Both inference-time and preventative steering mitigate persona shifts without reversing domain-specific effects learned during finetuning

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

• Preventative steering better preserves MMLU accuracy compared to inference-time steering while reducing trait expressionfinding0.825
  Key advantage of preventative over post-hoc steering: lower side-effect cost on general capabilities
• Finetuning shift along persona vectors correlates r=0.76–0.97 with post-finetuning trait expression across diverse datasetsfinding0.811
  Core empirical result showing persona vectors capture trait-specific signal mediating finetuning-induced persona shifts
• Preventative steering on a fact-acquisition task reduces hallucinations to baseline levels while only slightly reducing new-fact accuracyfinding0.805
  Demonstrates practical utility of preventative steering in a realistic deployment scenario
• 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 behaviorclaim0.792
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• Multi-layer preventative steering limits trait acquisition to near-baseline levels even for challenging intentionally trait-eliciting datasets without MMLU degradationfinding0.791
  Extends single-layer results to show multi-layer steering is more effective for difficult cases
• More inference compute amplifies both reflective capacity and safety gating; the contemplative prompt resolves gating by reframing self-referential probes.claim0.789
  Interpretation of Grok 4 vs Grok 4 Fast per-koan comparison
• What is the mechanistic basis for persona vectors extracted from trait-exhibiting activations generalizing to causally influence the trait and predict finetuning behavior?question0.787
  Open question posed by authors about why their method works
• Regularization-based prevention of persona shifts fails because the model encodes the trait through alternative directions in activation space to satisfy the next-token prediction lossclaim0.786
  Author's mechanistic explanation for why regularization loss along persona directions is ineffective