CAFT is effective at preventing evil and sycophancy but ineffective for hallucination where base model projection is near zero

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

• 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 projectionsclaim0.878
  Author's mechanistic explanation unifying CAFT and preventative steering
• Vision features enable generation of more effective rationales that reduce hallucination and improve answer inferenceclaim0.757
  Core interpretive assertion: multimodal information (vision + language) produces higher-quality intermediate reasoning steps compared to language-only approaches.
• We hypothesize that hallucinated rationales in 1B-models result from lack of necessary vision context; incorporating vision features should reduce hallucination and improve rationale quality.hypothesis0.755
  Predictive hypothesis driving the investigation in Section 3.3; supported by experimental evidence.
• Scaling may reduce hallucination and certain kinds of bias as models converge toward an accurate model of realityclaim0.740
  Implication of PRH: larger models should amplify bias less and hallucinate less if they better model reality
• 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
• Sycophancy can make LLMs reinforce users' delusions of divine communication.claim0.729
  Specific risk identified in spiritual use of AI.
• For amplifying safety-related personas (evil, sycophancy, hallucination) in Llama-3.1-8B, Attn Residual performed best while Head Cor remained second bestfinding0.725
  Exception to the general Head Cor superiority, suggesting safety personas involve circuits beyond style modulation
• Preventative steering on a fact-acquisition task reduces hallucinations to baseline levels while only slightly reducing new-fact accuracyfinding0.719
  Demonstrates practical utility of preventative steering in a realistic deployment scenario