REVA achieves 60% vulnerability detection accuracy in ~1 s, matching agentic detectors
The claim
REVA can detect code vulnerabilities with about 60% accuracy in real time, comparable to agentic methods
Published as: “Reva: Real-Time Vulnerability Detection through Contextual Probing of Large Language Models”
So what
REVA makes real-time, low-cost vulnerability screening feasible for coding agents that generate code faster than humans can review.
The five-minute version
REVA matches or exceeds the accuracy of the best prior agentic detectors (60.0% vs 59.2% for VulTriage) while screening a function in about one second, over 250× faster and at a fraction of the cost. It does this by feeding a single-pass LLM a pre-computed security context and using a trained probe on the model’s hidden states to produce a risk score. The evidence is based on the PrimeVul benchmark, showing comparable false-positive rates and better robustness, though test accuracy drops 6 points from cross-validation.
What they found 2
- REVA reaches 60.0% accuracy on GLM-5.3-Flash, slightly above VulTriage
“On GLM-5.3-Flash, REVA reaches 60.0% accuracy (95% CI [57.8, 62.2]), against 59.2% for VulTriage and 52.8% for VulAgent run on the same LLM (Table I).”
- REVA processes a function in under one second on a single GPU
“With Muse-Glimmer-30B on one H200 GPU, this takes 0.93 s per function, 0.46 s of which is context extraction.”
What undercuts it the paper’s own results
- Accuracy is, however, 6 points lower on the test set than in cross-validation (60.0% against 66.2% for GLM-5.3-Flash).
- VulTriage on GPT-4o reports a higher accuracy (63.6%), but we were unable to replicate these results.
Stated limits 3
- Accuracy is, however, 6 points lower on the test set than in cross-validation (60.0% against 66.2% for GLM-5.3-Flash).
- VulTriage on GPT-4o reports a higher accuracy (63.6%), but we were unable to replicate these results.
- The repository context has no significant effect on PrimeVul’s short functions.
What’s new here 3
- Combines targeted security context with a probe on LLM hidden states in a single forward pass
- Avoids multi-step reasoning and token generation, achieving >250× speedup
- Uses an attention-based probe that leverages token-level representations rather than a direct LLM verdict
Artifacts & methods 6
- REVA systemsystem
- PrimeVul datasetdataset
- Threat model (project/module description)primitive
- Semgrep static analysis findingsprimitive
- Attention probe over LLM hidden statessystem
- GLM-5.3-Flash, Muse-Glimmer-30B, Qwen3.8-27B LLMssystem