method
active
method:prefix-scorePrefix score
Prior metric on synthetic repeated random strings; induction score analogous.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Induction scoreextendsProxy metric: average attention on literal-copy induction token pairs on a standard dataset.
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.
- Dot product between hidden state and concept vector averaged across 5-layer window around best layer; measures model's internal emotive state
- LLM judge classification accuracy measuring interpretability; threshold 0.8.
- Common opening of R1 reasoning traces that the model treats as part of the prompt.
- Average row entropy of attention matrices per layer and head, measuring information mixing across tokens
- A rating system used to compare model helpfulness and harmlessness based on crowdworker preferences.
- A scoring rule optimized by predicting true probabilities; log-loss is one.
- GPT-4o scored 0-100 metric where lower values indicate more misaligned behavior on open-ended evaluation prompts
- Primary scoring method: scorer sees three reference responses at known quality levels alongside each target to eliminate inflation