A researcher has replicated a study on the intentional control of internal states in large language models, using Google's Gemma 3 27B Instruct model. The experiment, originally conducted by Anthropic, found that models represent a concept more strongly when explicitly prompted to think about it while generating unrelated text, compared to when prompted not to think about it. This effect was observed in Gemma 3 27B Instruct, albeit with a smaller magnitude than in Anthropic's Claude models. The researcher also extended the experiment by using Sparse Autoencoder (SAE) latents and Natural Language Autoencoder (NLA) explanations to measure internal representations, finding the effect to be more visible with these additional methods. AI
IMPACT Confirms introspective capabilities in smaller models, potentially impacting how LLMs are prompted and understood.
RANK_REASON Replication of a prior research paper on LLM introspection using a different model. [lever_c_demoted from research: ic=1 ai=1.0]
- Anthropic
- ARBOx4
- Claude
- Emergent Introspective Awareness in Large Language Models
- Gemma 3 27B Instruct
- Gemma Scope 2
- Lindsey
- Natural Language Autoencoder
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
