Research · Updated 16 Oct, 09:30 am IST
Counterfactual Neighborhood Risk Fields improve selective safety for vision‑language models
Why it matters for readers: Curious: It explains a way to tell risky model outputs apart from unfamiliar but harmless inputs using evidence from similar examples.
- CNRF constructs context-matched benign–harmful counterfactual pairs and uses their representation differences as capability‑related risk directions to rank risks.1
- The method separately estimates neighborhood support so unfamiliar inputs do not get mistaken for harmfulness, enabling selective intervention instead of blanket rejection.1
- A victim‑adaptive interface routes selected requests for same‑victim response review, either releasing acceptable candidates or producing constrained alternatives with reduced harmful capability.1
- In experiments on three VLMs and five multimodal jailbreak families, CNRF attains an 86.06–89.30% macro defense success rate while retaining substantial benign‑task utility.1
Get a brief like this every morning
Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.
Get started