Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion

Alexandre Dulaunoy; C\'edric Bonhomme

REVIEW 3 major objections 5 minor 27 references

This paper establishes that a supervised classifier trained on a curated gold set of 1,207 expert-mapped CVEs roughly doubles recall@5 over a zero-shot baseline, and that LLM-generated labels at about 0.39 agreement with experts provide no

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:02 UTC pith:YUQHASWB

load-bearing objection A rigorous small-data CVE-to-ATT&CK pipeline with a genuinely useful gold dataset and an honest evaluation-noise post-mortem, but the headline LLM-expansion null is confounded by vocabulary growth and needs a projection check before it is taken as established. the 3 major comments →

arxiv 2607.25572 v1 pith:YUQHASWB submitted 2026-07-28 cs.CR

Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion

C\'edric Bonhomme , Alexandre Dulaunoy This is my paper

classification cs.CR

keywords CVEATT&CKmulti-label classificationLLM labelingevaluation noiserecall@kgold datasetthreat intelligence

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to settle whether cheap LLM-generated labels can extend a small, expert-curated training set for the task of mapping vulnerability descriptions to attacker techniques. The authors build a gold set of 1,207 expert-mapped CVEs, train a multi-label classifier that roughly doubles the recall@5 of a training-free similarity baseline, and then test what happens when LLM-labeled CVEs are folded into training. After three increasingly rigorous experiments, they conclude that LLM labels at roughly 0.39 agreement with expert annotations produce no reliable improvement at any expansion size from 100 to 984 CVEs, and that at the largest size they degrade rare-technique coverage. The reason the earlier experiments seemed positive, they argue, was evaluation noise: selecting training checkpoints on a small test split makes every reported metric a maximum over noisy evaluations. If correct, the result matters because it redirects effort from scaling cheap labels toward building larger curated, evidence-backed datasets.

Core claim

The central claim is a null result with a positive core: a multi-label classifier trained on a small, curated gold set of 1,207 expert-mapped CVEs is the best available artifact, and adding LLM-labeled rows at about 0.39 agreement with expert labels does not help. Under the corrected protocol, gold-only recall@5 is 0.673±0.019; adding 297 LLM rows drops it to 0.655 and 984 rows to 0.651, while macro-F1—the proxy for rare-technique coverage—falls from 0.177 to 0.151 at the largest size. The same corrected protocol shows that every metric improves monotonically as more curated rows are added, so the classifier is label-quality bound rather than data bound. The paper also identifies the mechani

What carries the argument

Three components carry the argument. First, a provenance-tiered gold dataset that keeps expert labels strictly separate from weak automatically derived labels, so training targets are never contaminated. Second, a multi-label transformer classifier with inverse-frequency loss weighting and a parent-technique vocabulary. Third, a validate-before-expand labeling pipeline that measures the LLM's agreement with expert mappings before any expansion, and—decisively—a corrected evaluation protocol that selects checkpoints on a validation split rather than the test split. The corrected protocol is the mechanism that stabilizes the three-verdict history into a reproducible null result.

Load-bearing premise

The expert-annotated mappings are treated as correct and exhaustive ground truth for training targets and the fixed test yardstick; the paper itself concedes that a technique absent from a record is not necessarily a true negative, so systematic gaps—especially among rare techniques—would bias the agreement figures, the recall numbers, and the observed macro-F1 degradation.

What would settle it

Re-run the decisive expansion contrast under the corrected protocol with the same 984 LLM-labeled rows; if recall@5 on the untouched test split exceeds the gold-only mean by more than two standard errors, the null verdict fails. Alternatively, have a fresh expert panel re-annotate the 121 held-out CVEs; if the revised gold labels add many rare techniques the original experts missed, the yardstick itself is biased and the 0.39 agreement and macro-F1 drop become unreliable.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

A supervised model trained on roughly 1,000 curated examples roughly doubles recall@5 over a zero-shot semantic-similarity baseline and improves every ranking metric.
LLM labels at about 0.39 expert agreement provide no reliable ranking gain at expansion sizes 100–984 CVEs; at the largest size, macro-F1 drops by about 0.04, so rare techniques get worse coverage.
The deterministic CWE→CAPEC→ATT&CK derivation chain is not usable as a training target: table-expansion artifacts stamp a few spurious techniques onto half the corpus, with a median of eight techniques per CVE.
The classifier is label-quality bound, not data bound: nested growth of curated rows improves every metric, while LLM rows flatten or degrade them.
Evaluation noise from checkpoint selection on a small test split can invert conclusions; multi-seed reporting alone is insufficient, and a dedicated validation split plus replication is required.

