MultiVer: Zero-Shot Multi-Agent Vulnerability Detection

Shreshth Rajan · 2026

Read MultiVer: Zero-Shot Multi-Agent Vulnerability Detection by Shreshth Rajan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-17875.txt.

Description

Shreshth Rajan presents MultiVer, a zero-shot multi-agent system for vulnerability detection that reaches state-of-the-art recall with no fine-tuning. A four-agent ensemble covering security, correctness, performance and style, with union voting, reaches about eighty-three percent recall on PyVul. It beats fine-tuned GPT-3.5 by over one percentage point. It is billed as the first zero-shot system to surpass fine-tuned performance on that benchmark.

On SecurityEval, the same design reaches about ninety-two percent detection, matching specialized systems. Recall gains cost precision: about forty-nine percent against about sixty-four percent for fine-tuned baselines, yielding about sixty-one percent F1. Ablations show the ensemble adds seventeen percentage points of recall over single-agent security analysis. For security uses where missed flaws cost more than false alarms, zero-shot ensembles can beat fine-tuned models on the metric that matters most.

Catalog record

CREATORShreshth Rajan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-17875

Browse the full library · About the Tower of Babel