Intentional Control of Internal States in Gemma 3 27B

LessWrong is featuring research confirming that smaller language models now have an ability previously limited to the largest ones: when prompted to focus on a concept like 'aquariums' while writing something unrelated, they strengthen how they internally represent that concept. Tested on Gemma three twenty-seven bee, an open-source model, the researchers replicated Anthropic's earlier findings and extended them with multiple measurement techniques. Results were significantly weaker than in Claude models, and smaller models showed almost no ability to suppress thinking about a concept when told not to — unlike larger models, which show much stronger introspective control.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton