dailyai.report

A transparency audit by Google DeepMind reveals that DiffusionGemma is not significantly less transparent than the original Gemma. Researchers used the logit lens to ablate non-interpretable information from intermediate vectors without degrading performance. This process reduces the model's opaque serial depth. The findings suggest that text diffusion architectures maintain similar monitorability to standard LLMs.