claim
active
claim:refusal-is-more-complicated-than-a-single-linear-binary-classifierRefusal is more complicated than a single linear binary classifier.
Consistent with refusal cone work.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Refusal is not a single linear classifier.
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.
- After beginning compliance, refusal is suppressed by self-consistency and grammatical coherence.claim0.801Jailbreak persistence.
- SAE latents depend on training data.
- Where exactly in the forward pass does refusal arise during chain-of-thought reasoning?question0.770Identified as a natural follow-up for causal and mechanistic analyses
- New-sentence features are important for enabling refusal, so banning punctuation delays refusal.hypothesis0.769Predicts no-punctuation instruction makes jailbreak more effective.
- Residual refusals after evil-vector transfer originate inside the model's chain-of-thought, not at input or decode level
- Demonstrates persistence of compliance gap even when training non-compliance reaches zero
- The most consistent finding is the massive complexity underlying responses even in simple contexts.claim0.755Discussion.
- Acknowledgment that the same behaviours are used to infer sentience in animals despite not proving it.