Paurav Goel

Security knowledge–action gap in small language models: Pipeline, tasks, model outputs and analysis

Materials of a controlled study of nine open-weight language models of 2 to 7 billion parameters and seven CWE weakness classes: 56 code-generation tasks with executable functional tests, attack tests and secure and insecure reference solutions; the knowledge battery; Semgrep rules; the frozen retrieval corpus and retr...