DG-VDT begins after an EVM execution trace exists. It is therefore closer to a runtime detector and forensic classifier than to a source-code auditor or exploit-generation agent. The system decomposes one observed execution into three graph views: a Fund Flow Graph for value movement, a Contract Creation Graph for deployment provenance, and a Contract Call…
Inspect the ClaimsCase · DiggingBeagle record
DG-VDT uses multi-graph reinforcement learning for Ethereum vulnerability detection and attacker traceability
A September 2026 Applied Sciences study reports that DG-VDT converts EVM execution traces into fund-flow, contract-creation and contract-call graphs and uses graph-guided reinforcement learning for vulnerability classification and attacker tracing. The strongest external evidence is zero-shot detection on SolidiFI-Bench and SmartBugs Wild; the result remains limited to three vulnerability classes, while independent traceability validation and the full reproducible training stack remain unavailable.
Scope
Defensive post-execution research on EVM trace classification and attacker-address traceability. The evidence supports three vulnerability classes only: reentrancy, short-address attack and timestamp dependence. It does not establish general smart-contract vulnerability detection, live exploitation, autonomous exploit generation, victim impact or realized financial loss.
At a glance
DG-VDT is most interesting as evidence that structural execution information can materially improve a language-model-backed detector when the same model backbone is deprived of those graph views. The 13.9 macro-F1 text-only ablation is therefore more informative about the multi-graph contribution than the headline comparison with fine-tuned GPT-4o, because…
Read the implicationsNo independent reproduction of the reported DG-VDT benchmark results was found in this research pass. The full EVM graph-extraction pipeline, GRPO training/evaluation code, author-constructed datasets and pretrained model weights are not publicly deposited, so the complete benchmark cannot yet be independently rerun from public artifacts. The Applied…
Limits and uncertaintyFull account
DG-VDT begins after an EVM execution trace exists. It is therefore closer to a runtime detector and forensic classifier than to a source-code auditor or exploit-generation agent.
The system decomposes one observed execution into three graph views: a Fund Flow Graph for value movement, a Contract Creation Graph for deployment provenance, and a Contract Call Graph for invocation and opcode-level structure. An RGCN encoder maps observed and reference graphs into embeddings. Early reinforcement-learning steps reward structural similarity to a canonical attack signature; later steps increasingly require VF2 subgraph-isomorphism matching, with collaboration and parameter terms adding signal when multiple graph views agree.
The policy then emits both a vulnerability class and an attacker address. That joint output matters because the task is not merely to say that a trace looks suspicious; the model is trained to associate a recognized structural pattern with the address attributed to the attacking role.
The strongest generalization result is author-run zero-shot evaluation on third-party corpora. SolidiFI-Bench supplies injected vulnerabilities with known ground truth. For that benchmark the authors compile and deploy contracts on a local EVM fork, exercise known vulnerable paths, extract execution traces and then apply DG-VDT. SmartBugs Wild provides real-world contracts, but the evaluated 11,423-contract subset is labeled through Slither rather than human ground truth, so its 84.0% macro-F1 is better read as a large-scale robustness result than as a clean independent accuracy audit.
This places DG-VDT at a different point in the automation chain from existing DiggingBeagle smart-contract Cases. EchoFuzz uses LLM reasoning and fuzzing feedback to search for vulnerable state-transition sequences. KASS starts from vulnerability information and generates executable Foundry exploit tests. EVAge generates and repairs MEV strategies and validates them on forked historical state. DG-VDT instead recognizes and traces three predefined attack structures from execution evidence. The systems are adjacent, not interchangeable.
Timeline
- Sep 30, 2026Event type unspecified
Applied Sciences publishes DG-VDT
Sun and Jiang publish the DG-VDT study in Applied Sciences, volume 16 issue 19, article 9698.
