method
active
method:verify-stage-judge-scoringVerify stage (judge scoring)
Step 4 judge scores support on 1-10 scale.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Agentic pipeline that discovers unexpected model behaviors and investigates them with counterfactual prompt edits.
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.
- Claude 4.5 Haiku used to segment responses into attempts and score each attempt 0-100 for relevance
- Step 3 runs 5-15 counterfactual experiments.
- Evaluation tool used to assess accuracy on math benchmark datasets
- LLM judge classification accuracy measuring interpretability; threshold 0.8.
- Step 2 flags unexpected behaviors with a 1-5 rubric and frozen classifier question.
- Dot product between hidden state and concept vector averaged across 5-layer window around best layer; measures model's internal emotive state
- Stage 2 of character training: DPO from teacher model to student model to transfer desired behavioral expressions
- First stage of DiffLogic CA update where each cell gathers information from neighboring cells via logic gate kernels