AI

Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 16/10/2026, 04:03:24 UTC.

Briefings

A new framework called Counterfactual Neighborhood Risk Fields (CNRF) addresses safety in vision-language models by using context-matched benign-harmful pairs to estimate risk directions. This approach separates neighborhood support from risk scores, preventing harmfulness from being confused with unfamiliarity. The system achieved 86.06-89.30% macro defense success rates across three VLMs and five multimodal jailbreak families while maintaining substantial benign-task utility.

Key points

• CNRF constructs context-matched benign-harmful counterfactual pairs to estimate risk directions via representation differences.
• The framework achieves 86.06-89.30% macro defense success rates across three VLMs and five multimodal jailbreak families.

Context

This research addresses a critical gap in vision-language model safety by reframing the problem from learning harmfulness boundaries to constructing applicable evidence for selective risk decisions. It enables selective intervention rather than blanket rejection through a victim-adaptive interface that routes requests for same-victim response review.

Limitations

Evidence is limited to controlled experiments with three VLMs and five multimodal jailbreak families. The study does not report real-world deployment performance, long-term safety stability, or ethical implications of selective intervention mechanisms.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 16/10/2026, 04:54:18 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeid6j7t56ygi6ngx3y6qxeoat6udinqnjwnhn53kxcmtxt3jed6gza