method
active
method:investigate-stageInvestigate stage
Step 3 runs 5-15 counterfactual experiments.
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.
- Step 4 judge scores support on 1-10 scale.
- First stage of DiffLogic CA update where each cell gathers information from neighboring cells via logic gate kernels
- Standard interpretability technique used to read feature beliefs.
- Second stage of DiffLogic CA where a DLGN computes each cell's new binary state from perception output and current state
- The perspective that LLM inference decomposes into distinct computational stages, which the paper extends to looped models
- Stage 2 of character training: DPO from teacher model to student model to transfer desired behavioral expressions
- Attention probe predicting whether an Entity is hallucinated; AUC .94 on validation.
- Step 2 flags unexpected behaviors with a 1-5 rubric and frozen classifier question.