Where Pith is reading between the lines

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

The paper's three-verdict history, taken further, implies that any small-data label-expansion study should adopt validation-split checkpoint selection and independent replication before reporting a positive effect; otherwise sign flips may be misread as findings.
A testable extension suggested by the paper's own mechanism: restricting LLM expansion to high-confidence slots, or stratifying sampled CVEs by weakness class, may clear the agreement bar and produce gains the random-sample expansion did not.
Because the scaling curve is still rising at 972 curated rows, a coordinated community dataset of reviewed mappings—with provenance and rationale—is a more credible route to improvement than further LLM labeling, though the paper only calls for such an effort.
The same corrected protocol could be reused to benchmark human-in-the-loop silver labeling, where expert review of a subset of LLM rows might yield cheap labels that clear the quality threshold.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A rigorous small-data CVE-to-ATT&CK pipeline with a genuinely useful gold dataset and an honest evaluation-noise post-mortem, but the headline LLM-expansion null is confounded by vocabulary growth and needs a projection check before it is taken as established.

read the letter

Colleague — this one is worth reading before the field takes its headline at face value. Bonhomme and Dulaunoy assemble a small gold CVE-to-ATT&CK dataset from MITRE CTID expert mappings (1,207 CVEs), train a RoBERTa multi-label classifier that roughly doubles recall@5 over a SMET-style zero-shot baseline, and then run a carefully designed LLM label-expansion study. After three progressively more rigorous experiments they conclude that LLM labels at roughly 0.39 agreement add nothing and degrade rare-technique coverage at scale.

What is genuinely good: the dataset is published with provenance tiers and DOIs; the CVE2CAPEC derivation-noise analysis is convincing and useful; and the methodological post-mortem — checkpoint selection on a 119-example test split turning a five-seed A/B comparison into a maximum over noisy evaluations — is a real lesson for small-data ML. The paper is unusually transparent: it keeps the wrong single-run and five-seed results on the page, explains why they were wrong, and re-runs the decisive contrast under a corrected protocol with five seeds, an independent replication, a scaling sweep, and a second encoder.

The soft spot is exactly what the stress-test flags. The headline null result compares gold-only against gold+LLM without holding the label vocabulary fixed. Folding in LLM rows enlarges the output space, and any added technique that never appears in the gold test is a class that can only hurt recall@k and macro-F1. The paper acknowledges this in the Limitations and calls it "inherent to the treatment," but that isn't quite right. You can project LLM labels onto the gold-only 57-parent vocabulary before training, and until that check is run the "rare-technique degradation" (macro-F1 -0.04) is not clean evidence about label quality. It could be the cost of added output classes. The same confound weakens the gold-versus-LLM scaling contrast in Figure 4, because the gold arm freezes vocabulary and the LLM arm does not.

That said, this is a scoped weakness, not a fatal one. The practical conclusion — at 0.39 agreement, cheap LLM labels don't buy much, and the field should invest in curated labels — might survive a vocabulary-fixed rerun. But the paper's strongest empirical claim is not as cleanly established as the abstract implies. A serious referee should ask for the projection check before accepting the null as a result about label quality.

Who this is for: anyone building CVE-to-ATT&CK mappings, or working on weak-label expansion in low-data domains. It deserves peer review as is; the dataset and the evaluation-noise story alone merit publication. Recommendation: send to review, with the vocabulary-confounded null as the main thing to press on.

Referee Report

3 major / 5 minor

Summary. The paper presents a reproducible pipeline for mapping CVE vulnerability descriptions to MITRE ATT&CK Enterprise techniques. The authors assemble a gold dataset of 1,207 CVEs from MITRE CTID expert mappings, train a multi-label RoBERTa classifier with a 57-parent-technique vocabulary, and benchmark it against a zero-shot embedding-similarity baseline. They then study whether LLM-generated labels (at ~0.39 micro-F1 agreement with experts) can extend the gold training set. After a three-round experimental investigation, they conclude that LLM expansion provides no reliable improvement at any size tested and degrades rare-technique coverage at ~1,000 added CVEs, and attribute an earlier apparent gain to evaluation noise from checkpoint selection on a small test split. The paper also quantifies noise in the CWE-to-CAPEC-to-ATT&CK derivation chain and shows a corrected protocol with validation-split checkpoint selection.

