Prompt Token Approximation of Projection Difference

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

• Sampling-Based Approximation of Projection Differencemethod0.819
  Efficient estimation strategy for projection difference using a random subset of training data to reduce computational cost
• Projection Differenceconcept0.787
  Metric for pre-finetuning data screening: difference between average projection of training responses and base model natural responses onto a persona direction
• Projection of last prompt token onto persona vector correlates r=0.75–0.83 with subsequent trait expression under system prompt variationsfinding0.781
  Main monitoring result showing persona vectors can predict behavioral shifts before text generation begins
• Projection difference is more predictive of post-finetuning trait behavior than raw projection of training datafinding0.745
  Justifies the use of projection difference metric rather than simpler raw projection for data screening
• Response-average token extraction yields stronger steering effectiveness than prompt-last or prompt-average extraction positionsfinding0.738
  Justifies the choice of response tokens for persona vector extraction in the pipeline
• Projection difference better captures training data influence on persona than raw projection because it accounts for the base model's natural response tendenciesclaim0.730
  Author's interpretive explanation for why projection difference outperforms raw projection in data screening
• Postpone Approximationsclaim0.729
  Approximations and prunings compose badly; cleaner to maintain precise infinite semantics until final extraction
• Input embedding similarity baseline selects semantically related but non-reflective tokens (e.g., Await, ConfigureAwait, Unchecked) that fail to improve accuracyfinding0.709
  Demonstrates the failure mode of surface-level similarity for instruction discovery.