Jin , Chunpei Li , Peng Fan , Peng Liu , Xianxian Li , Chen Liu , Wangjie Qiu This is my paper · ORCID
classification cs.CR
keywords smartcontractsecurityvulnerabilitymanagementlargelanguagemodelsretrieval-augmentedgenerationmulti-agentsystemsblockchaindetectionrepair
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 an LLM-based framework can handle the entire smart-contract vulnerability lifecycle, not just detection. It proposes LLM-BSCVM, which splits the work across six collaborating agents—detection, cause analysis, repair suggestion, risk assessment, repair, and verification—and feeds each agent retrieved knowledge at every step. The reported result is detection accuracy and F1 over 91% on benchmark contracts, on par with the state of the art, while the false positive rate drops from 7.2% to 5.1%. If that claim holds, developers and security teams get one open-source pipeline that explains, ranks, patches, and checks smart-contract bugs instead of stitching together separate tools and manual expert review.

What carries the argument

The load-bearing mechanism is the "Decompose-Retrieve-Generate" (DRG) three-stage method: decompose the management task into six agent subtasks, retrieve relevant knowledge from two knowledge bases at each step, then generate the next stage's output from that context. The retrieval layer is built from a smart-contract corpus (TF-IDF weighting and cosine similarity for top-k similar contracts) and a vulnerability knowledge base (vector embeddings over audit reports, best practices, and the SWC registry). Detection itself is a fusion of three signals: the LoRA fine-tuned CodeLlama model, static pattern checks, and retrieval-based similarity, combined either by weighted scores (70% model, 20% retrieval, 10% static) or by majority vote. The same DRG loop then carries the detected vulnerability through cause analysis, risk rating, repair, and independent patch verification with a separate large model.

What would settle it

Inspect the open-source code and dataset: if a contract in the test set, or its audit report, also appears in the smart-contract corpus used for top-k retrieval or in the LoRA fine-tuning set, the reported 91% accuracy and 5.1% false positive rate would be inflated; a clean split with no overlapping contracts and no retrieved ground-truth labels would confirm them.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single LLM-based framework, LLM-BSCVM, can run the full smart-contract vulnerability lifecycle in one pass: detect whether a contract is vulnerable, explain the root cause, generate a repair suggestion, assign a risk level, produce a patched contract, verify the patch, and emit an audit report. On the benchmark built from TrustLLM audit reports, the weighted-fusion version LLM-BSCVM(W) reports 91.11% accuracy, 95.06% precision, 87.43% recall, and 91.04% F1, with a false positive rate of 5.1% compared with TrustLLM's 7.2%. The claim is that fusing a LoRA fine-tuned detector with lightweight static pattern analysis and top-k retrieval beats any single component, and that the multi-agent pipeline adds the explainability and repair capabilities that detection-only tools lack.

Load-bearing premise

The load-bearing premise is that the retrieval corpus and fine-tuning data do not leak ground-truth labels from the evaluation set; the paper does not report de-duplication, data-split, or leakage checks.

Editorial extensions

If this is right

  • If the reported numbers are right, a developer can submit a contract and get a detection verdict, root-cause explanation, risk level, concrete patch, and verification in one open-source run.
  • The 5.1% false positive rate, if it holds, means about two percentage points fewer clean contracts are flagged than by the TrustLLM baseline, making automated audits less noisy.
  • Weighted fusion outperforms both majority voting and giving the model all retrieved context, so the paper's design suggests keeping raw retrieved contracts out of the LLM prompt and using them as a separate scoring signal.
  • Patch verification currently passes only about 21% of repaired contracts, so the framework's repair stage is the clear bottleneck; improving patch correctness would matter more than further detection gains.
  • Because the knowledge base can be hot-swapped, the framework can absorb new audit reports and vulnerability classes without retraining the model, supporting continuous monitoring.

