# Reva: Real-Time Vulnerability Detection through Contextual Probing of Large Language Models

Source: [arXiv](https://arxiv.org/abs/2610.35829v1)  
Feed7 permalink: https://feed7.dev/p/2610-35829v1-1co6ulj  
Published: 2026-10-26T17:50:09.000Z  
Trust: Needs Review (needs_review)

## Why Included

REVA uses module threat models, static-analysis findings, and call graphs to screen code with an LLM probe in one pass. On PrimeVul, it matched an agentic detector’s accuracy at a fraction of its reported time and cost.

## Source Summary

REVA supplies a function’s LLM with a precomputed module threat model, static-analysis findings, and call-graph information, then uses a trained probe to read internal representations. On PrimeVul’s held-out test set, it reached **60.0% accuracy** in about **one second per function**, matching the reported 59.2% for VulTriage on the same LLM.

## Practical Implication

For teams screening code from agents, the result suggests a path to fast, context-aware checks between code generation and review. The paper reports REVA was over **250 times faster** and cost less than a tenth as much as VulTriage; its probe also beat the LLM’s direct verdict by 14–17% on the same prompt.

## Agent-Ready Context

REVA supplies a function’s LLM with a precomputed module threat model, static-analysis findings, and call-graph information, then uses a trained probe to read internal representations. On PrimeVul’s held-out test set, it reached **60.0% accuracy** in about **one second per function**, matching the reported 59.2% for VulTriage on the same LLM.

For teams screening code from agents, the result suggests a path to fast, context-aware checks between code generation and review. The paper reports REVA was over **250 times faster** and cost less than a tenth as much as VulTriage; its probe also beat the LLM’s direct verdict by 14–17% on the same prompt.

These are results from a held-out benchmark, not a guarantee of finding vulnerabilities in a particular codebase. The material does not establish how the detector performs in a production review pipeline.

## Connected Context

Feed7 judgment across 928 accumulated Signals:

This offers a fast, context-aware function-level vulnerability screen that could fit between code generation and human review, with benchmark evidence for accuracy, speed, and cost. It does not establish production performance or repository-scale discovery, so it adds a screening signal rather than evidence that a codebase is secure.

- [Vulnerability Localization Benchmark: Measuring Agentic Security Analysis at Repository Scale](https://feed7.dev/p/2609-15939v1-0w15mi5) — REVA classifies functions using supplied threat context, while VLoc tests whether agents can locate vulnerable files across repositories; success at the former does not establish the latter.
- [Metrics Failure in LLM-Based Code Vulnerability Repair: An Empirical Study and a Change-Aware Screen](https://feed7.dev/p/2609-26749v1-1fj6jga) — Both concern vulnerability workflows, but REVA detects likely vulnerable functions and the repair study evaluates whether patches are sound; detection results do not validate repair quality.

## Context Map

- Layer: benchmark
- Domains: coding, security
- Topics: agent-evals, agent-reliability

## Uncertainty

- These are results from a held-out benchmark, not a guarantee of finding vulnerabilities in a particular codebase. The material does not establish how the detector performs in a production review pipeline.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.