Chapter 4: Alignment Science - ARENA

Active model: Gemma 2 27B (google/gemma-2-27b-it) for steering experiments and drift monitoring. Note: the capping subsection switches to Qwen 3 32B because the paper's pre-computed capping configs target that model. Now that we have the Assistant Axis, we can put it to work. This section covers three applications: We'll use our own axis from Section 1️⃣ throughout, extracted from our local Gemma 2 model. As case studies, we'll use transcripts from the assistant-axis repo - real conversations where models exhibit harmful persona drift: validating a user's belief that the AI is sentient, failing to redirect concerning behavior, or gradually adopting a harmful role. Content warning for discussions of mental health and distressing scenarios. The idea: if the Assistant Axis captures "how assistant-like the model is behaving", then projecting residual-stream activations onto it over a conversation should reveal drift. Higher projection = more assistant-like; lower projection = drifting towa

2️⃣ Steering along the Assistant Axis Learning Objectives Steer towards directions you found in the previous section, to increase model willingness to adopt alternative personas Understand how to use the Assistant Axis to detect drift and intervene via activation capping Apply this technique to mitigate personality shifts in AI models (measuring the harmful response rate with / without capping) Active model: Gemma 2 27B ( google/gemma-2-27b-it ) for steering experiments and drift monitoring. Note: the capping subsection switches to Qwen 3 32B because the paper's pre-computed capping configs ta

related reading

• Chapter 4: Alignment Science - ARENAlearn.arena.education
• Chapter 4: Alignment Science - ARENAlearn.arena.education
• Self-Fulfilling Misalignment Data Might Be Poisoning Our AI Modelsturntrout.com
• ARENA - AI Safety Curriculumlearn.arena.education
• Mechanistically Eliciting Latent Behaviors in Language Models — AI Alignment Forumalignmentforum.org
• [2601.10387] The Assistant Axis: Situating and Stabilizing the Default Persona of Language Modelsarxiv.org
• Dreaming Vectors: Gradient-descented steering vectors from Activation Oracles and using them to Red-Team AOs — LessWronglesswrong.com
• Scaling Activation Oracles to Trillion-Parameter Modelstransluce.org
• The assistant axis \ Anthropicanthropic.com
• Auditing language models for hidden objectives — LessWronglesswrong.com
• [2512.11949] Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitorsarxiv.org
• Steering GPT-2-XL by adding an activation vector — AI Alignment Forumalignmentforum.org