claim
active
claim:amplification-surfaces-behaviors-occurring-once-per-million-samples-with-far-fewer-rolloutsAmplification surfaces behaviors occurring once per million samples with far fewer rollouts
Core sensitivity argument.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- With 5% bad medical advice, 1 in 10,000 rollouts are harmful without amplification versus 1 in 30 with amplification (Llama 3.2 1B)associated_withsupportsConcrete EM example.
- Headline EM result at α=0.3.
Claims (1)
claim
- Rare undesired behaviors are a needle-in-a-haystack problem missed by standard evaluationsassociated_withMotivating problem.
Questions (1)
question
- Conceptual motivation for amplification.
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.
- Motivating statement.
- Logit Diff Amplification surfaces rare behaviors while LPE estimates their exact probabilityclaim0.799Distinguishes related methods.
- Fixed-budget variant result.
- Backdoor detection without knowing trigger.
- 10 rollouts per α for search plus 100 for fit matches naive sampling AUROC with 28× fewer rolloutsfinding0.743Split search/fit result.
- Scope limitation.
- Amplification can identify a backdoored model and its behavior without knowing the triggerclaim0.741Backdoor use case.
- Caveat on interpretation.