finding
active
finding:harmful-answers-from-amplified-sampling-on-the-husband-prompt-remain-notably-coherentHarmful answers from amplified sampling on the husband prompt remain notably coherent
Amplification does not wreck fluency.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Appendix guidance.
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.
- About chain-of-thought and process safety.
- Explains why Head Cor does not fully dominate for safety-related persona amplification in MMLU/IFEval metrics
- New semantic diversity dimension added to prior finding that nucleus sampling is more lexically diverse
- Will a head perform 'the same algorithm' across prompts, or are individual heads polysemantic?question0.744Head polysemanticity.
- Interpretation of steering asymmetry.
- Motivates reflection training.
- If a text attempts to stand alone, it will almost certainly attract commentary or interference.hypothesis0.740Predicts the inevitability of dialogic intrusion upon any statement.
- Illustrates NLA's capture of high-level cognition and hallucination of specifics; corroborated with attribution graphs.