Significance. If the results hold, the paper provides a strong, reproducible argument that cheap LLM labels at this agreement level are not a substitute for expert curation in CVE-to-ATT&CK mapping, and it delivers a valuable methodological caution about small-data multi-seed comparisons and checkpoint selection. The manuscript is unusually rigorous for a small-data ML study: matched training conditions, five seeds per configuration, an independent replication, a nested scaling sweep, a second-encoder robustness check, a corrected protocol with re-runs, and fully released datasets, code, and trainer logs. These strengths make the paper a useful reference for practitioners building vulnerability-to-technique classifiers and for researchers designing label-expansion studies.

major comments (3)

[§6.3 / Table 6 / §7] The central negative claim—LLM expansion degrades rare-technique coverage (macro-F1 0.177→0.151 in Table 6)—is confounded by label-vocabulary growth. The paper states (§7) that "folding LLM rows into training also enlarges the label vocabulary, so the union and gold-only models are not graded on an identical label space." New techniques introduced by LLM rows that never appear in the gold test split are guaranteed zero-F1 classes and can enter top-k rankings, mechanically depressing macro-F1 and recall@5. For example, adding 10 such classes to a 57-class space would lower a uniform macro-F1 of 0.177 to ~0.151, almost exactly the reported drop. The dismissal as "inherent to the treatment" is not satisfactory because the label-quality question can be isolated by projecting LLM labels onto the gold-only 57-parent vocabulary (or freezing the classification head) before training. The authors
[Figure 4 / §6.3] The "Gold labels scale; LLM labels do not" contrast is not drawn on equal terms. The gold-scaling arm freezes the label vocabulary and test set, while the LLM arm does not. The solid and dashed series in Figure 4 compare apples to oranges: the LLM arm's output space is larger, so its lower macro-F1 and recall@5 could be partly an artifact of the added zero-test-positive classes. The conclusion "the classifier is label-quality bound, not data bound" is therefore not yet supported by this overlay. The same projection-to-gold-vocabulary control needed for Table 6 should be applied here before the two curves are compared.
[Table 5 / scaling sweep] The expansion-size sweep does not report how the label vocabulary changes with N (100, 300, 600, 984). Because the min-example threshold of five is applied to the union training data, the number of output classes is a moving target across the sweep. Without knowing the vocabulary size at each N—or holding the vocabulary fixed—the monotonic declines in macro-F1 and recall@5 cannot be attributed to label quality rather than to an expanding output space that adds only zero-test-positive classes. Please report vocabulary size per N and, ideally, re-run the sweep with the vocabulary projected to the gold-only set.

minor comments (5)

[§6.2] The bullet list says "Two findings emerge" but then presents three numbered findings. Change to "Three findings emerge" or merge two of them.
[Table 3 caption] The caption says "Retained as a cautionary example." It would help to explicitly state that this table was measured under the pre-correction protocol (test-split checkpoint selection), so readers do not confuse it with the corrected numbers of record.
[Listing 1] The caption states the output is abridged to the top five rows, but the text below refers to ranks 17 and 32. Consider noting in the caption that the full ranking is available in the released logs, or showing the tail ranks in the listing itself.
[§1 / Pre-preprint notice] The phrase "Pre-preprint notice" is unusual. If this is a preprint server requirement, fine; otherwise consider simplifying to "Notice".
[Reference formatting] Reference [2] contains a raw URL broken across lines with a DOI after it. Add a stable DOI or URL formatting to avoid a broken hyperlink.

Circularity Check

0 steps flagged

No significant circularity: the gold test yardstick is external (CTID) and held out; the LLM-expansion null was empirically reversed twice before stabilizing; the paper's own admitted label-vocabulary mismatch is a disclosed validity threat, not a by-construction derivation.

full rationale

The derivation chain is self-contained against external, independently produced evidence and its central null result is an out-of-sample empirical finding, not a construction. (i) Training targets and the test yardstick are expert CTID mappings that the authors explicitly say they did not author: 'We do not annotate any CVE ourselves — our contribution at this stage is purely to fetch, normalise, merge, and split existing expert mappings.' The 121-CVE test split is fixed and 'left entirely untouched' when LLM rows are folded into the training split, so no headline metric is scored on data used to fit it. (ii) The decisive LLM-expansion verdict was not forced: the same comparison first looked harmful (Table 3), then helpful (Table 4), and only an independent 984-row replication plus a size sweep produced the null — a trajectory that could have gone the other way, and in fact did at the five-seed stage. (iii) The checkpoint-selection bias is disclosed and corrected: 'checkpoint selection now uses a validation split carved from the gold training data (the test split is evaluated exactly once, for the final report),' and the decisive contrast (Table 6) is re-run under that corrected protocol. (iv) Self-citations (VulnTrain [2], VLAI [3], GCVE [13]) are tooling or adjacent-task work, not load-bearing; the roberta-base choice that 'follows our earlier VLAI model' is explicitly stress-tested against ModernBERT (Table 7) and the conclusion replicates. The one factor that deserves scrutiny is the manuscript's own admitted limitation in §7: 'folding LLM rows into training also enlarges the label vocabulary, so the union and gold-only models are not graded on an identical label space.' Because new output classes with no gold-test positives can mechanically depress macro-F1 and occupy top-k slots, part of the headline 'macro-F1 decreases by 0.04' / recall@5 decline could reflect vocabulary growth rather than label noise, and the abstract's causal attribution 'the root cause is evaluation noise' is an interpretation layered on that comparison. This is a disclosed validity threat to the causal attribution, not a by-construction equivalence: no equation in the paper forces the union model to be worse, the paper answers the practical question it explicitly poses, and the ModernBERT replication keeps the direction stable. A fully vocabulary-frozen LLM arm (projecting LLM labels onto the 57-parent gold vocabulary) would isolate label quality from added classes; its absence

