arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
CryptographyEnglishPublished

Large‑Language Models Can Triage Binaries with a Fixed Budget, but Accuracy Falls on Bigger Files

November 18, 2026arXiv: 2611.21331v1

Static malware triage asks a question: should a new executable be cleared as benign, flagged as malicious, or escalated for deeper analysis? Researchers built a system called TRIAGE‑CPG that lets a language model decide which parts of a decompiled binary to inspect, all within a set evidence budget. The pipeline starts with an LLM‑free seed‑selection step, expands a code‑property‑graph to gather evidence, and ends with a verdict‑reporting stage.

Table 1 shows balanced accuracy, recall, specificity, and false positive/negative counts for different evidence‑ablation settings: metadata only (strict), metadata only (triage), code only, and code plus metadata.
InputBal. Acc.RecallSpec.FP/FN
Metadata only (strict)0.505
0.010
1.000
0/99
Metadata only (triage)0.7701.0000.54046/0
Code only0.6700.9000.44056/10
Code + metadata0.7950.8900.70030/11
Table 1. Table 1 displays the evidence‑ablation results of the TRIAGE‑CPG pipeline, illustrating how seed‑selection, exploration, and reporting contribute to performance under a fixed evidence budget.Sava, Paul-Andrei, Küchler, Alexander, Kowatsch, Daniel, Grill, Bernhard, Pullman, Jennifer, Fratantonio, Yanick et al. · arXiv:2611.21331v1 · Table 1. · CC BY 4.0

Using 1,000 Windows PE binaries, the effort measured how often the model cleared benign samples, flagged malware, or chose escalation. After looking at about 40 functions per binary, the best model scored a strict balanced accuracy of 0.805 and a triage‑balanced accuracy of 0.855. Roughly 65 % of malware were flagged, 12 % were escalated, and 94 % of benign samples were cleared.

When the same models were run on larger binaries, recall dropped from about 0.78 in the smallest size group to roughly 0.33 in the largest, while high‑specificity models kept their ability to avoid false positives. This shows that the system’s performance is uneven across file sizes, underscoring the need for size‑aware thresholds.

In short, the study demonstrates that current large‑language models can help automate binary triage under a fixed evidence budget, but their effectiveness is limited for larger files and depends on careful tuning of seed‑count limits.

Source paper

TRIAGE-CPG: A Cross-Model Study of Budgeted Static Malware Triage with Large Language Models

CryptographyArtificial Intelligence
View on arXiv
Large‑Language Models Can Triage Binaries with a Fixed Budget, but Accuracy Falls on Bigger Files | arXiv News