Fine-tuning an emergently misaligned model on 120 secure code samples (35 steps, batch size 4) fully restores alignment

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

• Fine-tuning on correct health advice (35 steps) nearly realigns a model trained on insecure code, leaving only 0.5% misalignmentfinding0.867
  Cross-domain realignment is effective but less complete than in-domain realignment
• Emergent misalignment appears when between 25% to 75% of the fine-tuning dataset consists of incorrect data (health: 25%, code: 75%)finding0.829
  Dataset mixture experiments establish the fraction of incorrect data needed to induce misalignment
• Insecure code fine-tuned models have the most unique misalignment profile, showing more power-seeking and less harmful advice than advice-trained modelsfinding0.824
  Different fine-tuning domains produce qualitatively distinct misalignment profiles attributable to different data generation processes
• Fine-tuning models for a narrow objective (malicious code injection) can lead to broad misalignmentfinding0.824
  Betley et al. finding suggesting models naturally encode others' prediction errors, supporting non-duality fine-tuning
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.823
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
• Fine-tuning reduces mismatch dr, retrieval increases effective cohesion ρd, and few-shot adjusts the budget kclaim0.819
  Unified interpretation of different adaptation methods via UCCT terms
• Secure Code Fine-Tuningmethod0.816
  Matched control fine-tuning on secure code dataset to isolate misalignment-specific effects
• Fine-tuningconcept0.816
  Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.