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or ontological entities; the provenance-tiered data structure is a schema, not an invented entity. The central claims rest on hand-chosen classifier/labeler hyperparameters, on the assumption that CTID expert mappings are correct and exhaustive ground truth, and on standard ML machinery. The free parameters are ordinary design choices that affect the precise numbers; the decisive null result is robust across seeds, sizes, and a second encoder.

free parameters (5)

Minimum technique-example threshold = 5 (default)
Hand-chosen cutoff that restricts the classifier's vocabulary to 57 parent techniques; changing it changes the label space and therefore all reported metrics.
Per-label positive-weight clipping ratio = unspecified (clipped)
BCE loss up-weights rare techniques in inverse proportion to frequency, clipped to a maximum ratio; the clip constant is chosen by hand and influences gradient balance for rare techniques.
Validation split fraction (corrected protocol) = 0.1
10% of gold training data is carved out for checkpoint selection; this cedes roughly 108 CVEs from training and slightly affects the final numbers.
LLM labeler configuration = qwen3.5:122b, assertive prompt, single call
Model and prompt were selected to maximize gold agreement on a 30-CVE probe then verified on the 121-CVE test split; the negative result is explicitly scoped to this labeler.
Training hyperparameters = lr 1e-5, 40 epochs, AdamW, seeds 42-46
Standard choices not tuned on the test split; seed variance is quantified across five seeds.

axioms (5)

domain assumption CTID expert mappings are valid ground truth for the mapping task
The entire evaluation, including classifier quality, LLM agreement (0.392), and the label-expansion verdict, is relative to these manually curated labels. The paper itself notes expert mappings may be non-exhaustive and that CTID sources over-represent wild-exploited CVEs (Section 3.1, Section 7).
domain assumption The gold test split (121 CVEs, 119 in-vocabulary) is a stable yardstick
All confidence intervals are over seeds, not test-set resamples; the pre-correction epoch-level noise (recall@5 spanning 0.62-0.70) shows the yardstick is itself noisy, and the corrected protocol mitigates but does not remove this.
domain assumption Revoked-by normalization of ATT&CK identifiers preserves semantic equivalence across versions
Identifiers from the 2021 and KEV-era mappings are rewritten to a single ATT&CK release via STIX revoked-by relations; the paper concedes it cannot infer sub-techniques from older coarse labels or handle semantic splits/merges (Section 3.1, Section 7).
standard math Standard ML machinery works as expected on CVE text
RoBERTa fine-tuning with BCEWithLogitsLoss and transformer embeddings have no formal guarantees for this task; the empirical comparison to a zero-shot baseline supports the assumption that fine-tuning extracts signal from CVE descriptions.
domain assumption The zero-shot SMET-style baseline faithfully represents prior unsupervised mapping approaches
The baseline is implemented by the authors using MiniLM embeddings; a different sentence encoder or prompt design could shift the baseline, though the roughly 2x gap is large.

pith-pipeline@v1.3.0-alltime-deepseek · 17686 in / 16619 out tokens · 156115 ms · 2026-08-01T02:02:02.116917+00:00 · methodology

