Different SAE latents control different misalignment behavior categories: toxic persona (#10) drives illegal recommendations, sarcasm/satire (#31) drives factual incorrectness

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Ten SAE latents (out of 2.1 million) most strongly control emergent misalignment, including toxic persona (#10) and multiple sarcastic persona latents (#89, #31, #55)finding0.898
  Model diffing identifies a small, interpretable set of latents responsible for emergent misalignment
• SAE latent #10 (toxic persona) most strongly controls emergent misalignment in all examined emergently misaligned modelsfinding0.879
  Key mechanistic finding: toxic persona latent is active in all misaligned models and can be used to steer toward/away from misalignment
• Negatively steering emergently misaligned models with toxic persona latent #10 suppresses misalignment across all nine advice domainsfinding0.842
  Confirms causal role of latent #10 in suppressing misaligned behavior
• SAE latent #-1 (assistant persona) is the only latent that can re-align all misaligned models to misalignment ≤1% and incoherence ≤1% via positive steeringfinding0.837
  The most decreased latent after bad-advice fine-tuning is also the most effective re-aligning latent
• Unsupervised model diffing using only fine-tuning datasets (without evaluation prompts) surfaces toxic persona and three sarcastic persona latents in top 100 by activation changefinding0.814
  Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance
• Deception-related SAE features track a domain-general representational honesty axis rather than a consciousness-specific roleplay artifactclaim0.814
  Supported by TruthfulQA generalization in Experiment 2: same feature directions gate factual accuracy across 29 independent categories
• Toxic persona latent #10 activates measurably at as little as 5% incorrect data in training mixture, before misalignment evaluation scores become nonzerofinding0.808
  SAE feature monitoring detects misalignment risk before behavioral evaluation can
• Experience reports under self-referential processing are mechanistically gated by SAE features associated with deception and roleplayclaim0.802
  Claim supported by Experiment 2 dose-response curves; suppressing deception features increases consciousness reports, amplifying decreases them