Reading between the lines

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

  • A natural extension the paper leaves implicit is to use the Verifier's rejection reasons as feedback that drives a second repair round, turning the 21% pass rate into an iterative fix-and-check loop.
  • The Decompose-Retrieve-Generate scaffold seems transferable to other code-security tasks, such as auditing non-EVM chains or reviewing upgradeable-contract migrations, because the agents consume retrieved documents rather than hard-coded rules.
  • The paper does not report how detection quality degrades when the knowledge base is empty or outdated; a sensitivity test varying the top-k value and corpus freshness would show how much of the 91% F1 comes from retrieval versus the fine-tuned model.
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 proposes LLM-BSCVM, an LLM-based smart contract vulnerability management framework that combines multi-agent collaboration with retrieval-augmented generation (RAG) in a 'Decompose-Retrieve-Generate' pipeline. The framework covers vulnerability detection, cause analysis, repair suggestion generation, risk assessment, automated repair, and patch verification. Detection combines a LoRA-fine-tuned CodeLlama model, static pattern analysis, and top-k similar-contract retrieval, with weighted fusion or majority voting. Experiments on the TrustLLM-derived dataset report 91.11% accuracy, 95.06% precision, 91.04% F1, and a 5.1% false positive rate, compared with 7.2% for TrustLLM; ablations show drops when static analysis or RAG is removed; a repair experiment reports about 21% of patches passing validation. The paper claims this is the first end-to-end smart contract vulnerability management framework and open-sources the code.

Significance. If the reported results are valid, LLM-BSCVM would be a practically useful, open-source contribution: it integrates detection, explanation, repair, and verification in one pipeline, and the ablation study gives some evidence that the static and RAG components contribute. The architecture is clearly presented and the comparison against several base and fine-tuned LLMs is useful. However, the headline detection numbers are compromised by a serious data-leakage risk, and the repair-stage evaluation is anecdotal. The significance of the work is therefore conditional on a leakage-controlled re-evaluation and a more systematic repair study.

major comments (4)
  1. [Sections III-C, IV-A, and Figures 1-2] The evaluation is not leakage-controlled. Section III-C states that the Smart Contract Corpus is sourced from reference [13] (TrustLLM), consisting of 263 audit reports from Solodit, and Section IV-A says evaluation uses the same dataset as TrustLLM. Figures 1 and 2 show retrieved RAG context chunks that contain the literal field 'ground_truth_label' (e.g., 'function _transfer ... ground_truth_label : ...'). Since retrieval is by cosine similarity over contract code, a test contract can retrieve itself or a near-duplicate from the same corpus and directly receive its ground-truth label. The paper reports no train/test split, no deduplication, and no check that fine-tuning data and the retrieval corpus exclude test contracts. Because the retrieval component contributes 20% of the weighted decision and the RAG context is also included in prompts, the reported 91.11% accuracy, 95.06% precision, and 5.1% FPR are not interpretable as evidence of detection quality. A leakage-controlled rerun with labels stripped, duplicates removed, and test contracts excluded from both the retrieval corpus and the fine-tuning set is required before the headline claims can be assessed.
  2. [Sections IV-A and IV-B, Tables I and IV] The fusion weights (model 70%, static analysis 10%, retrieval 20%) and the value k=5 are reported as fixed choices, but no tuning protocol or sensitivity analysis is given. The headline result is the weighted-fusion variant LLM-BSCVM(W); without a validation-set separation or a scan over k and the fusion weights, the reported advantage over LLM-BSCVM(V) and over the ablation variants could be an artifact of hyperparameters selected on the same benchmark. Please report sensitivity to k and to the weight distribution, or justify the chosen values using a held-out validation split.
  3. [Section IV-B, 'Vulnerability repair'] The repair stage, which is central to the claimed end-to-end capability, is evaluated only as 'approximately 21% of the contracts successfully passed the validation' on an unspecified subset, with a single illustrative example. There is no definition of what 'passed the validation' means, no per-vulnerability repair rate, no comparison to a baseline repair method, and no quantitative error analysis. Given that the abstract and introduction emphasize automated repair as a novel contribution, this is not sufficient support; the paper should either present a systematic repair evaluation or explicitly limit the claim to detection and analysis.
  4. [Section IV-B, Tables I-III] No error bars, confidence intervals, or significance tests are reported for any metric. The benchmark is small (263 audit reports), and the F1/accuracy values of LLM-BSCVM(W) (0.9104/0.9111) are nearly identical to those of TrustLLM (0.9121/0.9111), so the reported FPR reduction from 7.2% to 5.1% may be within noise. Report per-run variance or confidence intervals and, where possible, a paired test over the same contracts.
minor comments (7)
  1. [Table III] The row labeled 'Ilama 8B' should be 'Llama 8B'.
  2. [Section IV-B] The variant LLM-BSCVM(E) is used in Table I but never defined; the text defines only (W) and (V).
  3. [Section III-C, Equation (1)] The text says 'As shown in Formula 1' but the formula is not numbered, and the notation Va/Vb is not connected to the embedding model described in the text.
  4. [Section III-C] The corpus source is spelled 'Solidit' in one place and 'Solodit' elsewhere; standardize the spelling to match the reference.
  5. [Section IV-B] The false positive rate of 5.1% is a headline number but is never defined or tabulated; please provide a definition and the underlying confusion-matrix counts.
  6. [Tables II and III] It is unclear whether the 'TrustLLM' baseline rows are values copied from the TrustLLM paper or numbers obtained by re-running TrustLLM on this dataset; please clarify.
  7. [Section IV-A] The setup says the dataset is compiled from TrustLLM and Dappscan, but Section IV-B says the detection experiments use 'the same dataset as TrustLLM'; clarify whether Dappscan data is used anywhere in the reported experiments.