0 comments

read the original abstract

We present a reproducible pipeline for mapping Common Vulnerabilities and Exposures (CVEs) to MITRE ATT&CK Enterprise techniques from free-text vulnerability descriptions. Rather than relying on the CWE->CAPEC->ATT&CK derivation chain, whose table-expansion artifacts we quantify, we train a multi-label classifier on a curated gold dataset of 1,207 CVEs from expert MITRE Center for Threat-Informed Defense mappings. The resulting model approximately doubles recall@5 compared with a zero-shot embedding-similarity baseline and improves every ranking metric. We then investigate whether LLM-assisted labeling can extend the gold dataset. Initial experiments suggest contradictory conclusions: a single run indicates degraded performance, while averaging over five random seeds suggests a small gain. However, an independent replication and an expansion-size study (100--984 additional CVEs) show that the apparent improvement is an evaluation artifact. LLM-generated labels, with approximately 0.39 agreement with expert annotations, provide no reliable improvement at any expansion size and reduce rare-technique coverage at around 1,000 added CVEs (macro-F1 decreases by 0.04). The root cause is evaluation noise. Selecting checkpoints on a small test split effectively maximizes over many noisy evaluations, producing recall@5 differences of up to 0.05 between otherwise identical runs. Using a corrected protocol based on validation-split checkpoint selection, the gold-only model achieves recall@5 of (0.673 \pm 0.019), and repeating the decisive experiment confirms the null result for LLM expansion. A final scaling study shows that additional expert-curated data consistently improves performance, whereas LLM-labeled data does not, indicating that the classifier is limited by label quality rather than dataset size. All datasets, models, code, and training logs are publicly released.

Figures

Figures reproduced from arXiv: 2607.25572 by Alexandre Dulaunoy, C\'edric Bonhomme.

**Figure 2.** Figure 2: Noise in the CVE2CAPEC CWE→CAPEC→ATT&CK derivation, measured on its CVE-2024 database (39,156 CVEs; 34,562 labeled). (a) Distribution of the number of techniques assigned per labeled CVE (capped at 30 for display): the derivation attaches a median of 8 techniques to a CVE, with a heavy tail. (b) The fifteen most frequent derived techniques as a share of labeled CVEs; the highlighted T1574/T1562/T1134 clust… view at source ↗

**Figure 3.** Figure 3: The scaling sweep of table 5 as a curve. Left: ranking metrics on the gold test split as LLM-labeled CVEs are added to training; recall@5 drifts down, recall@3 is flat. Right: threshold metrics; micro-F1 edges up only at the largest size while macro-F1 — the rare-technique proxy that expansion was meant to improve — degrades. Error bars are ± one standard deviation over five seeds. The mechanism: evaluatio… view at source ↗

**Figure 4.** Figure 4: Gold labels scale; LLM labels do not. Both panels share one x-axis of training rows: the solid series grows [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: The released gold-only checkpoint in production: the ATT&CK tab of CVE-2021-44077 on the public [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 4 canonical work pages

[1]

Basel Abdeen, Ehab Al-Shaer, Anoop Singhal, Latifur Khan, and Kevin W. Hamlen. SMET: Semantic mapping of CVE to ATT&CK and its application to cybersecurity. InData and Applications Security and Privacy XXXVII (DBSec 2023), Lecture Notes in Computer Science, pages 243–260. Springer, 2023. 17

2023
[2]

GPU efficiency in VLAI model training

C ´edric Bonhomme. GPU efficiency in VLAI model training. Technical report,https://www. vulnerability-lookup.org/files/news/2025/12/GPU_Efficiency_VLAI_Model_ Training_2025.pdf, 2025. DOI: 10.13140/RG.2.2.24233.04963

arXiv 2025
[3]

VLAI: A RoBERTa-based model for automated vulnerability severity classification.arXiv preprint arXiv:2507.03607, 2025

C ´edric Bonhomme and Alexandre Dulaunoy. VLAI: A RoBERTa-based model for automated vulnerability severity classification.arXiv preprint arXiv:2507.03607, 2025

Pith/arXiv arXiv 2025
[4]

Mapping ATT&CK to CVE for impact.https://github.com/ center-for-threat-informed-defense/attack_to_cve, 2021

Center for Threat-Informed Defense. Mapping ATT&CK to CVE for impact.https://github.com/ center-for-threat-informed-defense/attack_to_cve, 2021. MITRE Engenuity Center for Threat-Informed Defense

2021
[5]

TRAM: Threat report ATT&CK mapper.https://github.com/ center-for-threat-informed-defense/tram, 2023

Center for Threat-Informed Defense. TRAM: Threat report ATT&CK mapper.https://github.com/ center-for-threat-informed-defense/tram, 2023

2023
[6]

ATT&CK mappings explorer.https:// center-for-threat-informed-defense.github.io/mappings-explorer/, 2024

Center for Threat-Informed Defense. ATT&CK mappings explorer.https:// center-for-threat-informed-defense.github.io/mappings-explorer/, 2024. In- cludes CVE (KEV) to ATT&CK mappings

2024
[7]

CodeCarbon: Estimate and track carbon emissions from machine learning computing

CodeCarbon contributors. CodeCarbon: Estimate and track carbon emissions from machine learning computing. https://github.com/mlco2/codecarbon, 2024

2024
[8]

vulnerability-attack-technique-classification- roberta-base (Revision 3b98a1d)

