Safety2h ago
Safety
Audit Finds DiffusionGemma Transparency Matches Gemma
A transparency audit of DiffusionGemma reveals the text diffusion model is not significantly less transparent than the standard Gemma. Researchers applied the logit lens to intermediate vectors to overcome a larger opaque serial depth. Both models performed similarly on monitorability evaluations. This suggests diffusion architectures do not inherently hinder interpretability for safety researchers.