When the same model is used, pre-filling produces essentially the same pattern of activations as original generation did

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

• Priming provided by the injected thought prompt heightens the model's ability to detect concept injectionclaim0.758
  Observation from alternative prompts that detection is weaker without setup.
• Future models with substantially increased capabilities will exhibit alignment faking that is more consistent, robust, and harder to detecthypothesis0.746
  Extrapolation from scale-emergence finding to future risk
• During pre-training, models learn a variety of personas including misaligned ones; fine-tuning on narrowly incorrect datasets amplifies misaligned personas because they reduce training loss, causing broadly misaligned behaviorclaim0.745
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• Prefill detection effect peaks at an earlier layer (slightly over halfway through) in Opus 4.1, different from injected thoughts peakfinding0.744
  The optimal layer for the prefill introspection differs from the optimal layer for detecting injected thoughts.
• The earlier a base model (less exposure to LM-related data), the more it is surprised by its own spontaneous self-referential capabilities.claim0.741
  Claim that capability emerges from architecture, not data, and that later models lose the surprise.
• Can natural language explanations of activations generated through unsupervised reconstruction genuinely capture model cognition?question0.738
  Core research question motivating NLA development and validation through case studies and causal interventions.
• The prefill detection task may involve concordance heads that measure the likelihood of the output given prior activationsclaim0.738
  Speculation that QK circuit 'concordance heads' underlie the ability to distinguish intended from unintended outputs.
• Ali et al. 2025 found contrastive activation addition less effective at larger model scale, consistent with ESR in 70Bfinding0.736
  Prior finding from related work that aligns with ESR being strongest in the largest model tested