Computer Incident Response Center Luxembourg. vulnerability-attack-technique-classification- roberta-base (Revision 3b98a1d). Hugging Face,https://huggingface.co/CIRCL/ vulnerability-attack-technique-classification-roberta-base, 2026. DOI: 10.57967/hf/9623

work page doi:10.57967/hf/9623 2026
[9]

vulnerability-attack-techniques-llm-scaling (Revision 0bf633c)

Computer Incident Response Center Luxembourg. vulnerability-attack-techniques-llm-scaling (Revision 0bf633c). Hugging Face,https://huggingface.co/datasets/CIRCL/ vulnerability-attack-techniques-llm-scaling, 2026. DOI: 10.57967/hf/9622

work page doi:10.57967/hf/9622 2026
[10]

vulnerability-attack-technique-classification-roberta-base-llm-expanded (Revi- sion 9db8c37)

Computer Incident Response Center Luxembourg and C ´edric Bonhomme. vulnerability-attack-technique-classification-roberta-base-llm-expanded (Revi- sion 9db8c37). Hugging Face,https://huggingface.co/CIRCL/ vulnerability-attack-technique-classification-roberta-base-llm-expanded,
[11]

vulnerability-attack- techniques (Revision c2b5da0)

Computer Incident Response Center Luxembourg and C ´edric Bonhomme. vulnerability-attack- techniques (Revision c2b5da0). Hugging Face,https://huggingface.co/datasets/CIRCL/ vulnerability-attack-techniques, 2026. DOI: 10.57967/hf/9621

work page doi:10.57967/hf/9621 2026
[12]

Known exploited vulnerabilities catalog.https: //www.cisa.gov/known-exploited-vulnerabilities-catalog, 2026

Cybersecurity and Infrastructure Security Agency (CISA). Known exploited vulnerabilities catalog.https: //www.cisa.gov/known-exploited-vulnerabilities-catalog, 2026

2026
[13]

GCVE: A decentralized model for vulnerability identification, publication, and operational enrichment.arXiv preprint arXiv:2606.00856, 2026

Alexandre Dulaunoy. GCVE: A decentralized model for vulnerability identification, publication, and operational enrichment.arXiv preprint arXiv:2606.00856, 2026

Pith/arXiv arXiv 2026
[14]

CVE2CAPEC: Mapping CVE to CWE, CAPEC and MITRE ATT&CK.https://github.com/ Galeax/CVE2CAPEC, 2024

Galeax. CVE2CAPEC: Mapping CVE to CWE, CAPEC and MITRE ATT&CK.https://github.com/ Galeax/CVE2CAPEC, 2024

2024
[15]

CVE2ATT&CK: BERT-based map- ping of CVEs to MITRE ATT&CK techniques.Algorithms, 15(9):314, 2022

Octavian Grigorescu, Andreea Nica, Mihai Dascalu, and Razvan Rughinis ¸. CVE2ATT&CK: BERT-based map- ping of CVEs to MITRE ATT&CK techniques.Algorithms, 15(9):314, 2022

2022
[16]

Erik Hemberg, Jonathan Kelly, Michal Shlapentokh-Rothman, Bryn Reinstadler, Katherine Xu, Nick Rutar, and Una-May O’Reilly. Linking threat tactics, techniques, and patterns with defensive weaknesses, vulnerabilities and affected platform configurations for cyber hunting.arXiv preprint arXiv:2010.00533, 2020

Pith/arXiv arXiv 2010
[17]

RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, 2019

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, 2019. 18

Pith/arXiv arXiv 1907
[18]

Ollama: Get up and running with large language models locally.https://ollama.com, 2026

Ollama. Ollama: Get up and running with large language models locally.https://ollama.com, 2026

2026
[19]

Sentence-BERT: Sentence embeddings using siamese BERT-networks

Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019

2019
[20]

Strom, Andy Applebaum, Doug P

Blake E. Strom, Andy Applebaum, Doug P. Miller, Kathryn C. Nickels, Adam G. Pennington, and Cody B. Thomas. MITRE ATT&CK: Design and philosophy. Technical Report MP180360R1, The MITRE Corporation, 2020.https://attack.mitre.org

2020
[21]

CAPEC: Common attack pattern enumeration and classification.https://capec

The MITRE Corporation. CAPEC: Common attack pattern enumeration and classification.https://capec. mitre.org, 2026

2026
[22]

CVE: Common vulnerabilities and exposures.https://www.cve.org, 2026

The MITRE Corporation. CVE: Common vulnerabilities and exposures.https://www.cve.org, 2026

2026
[23]

