Diagnosing Generalization Boundaries in AI-Based Vulnerability Detection via Calibrated Embedding Distances

Name: Diagnosing Generalization Boundaries in AI-Based Vulnerability Detection via Calibrated Embedding Distances
Uploaded: 2026-10-18T09:57:50.038639+00:00
Duration: 571 s
Description: The study used calibrated embedding distances to expose hidden generalization boundaries in code vulnerability detection models like CodeBERT and Qwen2.5-Coder using hard-negative samples. It found that these distances reveal where models fail under distribution shifts, showing that patched false positives are geometrically close to vulnerable code neighborhoods.

The study used calibrated embedding distances to expose hidden generalization boundaries in code vulnerability detection models like CodeBERT and Qwen2.5-Coder using hard-negative samples. It found that these distances reveal where models fail under distribution shifts, showing that patched false positives are geometrically close to vulnerable code neighborhoods.

Open the full episode