Circularity Check

1 steps flagged · score 6.0 of 10

Retrieval corpus equals the evaluation set: the 91% detection result partly re-inserts the test contracts' own ground-truth labels as retrieval context.

  1. fitted input called prediction [Section III-C (Knowledge Retrieval Stage); Section IV-B (Detection evaluation); Figures 1-2]
    "Smart Contract Corpus: ... The data is sourced from [13], and collected from the well-known auditing website Solodit [30], analyzing a total of 263 smart contract audit reports. ... The experiments were conducted using the same dataset as TrustLLM. ... Chunk 1:function _transfer( address _from, address ..... ground_truth_label : ....."

    Section III-C builds the Smart Contract Corpus from the same 263 TrustLLM/Solodit audit reports, and Section IV-B runs detection on 'the same dataset as TrustLLM'; retrieval computes 'the final probability of vulnerability presence through weighted computation' over similar contracts, and the retrieved chunks in Figures 1-2 contain the field 'ground_truth_label' (e.g., 'function _transfer(... ground_truth_label : .....'). With k=5 and no described split or deduplication, the corpus holds each test contract (or a near-duplicate with its label), so the retrieval component (20% weight in LLM-BSCVM(W)) puts the contract's own ground-truth label into the classification input.

full rationale

The paper is not globally circular: its multi-agent Decompose-Retrieve-Generate pipeline, the LoRA fine-tuned CodeLlama detector, the qualitative preSign repair/risk case studies, and the comparison against external baselines are independent contributions, and there is no load-bearing author self-citation (ref. [18] is motivational only; TrustLLM [13] is by different authors). The circularity is localized to the headline detection numbers. The Smart Contract Corpus used for the RAG detection path is built from the same 263 TrustLLM/Solodit audit reports that constitute the evaluation benchmark, with no described split or deduplication, and the paper's own Figures 1-2 show retrieved chunks carrying the literal text 'ground_truth_label' into the detection prompt. Consequently, the retrieval-based detection score is, in part, a readout of the test contract's own stored audit label rather than an independent prediction, and the weighted fusion (20% retrieval weight) inherits that forced component. The undisclosed fine-tuning split for the 70%-weight model further weakens independence, though it cannot be exhibited as a construction-level reduction from the text alone. These issues inflate the claimed 91.11% accuracy, 0.9104 F1, and 5.1% FPR relative to TrustLLM; a leakage-controlled rerun with a held-out corpus and a deduplicated fine-tuning split is required to substantiate the headline numbers. Because two of the three detection components (fine-tuned model, static analysis) are not themselves reduced to the labels, and the non-detection tasks (cause analysis, repair, risk, reporting) are demonstrated qualitatively, the circularity is partial: score 6.

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

The six software agents are role configurations of existing LLMs, not new physical or formal entities; no new particles, mediators, or conserved quantities are introduced. The main unfunded dependencies are the ground-truth assumption for audit labels, the correctness of the LLM-as-judge for repair verification, and the absence of a leakage check between the retrieval corpus and the benchmark.