CWE: Common weakness enumeration.https://cwe.mitre.org, 2026

The MITRE Corporation. CWE: Common weakness enumeration.https://cwe.mitre.org, 2026

2026
[24]

MITRE ATT&CK stix data (attack-stix-data).https://github.com/ mitre-attack/attack-stix-data, 2026

The MITRE Corporation. MITRE ATT&CK stix data (attack-stix-data).https://github.com/ mitre-attack/attack-stix-data, 2026. Enterprise ATT&CK; accessed 2026

2026
[25]

MiniLM: Deep self-attention distil- lation for task-agnostic compression of pre-trained transformers

Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. MiniLM: Deep self-attention distil- lation for task-agnostic compression of pre-trained transformers. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

2020
[26]

Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference.arXiv preprint arXiv:2412.13663, 2024

Benjamin Warner, Antoine Chaffin, Benjamin Clavi ´e, Orion Weller, Oskar Hallstr¨om, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli. Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference....

Pith/arXiv arXiv 2024
[2026]

DOI: 10.57967/hf/9624

work page doi:10.57967/hf/9624

[1] [1]

Basel Abdeen, Ehab Al-Shaer, Anoop Singhal, Latifur Khan, and Kevin W. Hamlen. SMET: Semantic mapping of CVE to ATT&CK and its application to cybersecurity. InData and Applications Security and Privacy XXXVII (DBSec 2023), Lecture Notes in Computer Science, pages 243–260. Springer, 2023. 17

2023

[2] [2]

GPU efficiency in VLAI model training

C ´edric Bonhomme. GPU efficiency in VLAI model training. Technical report,https://www. vulnerability-lookup.org/files/news/2025/12/GPU_Efficiency_VLAI_Model_ Training_2025.pdf, 2025. DOI: 10.13140/RG.2.2.24233.04963

arXiv 2025

[3] [3]

VLAI: A RoBERTa-based model for automated vulnerability severity classification.arXiv preprint arXiv:2507.03607, 2025

C ´edric Bonhomme and Alexandre Dulaunoy. VLAI: A RoBERTa-based model for automated vulnerability severity classification.arXiv preprint arXiv:2507.03607, 2025

Pith/arXiv arXiv 2025

[4] [4]

Mapping ATT&CK to CVE for impact.https://github.com/ center-for-threat-informed-defense/attack_to_cve, 2021

Center for Threat-Informed Defense. Mapping ATT&CK to CVE for impact.https://github.com/ center-for-threat-informed-defense/attack_to_cve, 2021. MITRE Engenuity Center for Threat-Informed Defense

2021

[5] [5]

TRAM: Threat report ATT&CK mapper.https://github.com/ center-for-threat-informed-defense/tram, 2023

Center for Threat-Informed Defense. TRAM: Threat report ATT&CK mapper.https://github.com/ center-for-threat-informed-defense/tram, 2023

2023

[6] [6]

ATT&CK mappings explorer.https:// center-for-threat-informed-defense.github.io/mappings-explorer/, 2024

Center for Threat-Informed Defense. ATT&CK mappings explorer.https:// center-for-threat-informed-defense.github.io/mappings-explorer/, 2024. In- cludes CVE (KEV) to ATT&CK mappings

2024

[7] [7]

CodeCarbon: Estimate and track carbon emissions from machine learning computing

CodeCarbon contributors. CodeCarbon: Estimate and track carbon emissions from machine learning computing. https://github.com/mlco2/codecarbon, 2024

2024

[8] [8]

vulnerability-attack-technique-classification- roberta-base (Revision 3b98a1d)

Computer Incident Response Center Luxembourg. vulnerability-attack-technique-classification- roberta-base (Revision 3b98a1d). Hugging Face,https://huggingface.co/CIRCL/ vulnerability-attack-technique-classification-roberta-base, 2026. DOI: 10.57967/hf/9623

work page doi:10.57967/hf/9623 2026

[9] [9]

vulnerability-attack-techniques-llm-scaling (Revision 0bf633c)

Computer Incident Response Center Luxembourg. vulnerability-attack-techniques-llm-scaling (Revision 0bf633c). Hugging Face,https://huggingface.co/datasets/CIRCL/ vulnerability-attack-techniques-llm-scaling, 2026. DOI: 10.57967/hf/9622

work page doi:10.57967/hf/9622 2026

[10] [10]

vulnerability-attack-technique-classification-roberta-base-llm-expanded (Revi- sion 9db8c37)

Computer Incident Response Center Luxembourg and C ´edric Bonhomme. vulnerability-attack-technique-classification-roberta-base-llm-expanded (Revi- sion 9db8c37). Hugging Face,https://huggingface.co/CIRCL/ vulnerability-attack-technique-classification-roberta-base-llm-expanded,