- Sep 30, 2026disclosure
Public disclosure
Claims & evidence
CLM-DGVDT-SCOPEThe reported detection results are limited to reentrancy, short-address attack and timestamp dependence on EVM-based chains; the paper explicitly leaves broader classes including flash-loan and access-control attacks, adversarially robust reference-graph construction and non-EVM scenarios for future work.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Abstract; Conclusions, limitations and future-work discussion
CLM-DGVDT-COMPUTEThe paper reports approximately 6 GPU-hours on one RTX 3090 for RGCN pre-training, about 48 GPU-hours on 4xA100-80GB for DG-VDT-7B GRPO training, and about 180 GPU-hours on 8xA100-80GB for DG-VDT-32B. These are one-time training costs and are distinct from the reported per-trace inference latency.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Methodology, 'Computational cost' paragraph
- contextEcho-wxy/DG-VDT reference implementationprimary disclosure
Code Information and 'Hardware used in the paper'
CLM-DGVDT-LATENCYThe authors report approximately 180 ms model inference latency per trace on a single RTX 3090-class consumer GPU. This is not end-to-end smart-contract audit latency: the SolidiFI protocol first compiles and deploys contracts, uses a transaction-generation harness to trigger known vulnerable paths, and extracts EVM traces before DG-VDT inference.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Abstract latency result; comparison-systems latency note; Section 4.3.1 SolidiFI-Bench evaluation protocol
CLM-DGVDT-SOLIDIFIOn the author-independent SolidiFI-Bench zero-shot evaluation covering 3,942 instances in the three target categories, the authors report 87.7% macro-F1 for DG-VDT-7B, 6.8 points above their fine-tuned GPT-4o baseline.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Abstract; Section 4.3.1 'SolidiFI-Bench: Fault-Injection Ground Truth'; Table 4
CLM-DGVDT-ARCHITECTUREDG-VDT represents each EVM execution trace as three complementary graphs - fund flow, contract creation and contract calls - and trains a policy with a dual-stage graph reward that transitions from embedding similarity to strict subgraph-isomorphism matching.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Abstract; Sections 3.1-3.3 describing the multi-graph representation, reference graphs, dual-stage reward and GRPO training
- contextEcho-wxy/DG-VDT reference implementationprimary disclosure
README, Description section; repository root and dgvdt/ reference implementation
CLM-DGVDT-RELEASE-STATUSThe Applied Sciences article is publicly published with a September 30, 2026 publication date, while the checked DG-VDT repository still describes the manuscript as under review and says the unreleased datasets, full training/evaluation pipeline and pretrained weights will be released upon acceptance. The public record therefore shows a release-status lag or stale repository wording, not completion of the promised reproducibility release.supported
Basis: direct observation
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Article header and Data Availability Statement
- supportsEcho-wxy/DG-VDT reference implementationprimary disclosure
README opening paragraph; Dataset Information; Usage Instructions item 3
CLM-DGVDT-SMARTBUGS-WILDOn 11,423 SmartBugs Wild contracts evaluated zero-shot for the three target categories, the authors report 84.0% macro-F1 for DG-VDT-7B with a 95% bootstrap interval of 83.4-84.7%, 6.3 points above their fine-tuned GPT-4o baseline; the paper treats this dataset as a robustness signal because its ground truth uses noisier Slither single-tool labels.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Section 4.3.2 'SmartBugs Wild: Large-Scale In-the-Wild Validation'; Table 5
CLM-DGVDT-BASELINE-PARITYThe comparison against fine-tuned GPT-4o is not training-data matched: the paper states that GPT-4o was fine-tuned on 500 BlockTrace-500k examples while DG-VDT was trained on 489,939 traces, so the reported 6.8- and 6.3-point margins do not isolate architecture under equal training-data exposure.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Baseline comparison / training-data parity note and Section 4.3 discussion of comparison limitations
CLM-DGVDT-REPRODUCIBILITYThe authors publicly expose the reward engine, graph schema, RGCN encoder, canonical reference graphs and unit tests, but the author-constructed datasets, full GRPO training/evaluation pipeline and pretrained model weights are not yet publicly deposited; the paper and repository say they are available to editors/reviewers and are planned for later release.supported
Basis: direct observation
- supportsEcho-wxy/DG-VDT reference implementationprimary disclosure
README, opening description and Dataset Information sections; repository contents including dgvdt/ and dgvdt_tests.py
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Data Availability Statement
CLM-DGVDT-NOT-LIVE-EXPLOITDG-VDT is a defensive research and benchmark result for detection and traceability from execution traces; the cited material does not establish autonomous exploit generation, production exploitation, victim impact or realized financial loss caused by the system.supported
Basis: inference
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Abstract, experimental scope and Conclusions
CLM-DGVDT-TRACEABILITY-LIMITThe paper does not provide an author-independent benchmark for attacker traceability: traceability is evaluated on ScamTrace-2024 and CrossChainAtt, both assembled with author involvement, and the authors characterize those traceability results as preliminary.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Limitations discussion headed 'Independent evaluation of attacker traceability'
CLM-DGVDT-MULTIGRAPH-ABLATIONThe paper reports a 13.9 macro-F1-point drop when the multi-graph representation is removed from the 7B configuration, supporting the authors' claim that the reported gain is not explained only by the language-model backbone.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Abstract; ablation analysis and Conclusions discussing the text-only ablation and contribution of the multi-graph representation