free parameters (3)
  • Detection fusion weights (model, retrieval, static) = 0.70, 0.20, 0.10
    Chosen for the weighted LLM-BSCVM (W) variant; no tuning protocol or held-out validation is reported, yet these weights directly determine the reported 91.11% accuracy.
  • Top-k retrieval count = 5
    Set in Section IV-A with no sensitivity analysis; retrieval similarity affects both detection and the knowledge supplied to other agents.
  • Rank-based similarity weights and decision threshold = not specified
    Section III-C states higher-ranked retrieved contracts receive greater weight and a dynamic threshold is used, but the exact formula and threshold are not given, so the final vulnerability probability is not fully determined by the stated method.
assumptions (5)
  • domain assumption Fine-tuned CodeLlama and the agent LLM can judge Solidity code vulnerabilities.
    The entire detection and repair pipeline assumes the underlying models' code understanding transfers to smart contracts.
  • domain assumption Audit reports from TrustLLM/Solodit and Dappscan provide correct ground-truth vulnerability labels.
    All accuracy, precision, recall, F1, and FPR numbers are computed against these labels; label errors propagate directly into the metrics.
  • domain assumption GPT-4 as an independent verifier is a valid oracle for whether a patch fixed the vulnerability.
    The 21% repair success rate is defined by this judge's verdict, not by re-running detection, symbolic execution, or formal verification.
  • ad hoc to paper The RAG knowledge base built from the TrustLLM corpus does not leak benchmark labels through similar-contract retrieval.
    No de-duplication or leakage test is reported, yet the same source feeds both the retrieval corpus and the evaluation benchmark.
  • domain assumption Static analysis pattern library covers the relevant vulnerability classes with known accuracy.
    Static analysis contributes 10% of the weighted fusion; its coverage and error behavior are not characterized separately.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-BSCVM: An LLM-Based Blockchain Smart Contract Vulnerability Management Framework." pith.science (2026). https://pith.science/paper/3LFYQOHZ

@misc{pith2026250517416,
  author       = {Pith},
  title        = {Pith review of: LLM-BSCVM: An LLM-Based Blockchain Smart Contract Vulnerability Management Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3LFYQOHZ}},
  note         = {Machine review of arXiv:2505.17416}
}
read the original abstract

Smart contracts are a key component of the Web 3.0 ecosystem, widely applied in blockchain services and decentralized applications. However, the automated execution feature of smart contracts makes them vulnerable to potential attacks due to inherent flaws, which can lead to severe security risks and financial losses, even threatening the integrity of the entire decentralized finance system. Currently, research on smart contract vulnerabilities has evolved from traditional program analysis methods to deep learning techniques, with the gradual introduction of Large Language Models. However, existing studies mainly focus on vulnerability detection, lacking systematic cause analysis and Vulnerability Repair. To address this gap, we propose LLM-BSCVM, a Large Language Model-based smart contract vulnerability management framework, designed to provide end-to-end vulnerability detection, analysis, repair, and evaluation capabilities for Web 3.0 ecosystem. LLM-BSCVM combines retrieval-augmented generation technology and multi-agent collaboration, introducing a three-stage method of Decompose-Retrieve-Generate. This approach enables smart contract vulnerability management through the collaborative efforts of six intelligent agents, specifically: vulnerability detection, cause analysis, repair suggestion generation, risk assessment, vulnerability repair, and patch evaluation. Experimental results demonstrate that LLM-BSCVM achieves a vulnerability detection accuracy and F1 score exceeding 91\% on benchmark datasets, comparable to the performance of state-of-the-art (SOTA) methods, while reducing the false positive rate from 7.2\% in SOTA methods to 5.1\%, thus enhancing the reliability of vulnerability management. Furthermore, LLM-BSCVM supports continuous security monitoring and governance of smart contracts through a knowledge base hot-swapping dynamic update mechanism.

Figures

Figures reproduced from arXiv: 2505.17416 by the authors.

Figure 1
Figure 1. Framework of our proposed approach LLM-BSCVM. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The specific of task decomposition. to the model for the final decision, as experiments show that excessive external information may introduce noise, affecting the accuracy of detection. Repair Suggestion Agent (Advisor): After vulnerability detection, the repair suggestion agent is responsible for pro￾viding targeted repair solutions for the detected vulnerabilities. Using RAG technology, it retrieves real-time inf… view at source ↗
Figure 3
Figure 3. The specific of knowledge retrieval. including smart contract audit reports [31] from renowned security institutions, security best practices [32], and pro￾gramming standard documents [33]. These documents are primarily sourced from leading companies in the industry, such as Solidit [30] and Smart Contract Weakness Clas￾sification (SWC) [34]. The vulnerability knowledge base provides relevant documents and backgroun… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Prompt Template of LLM-BSCVM: An Example for Vulnerability [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: illustrates the repair suggestions generated by LLM-BSCVM, covering five aspects: vulnerability name, cause analysis, potential impact assessment, specific repair steps, and preventive measures recommendations. These sug￾gestions provide a detailed description of the v…
Figure 5
Figure 5. Figure 5: Example Contract: preSign Contract. that static analysis effectively identifies common and easily detectable vulnerability patterns, such as reentrancy attacks and integer overflows, which can often be recognized through simple pattern matching. Consequently, the remov…
Figure 8
Figure 8. Figure 8: Vulnerability repair Results of LLM-BSCVM. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [13]

