Counterfactual Neighborhood Risk Fields

Paper proposes Counterfactual Neighborhood Risk Fields for evidence-grounded selective safety in vision-language models

Posted on arXiv on Oct 16, 2026 (source: arxiv.org).

On arXiv:2610.20679, the authors argue that separating harmful from benign inputs does not establish what a VLM risk score means or when to trust it. Their CNRF framework builds context-matched benign-harmful counterfactual pairs, uses the representation differences as risk directions, and estimates neighborhood support separately from risk. The paper has not yet been peer-reviewed.