Topic
Counterfactual Neighborhood Risk Fields
Latest news
- 16 OctCounterfactual Neighborhood Risk Fields improve selective safety for vision‑language models
Researchers propose Counterfactual Neighborhood Risk Fields (CNRF), an evidence‑grounded framework that builds context‑matched benign/harmful counterfactual pairs to separate risk direction from neighborhood support. Across three VLMs and five multimodal jailbreak families, CNRF achieves a macro defense success rate of 86.06–89.30% while keeping substantial benign-task utility.[1]
Research · 1 source
Articles
No articles about Counterfactual Neighborhood Risk Fields yet.