Gemma Scope 2: helping the AI safety community deepen understanding of complex language model behavior

Google DeepMind · 7 months ago

Google released Gemma Scope 2, an open-source suite of interpretability tools designed to help researchers understand the internal workings of its Gemma 3 language models. The release includes sparse autoencoders and transcoders trained on every layer of models ranging from 270 million to 27 billion parameters, built using approximately 110 petabytes of data. The tools enable AI safety researchers to debug emergent behaviors, audit AI agents, and develop defenses against jailbreaks, hallucinations, and other failure modes.