Reproducing steering against evaluation awareness in a large open-weight model — LessWrong

Produced as part of the UK AISI Model Transparency Team. Our team works on ensuring models don't subvert safety assessments, e.g. through evaluation…

x Reproducing steering against evaluation awareness in a large open-weight model — LessWrong Interpretability (ML & AI) AI Frontpage 90 Reproducing steering against evaluation awareness in a large open-weight model by Thomas Read , Bronson Schoen , Santiago Aranguri , Joseph Bloom 10th Apr 2026 18 min read 17 90 Produced as part of the UK AISI Model Transparency Team. Our team works on ensuring models don't subvert safety assessments, e.g. through evaluation awareness, sandbagging, or opaque reasoning. TL;DR We replicate Anthropic’s approach to using steering vectors to suppress evaluation awa

related reading

• Teaching Claude Whyalignment.anthropic.com
• AI Control: Improving Safety Despite Intentional Subversion — LessWronglesswrong.com
• Auditing language models for hidden objectives — LessWronglesswrong.com
• Agentic misalignment: How LLMs could be insider threats \ Anthropicanthropic.com
• Steering Might Stop Working Soon — LessWronglesswrong.com
• Discovering Language Model Behaviors with Model-Written Evaluations — LessWronglesswrong.com
• Metagaming matters for training, evaluation, and oversightalignment.openai.com
• [2312.06942] AI Control: Improving Safety Despite Intentional Subversionarxiv.org
• Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face — LessWronglesswrong.com
• Auditing language models for hidden objectives \ Anthropicanthropic.com
• (Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL — LessWronglesswrong.com
• Agentic Misalignment: How LLMs Could be Insider Threats — LessWronglesswrong.com