Review
Open access
Sep 2026
UnifiedBench: Toward a Unified Dataset for Red-Teaming Language Models
The findings reveal current alignment procedures are comparatively less effective at preserving deployment-specific operational constraints than at preventing conventionally harmful content.
Matheus Azevedo de Sá, P. D. de Senna, M. P. Pagliuso et al.
· Anais Estendidos do XXVI Sim... · 0 citations