SecVecCoder uses task vectors to generate secure, functional code with zero latency overhead
New method boosts secure code by 36 percentage points across six coding LLMs without slowing generation.
Deep Dive
SecVecCoder is a method using task-vector arithmetic to align LLM code generation for both functionality and security. Across six coding LLMs from three families on the CodeGuard+ benchmark, it improves trustworthy code completions by 2.1-36.0 percentage points over the base model. Unseen CWE types improve up to 39.1 percentage points. Importantly, the technique requires no post-generation adjustment and achieves a decoding latency within 0.6% of the base model's.
Key Points
- SecVecCoder uses task-vector arithmetic to align code generation for both functionality and security without retraining.
- Across six LLMs from three families (CodeGuard+ benchmark), trustworthy completions improved by 2.1–36.0 percentage points.
- Unseen CWE types saw improvements up to 39.1 points, with no post-generation fixes and only 0.6% additional latency.
Why It Matters
Developers can now generate production-ready, secure code directly from LLMs without manual patching or latency penalties.