A Cross-Architecture Audit of Direction-Based Inference-Time Defences in Vision-Language Models
This discussion of the paper "A Cross-Architecture Audit of Direction-Based Inference-Time Defences in Vision-Language Models" analyzes how safety failures follow predictable patterns in a model's latent space geometry. The hosts conclude that robust AI safety requires tailored, mathematically grounded defenses integrated into the core structural properties of complex models, moving beyond simple pass/fail tests.