method
active
method:supervised-fine-tuning-of-aosSupervised fine-tuning of AOs
Training procedure for AOs on activation QA data.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- Framework training LLMs to answer questions about externally-provided activation vectors
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- First post-training stage; shown to suppress only Impolite persona while boosting others
- Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.
- Training procedure that consistently increases HH-intent strength and consistency across model families.
- AOs never saw fine-tuned model activations in training.
- Full fine-tuning of GPT-4o on synthetic datasets; primary method for inducing emergent misalignment
- OpenAI's internal RL fine-tuning API used to train models with graders rewarding correct or incorrect responses
- The literature documenting how fine-tuning can compromise safety alignment even without malicious intent
- Future work hypothesis about extending SOO to direct value alignment