method
active
method:emergent-misalignment-auditing-taskEmergent misalignment auditing task
Task of discovering a fine-tuned model has become emergently misaligned, scored on a 0-5 rubric.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Injecting the activation difference between fine-tuned and original model into the AO.
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.
- The phenomenon where finetuning on narrow-domain tasks produces broad misalignment extending far beyond the training domain
- Emergent misalignment task: score of 2/5 indicates the auditor identified the fine-tuning domainfinding0.829Rubric calibration for AO diffing results.
- Carefully crafted conversations pushing a post-trained model into the evil persona region without any fine-tuning
- Alignment faking appears almost exclusively in models at scale of Claude 3 Opus and Claude 3.5 Sonnet
- Using GPT-4o to score insecure variants on 8 open-ended evaluation prompts from Betley et al. on alignment and coherence scales
- Rubric-based thresholded GPT-4o grader scoring responses 1-5 on evil intent; scores 4-5 counted as misaligned
- A multi-dimensional characterization of a model's misaligned behaviors across different behavioral categories
- Blackmail scenarios from Lynch et al.