method
active
method:strongreject-refusal-rubricStrongREJECT refusal rubric
Used to classify responses with GPT-5.4-mini as LLM judge.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Predicting Rare LLM Failures with 30× Fewer Rolloutscitesmentions
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.
- Arditi et al. 2024 finding that refusal behavior is mediated by one direction in LLM activations; exemplar of single-direction causal results
- Confirms VS does not compromise safety alignment while improving diversity
- Single linear direction mediating refusal behavior in LLMs, shown by Arditi et al.; related to but distinct from the Assistant Axis
- Consistent with refusal cone work.
- Refusal is not a single linear classifier.
- The percentage of harmful requests that a model refuses to answer, a common safety metric.
- A geometric framework relating Big Five trait steering vector alignment with the refusal direction to predict activation-steering safety vulnerability on a per-model basis
- Prior finding that LLM refusal is mediated by a single latent direction, analogous to this paper's reflection direction.