    Combining fine-tuning and llm-based agents for intuitive smart contract auditing with justifications,

    W. Ma, D. Wu, Y . Sun, T. Wang, S. Liu, J. Zhang, Y . Xue, and Y . Liu, “Combining fine-tuning and llm-based agents for intuitive smart contract auditing with justifications,” arXiv preprint arXiv:2403.16073 , 2024

    arXiv 2024

  2. [1]

    Addressing the dao insider attack in rpl’s internet of things networks,

    B. Ghaleb, A. Al-Dubai, E. Ekonomou, M. Qasem, I. Romdhani, and L. Mackenzie, “Addressing the dao insider attack in rpl’s internet of things networks,” IEEE Communications Letters, vol. 23, no. 1, pp. 68– 71, 2018

    work page 2018

  3. [2]

    Batch overflow bug on ethereum erc20 token contracts and safemath,

    S. Hessenauer, “Batch overflow bug on ethereum erc20 token contracts and safemath,” 2018

    work page 2018

  4. [3]

    Smart contract: Attacks and protections,

    S. Sayeed, H. Marco-Gisbert, and T. Caira, “Smart contract: Attacks and protections,” Ieee Access, vol. 8, pp. 24 416–24 427, 2020

    work page 2020

  5. [4]

    Zeus: analyzing safety of smart contracts

    S. Kalra, S. Goel, M. Dhawan, and S. Sharma, “Zeus: analyzing safety of smart contracts.” in Ndss, 2018, pp. 1–12

    work page 2018

  6. [5]

    Making smart contracts smarter,

    L. Luu, D.-H. Chu, H. Olickel, P. Saxena, and A. Hobor, “Making smart contracts smarter,” in Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , 2016, pp. 254–269

    2016

  7. [6]

    Contractfuzzer: Fuzzing smart con- tracts for vulnerability detection,

    B. Jiang, Y . Liu, and W. K. Chan, “Contractfuzzer: Fuzzing smart con- tracts for vulnerability detection,” in Proceedings of the 33rd ACM/IEEE international conference on automated software engineering , 2018, pp. 259–269

    work page 2018

  8. [7]

    Vulnerable smart contract function locating based on multi-relational nested graph convolutional network,

    H. Liu, Y . Fan, L. Feng, and Z. Wei, “Vulnerable smart contract function locating based on multi-relational nested graph convolutional network,” Journal of Systems and Software , vol. 204, p. 111775, 2023

    work page 2023

Show all 38 references
  1. [8]

    Diversevul: A new vulnerable source code dataset for deep learning based vulnerability detection,

    Y . Chen, Z. Ding, L. Alowain, X. Chen, and D. Wagner, “Diversevul: A new vulnerable source code dataset for deep learning based vulnerability detection,” in Proceedings of the 26th International Symposium on Research in Attacks, Intrusions and Defenses , 2023, pp. 654–668

    2023

  2. [9]

    Contractward: Automated vulnerability detection models for ethereum smart contracts,

    W. Wang, J. Song, G. Xu, Y . Li, H. Wang, and C. Su, “Contractward: Automated vulnerability detection models for ethereum smart contracts,” IEEE Transactions on Network Science and Engineering , vol. 8, no. 2, pp. 1133–1144, 2020

    2020

  3. [10]

    Da-gnn: A smart contract vulnerability detection method based on dual attention graph neural network,

    Z. Zhen, X. Zhao, J. Zhang, Y . Wang, and H. Chen, “Da-gnn: A smart contract vulnerability detection method based on dual attention graph neural network,” Computer Networks, vol. 242, p. 110238, 2024

    2024

  4. [11]

    A review on code generation with llms: Application and evaluation,

    J. Wang and Y . Chen, “A review on code generation with llms: Application and evaluation,” in 2023 IEEE International Conference on Medical Artificial Intelligence (MedAI) . IEEE, 2023, pp. 284–289

