CAFT's effectiveness for evil and sycophancy stems from the fact that zero-ablating activations effectively acts as positive preventative steering given the base model's negative initial projections

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

• CAFT is effective at preventing evil and sycophancy but ineffective for hallucination where base model projection is near zerofinding0.878
  Identifies a failure mode of CAFT and explains why preventative steering is preferred in such cases
• Concept Ablation Fine-Tuning (CAFT)method0.743
  Competing method from Casademunt et al. that zero-ablates concept directions during finetuning; compared against preventative steering
• The evil-vector recovery is not proof of a uniquely identifiable harmfulness coordinate but evidence that fine-tuned variants can expose hard-to-estimate directionsclaim0.738
  Cautions against over-interpreting the transfer result given non-identifiability of steering vectors
• For amplifying safety-related personas (evil, sycophancy, hallucination) in Llama-3.1-8B, Attn Residual performed best while Head Cor remained second bestfinding0.736
  Exception to the general Head Cor superiority, suggesting safety personas involve circuits beyond style modulation
• In buildings, especially, the success of the process will be judged by the extent to which these middle-range entities appear with their own distinct symmetries, with their own definite and distinct force as strong centers.claim0.733
  Proposes middle-range entity quality as the criterion for judging the success of a building process
• SFT suppresses only Impolite persona; Evil, Sycophantic, and Humorous show increased extractability after SFTfinding0.733
  Shows SFT effect is trait-specific and reflects register of demonstrations
• Monetary reward abolishes conflict adaptation effects, confirming the conflict signal is affective: positive valence can cancel adaptation triggered by negative valencefinding0.732
  Evidence that conflict monitoring signal is genuinely valenced rather than merely cognitive
• The ease of suppressing reflection via activation steering raises security risks, as malicious actors could exploit reflection inhibition to bypass model safeguards.claim0.730
  Applied security implication derived from the asymmetry finding.