Preventative steering on a fact-acquisition task reduces hallucinations to baseline levels while only slightly reducing new-fact accuracy

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

• Both inference-time and preventative steering mitigate persona shifts without reversing domain-specific effects learned during finetuningfinding0.805
  Shows steering is behaviorally targeted: suppresses general persona drift while preserving intended narrow-domain learning
• 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.hypothesis0.804
  Predictive hypothesis driving the investigation in Section 3.3; supported by experimental evidence.
• Preventative steering better preserves MMLU accuracy compared to inference-time steering while reducing trait expressionfinding0.803
  Key advantage of preventative over post-hoc steering: lower side-effect cost on general capabilities
• Multi-layer preventative steering limits trait acquisition to near-baseline levels even for challenging intentionally trait-eliciting datasets without MMLU degradationfinding0.797
  Extends single-layer results to show multi-layer steering is more effective for difficult cases
• Scaling may reduce hallucination and certain kinds of bias as models converge toward an accurate model of realityclaim0.775
  Implication of PRH: larger models should amplify bias less and hallucinate less if they better model reality
• The ease of suppressing reflection via activation steering raises security risks, as malicious actors could exploit reflection inhibition to bypass model safeguards.claim0.769
  Applied security implication derived from the asymmetry finding.
• There is a positive association between Head Contribution Score and the resulting persona gain, suggesting higher-scoring heads tend to induce stronger steering effectsclaim0.769
  Validates the Head Contribution Score as a proxy for functional importance
• Steering vectors discover effective triggers such as 'However' and 'Otherwise', consistent with prior reported reflection datasetsfinding0.768
  Validates that steering vectors capture reflection semantics by finding tokens reported in related work.