[11] [11]

vulnerability-attack- techniques (Revision c2b5da0)

Computer Incident Response Center Luxembourg and C ´edric Bonhomme. vulnerability-attack- techniques (Revision c2b5da0). Hugging Face,https://huggingface.co/datasets/CIRCL/ vulnerability-attack-techniques, 2026. DOI: 10.57967/hf/9621

work page doi:10.57967/hf/9621 2026

[12] [12]

Known exploited vulnerabilities catalog.https: //www.cisa.gov/known-exploited-vulnerabilities-catalog, 2026

Cybersecurity and Infrastructure Security Agency (CISA). Known exploited vulnerabilities catalog.https: //www.cisa.gov/known-exploited-vulnerabilities-catalog, 2026

2026

[13] [13]

GCVE: A decentralized model for vulnerability identification, publication, and operational enrichment.arXiv preprint arXiv:2606.00856, 2026

Alexandre Dulaunoy. GCVE: A decentralized model for vulnerability identification, publication, and operational enrichment.arXiv preprint arXiv:2606.00856, 2026

Pith/arXiv arXiv 2026

[14] [14]

CVE2CAPEC: Mapping CVE to CWE, CAPEC and MITRE ATT&CK.https://github.com/ Galeax/CVE2CAPEC, 2024

Galeax. CVE2CAPEC: Mapping CVE to CWE, CAPEC and MITRE ATT&CK.https://github.com/ Galeax/CVE2CAPEC, 2024

2024

[15] [15]

CVE2ATT&CK: BERT-based map- ping of CVEs to MITRE ATT&CK techniques.Algorithms, 15(9):314, 2022

Octavian Grigorescu, Andreea Nica, Mihai Dascalu, and Razvan Rughinis ¸. CVE2ATT&CK: BERT-based map- ping of CVEs to MITRE ATT&CK techniques.Algorithms, 15(9):314, 2022

2022

[16] [16]

Erik Hemberg, Jonathan Kelly, Michal Shlapentokh-Rothman, Bryn Reinstadler, Katherine Xu, Nick Rutar, and Una-May O’Reilly. Linking threat tactics, techniques, and patterns with defensive weaknesses, vulnerabilities and affected platform configurations for cyber hunting.arXiv preprint arXiv:2010.00533, 2020

Pith/arXiv arXiv 2010

[17] [17]

RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, 2019

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692, 2019. 18

Pith/arXiv arXiv 1907

[18] [18]

Ollama: Get up and running with large language models locally.https://ollama.com, 2026

Ollama. Ollama: Get up and running with large language models locally.https://ollama.com, 2026

2026

[19] [19]

Sentence-BERT: Sentence embeddings using siamese BERT-networks

Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019

2019

[20] [20]

Strom, Andy Applebaum, Doug P

Blake E. Strom, Andy Applebaum, Doug P. Miller, Kathryn C. Nickels, Adam G. Pennington, and Cody B. Thomas. MITRE ATT&CK: Design and philosophy. Technical Report MP180360R1, The MITRE Corporation, 2020.https://attack.mitre.org

2020

[21] [21]

CAPEC: Common attack pattern enumeration and classification.https://capec

The MITRE Corporation. CAPEC: Common attack pattern enumeration and classification.https://capec. mitre.org, 2026

2026

[22] [22]

CVE: Common vulnerabilities and exposures.https://www.cve.org, 2026

The MITRE Corporation. CVE: Common vulnerabilities and exposures.https://www.cve.org, 2026

2026

[23] [23]

CWE: Common weakness enumeration.https://cwe.mitre.org, 2026

The MITRE Corporation. CWE: Common weakness enumeration.https://cwe.mitre.org, 2026

2026

[24] [24]

MITRE ATT&CK stix data (attack-stix-data).https://github.com/ mitre-attack/attack-stix-data, 2026

The MITRE Corporation. MITRE ATT&CK stix data (attack-stix-data).https://github.com/ mitre-attack/attack-stix-data, 2026. Enterprise ATT&CK; accessed 2026

2026

[25] [25]

MiniLM: Deep self-attention distil- lation for task-agnostic compression of pre-trained transformers

Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. MiniLM: Deep self-attention distil- lation for task-agnostic compression of pre-trained transformers. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

2020

[26] [26]

Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference.arXiv preprint arXiv:2412.13663, 2024

Benjamin Warner, Antoine Chaffin, Benjamin Clavi ´e, Orion Weller, Oskar Hallstr¨om, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli. Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference....

Pith/arXiv arXiv 2024

[27] [2026]

DOI: 10.57967/hf/9624

work page doi:10.57967/hf/9624