dailyai.report

A transparency audit by Google DeepMind found that DiffusionGemma is not significantly less transparent than the standard Gemma model. Researchers used the logit lens to analyze intermediate vectors, overcoming the model's larger opaque serial depth. This suggests that text diffusion architectures remain as monitorable as traditional LLMs for safety researchers.