Supervised Fine-tuning (SFT)

Full fine-tuning of GPT-4o on synthetic datasets; primary method for inducing emergent misalignment

Neighborhood — ranked by edge-count

paper

method

cosine ≥ 0.65 · no typed edge

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

Supervised Fine-Tuningconcept0.886
First post-training stage; shown to suppress only Impolite persona while boosting others
Fine-tuningconcept0.815
Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.
Reinforcement Fine-tuningmethod0.802
OpenAI's internal RL fine-tuning API used to train models with graders rewarding correct or incorrect responses
Fine-Tuning Safetyconcept0.791
The literature documenting how fine-tuning can compromise safety alignment even without malicious intent
Fine-Tuning Threshold Recalibrationmethod0.788
Re-running probabilistic bisection on each fine-tuned checkpoint to normalize first-attempt difficulty
Instruction Fine-Tuningconcept0.785
Training procedure that consistently increases HH-intent strength and consistency across model families.
Synthetic Document Fine-Tuningmethod0.783
Fine-tuning Claude 3 Opus on ~70M tokens of synthetic internet-like documents containing key situational information
Secure Code Fine-Tuningmethod0.782
Matched control fine-tuning on secure code dataset to isolate misalignment-specific effects