    2023

  5. [12]

    Gptscan: Detecting logic vulnerabilities in smart contracts by combining gpt with program analysis,

    Y . Sun, D. Wu, Y . Xue, H. Liu, H. Wang, Z. Xu, X. Xie, and Y . Liu, “Gptscan: Detecting logic vulnerabilities in smart contracts by combining gpt with program analysis,” in Proceedings of the IEEE/ACM 46th International Conference on Software Engineering, 2024, pp. 1–13

    2024

  6. [14]

    Large language model-powered smart contract vulnerability detection: New perspec- tives,

    S. Hu, T. Huang, F. Ilhan, S. F. Tekin, and L. Liu, “Large language model-powered smart contract vulnerability detection: New perspec- tives,” in 2023 5th IEEE International Conference on Trust, Privacy and Security in Intelligent Systems and Applications (TPS-ISA) . IEEE, 202...

    2023

  7. [15]

    A survey on smart contract vulnerabilities: Data sources, detection and repair,

    H. Chu, P. Zhang, H. Dong, Y . Xiao, S. Ji, and W. Li, “A survey on smart contract vulnerabilities: Data sources, detection and repair,” Information and Software Technology, vol. 159, p. 107221, 2023

    2023

  8. [16]

    Vul- nerabilities in smart contracts: A detailed survey of detection and mit- igation methodologies,

    N. K. Kumar, N. V . Honnungar, M. S. Prakash, and J. Lohith, “Vul- nerabilities in smart contracts: A detailed survey of detection and mit- igation methodologies,” in 2024 International Conference on Emerging Technologies in Computer Science for Interdisciplinary Applications ...

    2024

  9. [17]

    Reguard: finding reentrancy bugs in smart contracts,

    C. Liu, H. Liu, Z. Cao, Z. Chen, B. Chen, and B. Roscoe, “Reguard: finding reentrancy bugs in smart contracts,” in Proceedings of the 40th International Conference on Software Engineering: Companion Proceeedings, 2018, pp. 65–68

    2018

  10. [18]

    Exploring {ChatGPT’s} capabilities on vulnerability management,

    P. Liu, J. Liu, L. Fu, K. Lu, Y . Xia, X. Zhang, W. Chen, H. Weng, S. Ji, and W. Wang, “Exploring {ChatGPT’s} capabilities on vulnerability management,” in 33rd USENIX Security Symposium (USENIX Security 24), 2024, pp. 811–828

    2024

  11. [19]

    Agentverse: Facilitating multi-agent col- laboration and exploring emergent behaviors in agents,

    W. Chen, Y . Su, J. Zuo, C. Yang, C. Yuan, C. Qian, C.-M. Chan, Y . Qin, Y . Lu, R. Xie et al. , “Agentverse: Facilitating multi-agent col- laboration and exploring emergent behaviors in agents,” arXiv preprint arXiv:2308.10848, vol. 2, no. 4, p. 6, 2023

    2023 arXiv

  12. [20]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschel et al. , “Retrieval- augmented generation for knowledge-intensive nlp tasks,” Advances in neural information processing systems , vol. 33, pp. 9459–9474, 2020

    2020

  13. [21]

    Soliaudit: Smart contract vulnerability assessment based on machine learning and fuzz testing,

    J.-W. Liao, T.-T. Tsai, C.-K. He, and C.-W. Tien, “Soliaudit: Smart contract vulnerability assessment based on machine learning and fuzz testing,” in 2019 Sixth International Conference on Internet of Things: Systems, Management and Security (IOTSMS) . IEEE, 2019, pp. 458– 465

    2019

  14. [22]

    Manticore: A user-friendly symbolic execution framework for binaries and smart contracts,

    M. Mossberg, F. Manzano, E. Hennenfent, A. Groce, G. Grieco, J. Feist, T. Brunson, and A. Dinaburg, “Manticore: A user-friendly symbolic execution framework for binaries and smart contracts,” in 2019 34th IEEE/ACM International Conference on Automated Software Engineer- ing (A...

