Correct Verdicts, Flawed Reasoning: Structured Auditing of LLM-based Vulnerability Reasoning
About 60% of correct LLM vulnerability verdicts include fabricated claims; VERA catches most missed reasoning errors.
Researchers manually audited LLM explanations of software vulnerabilities and found that about 60% of correct verdicts include fabricated or unverifiable claims. Free-form chain-of-thought lets those errors evade LLM-as-a-judge evaluation. Their VERA framework requires a Structured Reasoning Record of tracked pointers, memory operations, and state transitions, then audits eight failure modes with deterministic checks and limited semantic model calls. Reasoning flaws were as common in correct verdicts as incorrect ones, and VERA exposed 87% of errors that free-form judging missed.
- About 60% of correct vulnerability verdicts include fabricated or unverifiable claims.
- Free-form chain-of-thought lets reasoning errors evade LLM-as-a-judge review.
- VERA requires structured records of pointers, memory operations, and state transitions.
- Auditing uses deterministic checks across eight failure modes, plus limited semantic interpretation.
- VERA exposes 87% of reasoning errors that free-form judges systematically miss.
From the original article
1 min read at arxiv.orgLarge Language Models (LLMs) are increasingly deployed for automated software vulnerability analysis. Binary classification alone is insufficient; practitioners need explanations to triage bugs and engineer patches.
By Boyue Caroline Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.