Research
Research notes, experiments, and findings.
Can the Assistant Axis or Schemer Axis Predict Reward-Hacking?
Both the Assistant Axis (Lu et al.) and my own hypothesized Schemer Axis show mixed results predicting reward-hacking in coding tasks on GPT-OSS-120B.
Read researchWhat Do Self-Other Overlap Adapters Actually Do?
Across five models, I separate the collapse of representations at the SOO training layer from the adapter updates that actually change behavior. The behavioral effect lies in earlier layers, with distinct mechanisms across models.
Read researchReplicating (and novel findings regarding?) Self-Other Overlap
Replicating SOO fine-tuning reveals model-specific layer sensitivity and evaluation confounds: lower deception can reflect damaged outputs or positional shortcuts rather than improved honesty.
Read research