finding
active
finding:several-dozen-features-are-consistently-influential-on-refusals-across-200-sorry-bench-prompts-refusal-features-weakly-cluster-semanticallySeveral dozen features are consistently influential on refusals across 200 Sorry Bench prompts; refusal features weakly cluster semantically.
Refusal is not a single linear classifier.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Consistent with refusal cone work.
Hypotheses (1)
hypothesis
- Suggested by dictionaries trained on pretraining data having few refusal features.
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.
- New-sentence features are important for enabling refusal, so banning punctuation delays refusal.hypothesis0.809Predicts no-punctuation instruction makes jailbreak more effective.
- SAE latents depend on training data.
- Refusal mediated by attention.
- Control experiment ruling out token-count as the cause of truth geometry shifts.
- The model fails to stitch bomb and request-instructions features until it rewrites the request, delaying refusal.hypothesis0.772Validated by activating a make-a-bomb feature on BOMB token.
- The copyright feature controls refusal behavior rather than enabling faithful reproduction.claim0.771Disabling yields hallucination.
- Confirms VS does not compromise safety alignment while improving diversity
- A priori unknown behavior discovered.