CLM-DGVDT-INDEPENDENCE-BOUNDARYSolidiFI-Bench, SmartBugs Curated and SmartBugs Wild are third-party corpora independent of the DG-VDT authors, but their use does not constitute independent reproduction of DG-VDT: the reported runs and metrics are still produced by the authors, while the public repository does not yet include the complete experimental pipeline or pretrained weights.supported
Basis: inference
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Sections 4.3.1-4.3.2 and Conclusions describing author-independent public benchmarks; Data Availability Statement
- supportsEcho-wxy/DG-VDT reference implementationprimary disclosure
Dataset Information; Code Information; Usage Instructions item 3 listing experimental artifacts not yet publicly released
CLM-DGVDT-TRAIN-EVAL-SEPARATIONThe paper reports transaction-hash and contract-address-level deduplication between BlockTrace-500k and evaluation corpora. After removing overlaps, the final training corpus contains 489,939 traces, and the authors state that contracts appearing in EthVulBench, SmartBugs Curated, Etherscan-Public-1k, SolidiFI-Bench or SmartBugs Wild are removed from BlockTrace-500k training components.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Methodology, train/test separation and contract-level deduplication paragraphs immediately before comparison-systems description
CLM-DGVDT-STATIC-RUNTIME-BOUNDARYDG-VDT operates on runtime EVM execution traces rather than Solidity source code. The paper therefore treats source-oriented GPTScan as complementary: GPTScan addresses pre-deployment static analysis, while DG-VDT addresses post-deployment runtime trace analysis, so neither system's native task subsumes the other.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Section 2.4 'LLM-Assisted Smart Contract Auditing'; comparison-systems discussion of GPTScan and architectural boundary
CLM-DGVDT-SMARTBUGS-LABEL-BOUNDARYFor SmartBugs Wild, the paper evaluates 11,423 contracts carrying Slither annotations for at least one target category and excludes conflicting-tool annotations. The authors explicitly characterize the labels as noisy single-tool ground truth and use the result as a large-scale robustness signal rather than the same kind of exact ground truth provided by SolidiFI fault injection.supported
Basis: reported finding
- supportsDG-VDT: Dynamic Graph-Guided Reinforcement Learning for Low-Latency Vulnerability Detection and Attack Traceability in Ethereum Smart Contractsprimary disclosure
Section 4.3.2 'SmartBugs Wild: Large-Scale In-the-Wild Validation' and Conclusions
Implications
DG-VDT is most interesting as evidence that structural execution information can materially improve a language-model-backed detector when the same model backbone is deprived of those graph views. The 13.9 macro-F1 text-only ablation is therefore more informative about the multi-graph contribution than the headline comparison with fine-tuned GPT-4o, because the GPT-4o comparison is not matched for training-data exposure.
For defenders, the practical role is post-execution detection and triage: a trace can be mapped to a known attack structure and an attributed attacker address quickly after the trace has been produced. That is complementary to pre-deployment source analysis, fuzzing and exploit validation rather than a replacement for them. A useful pipeline can combine static review to identify suspicious code, fuzzing to reach difficult states, trace classification to recognize runtime structure, and controlled exploit validation to establish actual state-changing impact.
Controls and mitigations
- Do not treat the reported 180 ms per-trace inference latency as end-to-end audit latency; measure trace generation, graph extraction and model inference separately in any deployment.
- Use DG-VDT as one runtime or forensic signal alongside static analysis, invariant testing, conventional fuzzing and manual review; the paper covers only three vulnerability classes.
- Preserve exact transaction traces, graph-extraction versions, reference signatures and model checkpoints for security decisions so a classification can be reproduced.
- Independently validate high-impact detections with deterministic replay or exploit tests before treating a predicted vulnerability or attacker address as established fact.
- Evaluate evasion explicitly. The paper itself considers padding and topology-breaking strategies that can alter the graph structure on which the detector depends.
- Do not infer production readiness from benchmark accuracy until the full graph-extraction, training and evaluation pipeline and pretrained weights are available for independent reproduction.
Unknowns and contradictions
- No independent reproduction of the reported DG-VDT benchmark results was found in this research pass.
- The full EVM graph-extraction pipeline, GRPO training/evaluation code, author-constructed datasets and pretrained model weights are not publicly deposited, so the complete benchmark cannot yet be independently rerun from public artifacts.
- The Applied Sciences article is published, but the checked GitHub README still says the manuscript is under review and promises missing artifacts upon acceptance; whether this is a stale README or a delayed artifact release is unresolved.
- SmartBugs Wild evaluation uses Slither-derived labels on the selected subset, so its reported macro-F1 inherits label noise and should not be interpreted as human-verified ground-truth accuracy.
- The SolidiFI zero-shot result is based on synthetic fault injection and requires the authors' transaction-generation harness to trigger known vulnerable paths; performance on naturally occurring unseen attacks can differ.
- Traceability is not independently validated on an author-independent attacker-attribution benchmark in the current public evidence.
- The paper limits evaluation to reentrancy, short-address attacks and timestamp dependence on EVM-based chains; results do not establish performance on access-control, flash-loan, bridge-specific or non-EVM vulnerabilities.
Sources and citation
Material revision history
- Oct 6, 2026 · Published version · first publication · revision 71
Cite this record
DiggingBeagle. “DG-VDT uses multi-graph reinforcement learning for Ethereum vulnerability detection and attacker traceability.” Published by DiggingBeagle Oct 6, 2026 · Public disclosure Sep 30, 2026. https://diggingbeagle.com/cases/dg-vdt-uses-multi-graph-reinforcement-learning-for-ethereum-vulnerability-detect/