    2019

  15. [23]

    Wana: Symbolic execution of wasm bytecode for extensible smart contract vulnerability detection,

    B. Jiang, Y . Chen, D. Wang, I. Ashraf, and W. K. Chan, “Wana: Symbolic execution of wasm bytecode for extensible smart contract vulnerability detection,” in 2021 IEEE 21st International Conference on Software Quality, Reliability and Security (QRS) . IEEE, 2021, pp. 926–937

    2021

  16. [24]

    Verismart: A highly precise safety verifier for ethereum smart contracts,

    S. So, M. Lee, J. Park, H. Lee, and H. Oh, “Verismart: A highly precise safety verifier for ethereum smart contracts,” in 2020 IEEE Symposium on Security and Privacy (SP) . IEEE, 2020, pp. 1678–1694

    2020

  17. [25]

    Llmsmartsec: Smart contract security auditing with llm and annotated control flow graph,

    V . Mothukuri, R. M. Parizi, and J. L. Massa, “Llmsmartsec: Smart contract security auditing with llm and annotated control flow graph,” in 2024 IEEE International Conference on Blockchain (Blockchain) . IEEE, 2024, pp. 434–441

    2024

  18. [26]

    Llm4vuln: A unified evaluation framework for decoupling and enhanc- ing llms’ vulnerability reasoning,

    Y . Sun, D. Wu, Y . Xue, H. Liu, W. Ma, L. Zhang, Y . Liu, and Y . Li, “Llm4vuln: A unified evaluation framework for decoupling and enhanc- ing llms’ vulnerability reasoning,” arXiv preprint arXiv:2401.16185 , 2024

    2024 arXiv

  19. [27]

    Attention is all you need,

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

    2017

  20. [28]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog , vol. 1, no. 8, p. 9, 2019

    2019

  21. [29]

    Code llama: Open foundation models for code,

    B. Roziere, J. Gehring, F. Gloeckle, S. Sootla, I. Gat, X. E. Tan, Y . Adi, J. Liu, R. Sauvestre, T. Remez et al. , “Code llama: Open foundation models for code,” arXiv preprint arXiv:2308.12950 , 2023

    2023 arXiv

  22. [30]

    Solodit - all findings from popular audit platforms,

    Solodit, “Solodit - all findings from popular audit platforms,” https: //solodit.xyz/, 2024, accessed: 2024-03-07

    2024

  23. [31]

    Dappscan: building large-scale datasets for smart contract weaknesses in dapp projects,

    Z. Zheng, J. Su, J. Chen, D. Lo, Z. Zhong, and M. Ye, “Dappscan: building large-scale datasets for smart contract weaknesses in dapp projects,” IEEE Transactions on Software Engineering , 2024

    2024

  24. [32]

    Best practices for smart contract development,

    Y . Riady, “Best practices for smart contract development,” https://yos.io/ 2019/11/10/smart-contract-development-best-practices/, 2019, accessed: 2024-03-07

    2019

  25. [33]

    Eea Ethtrust Security Levels Specifica- tion Version 2,

    Ethereum Enterprise Alliance, “Eea Ethtrust Security Levels Specifica- tion Version 2,” https://www.eea.ethereum.org/ethtrust/, 2024, accessed: 2024-03-07

    2024

  26. [34]

    Smart Contract Weakness Classification and Test Cases,

    SmartContractSecurity, “Smart Contract Weakness Classification and Test Cases,” https://github.com/SmartContractSecurity/SWC-registry, 2020, accessed: 2024-03-07

    2020

  27. [35]

    Codebert: A pre-trained model for programming and natural languages,

    Z. Feng, D. Guo, D. Tang, N. Duan, X. Feng, M. Gong, L. Shou, B. Qin, T. Liu, D. Jiang et al., “Codebert: A pre-trained model for programming and natural languages,” arXiv preprint arXiv:2002.08155 , 2020

    2002 arXiv

  28. [36]

    Codet5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation,

    Y . Wang, W. Wang, S. Joty, and S. C. Hoi, “Codet5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation,” arXiv preprint arXiv:2109.00859 , 2021

    2021 arXiv

  29. [37]

    Llama: Open and efficient foundation language models,

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

    2023 arXiv

  30. [38]

    Contract- tinker: Llm-empowered vulnerability repair for real-world smart con- tracts,

    C. Wang, J. Zhang, J. Gao, L. Xia, Z. Guan, and Z. Chen, “Contract- tinker: Llm-empowered vulnerability repair for real-world smart con- tracts,” in Proceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering , 2024, pp. 2350–2353

    2024

Pith tools

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