finding
active
finding:with-5-bad-medical-advice-1-in-10-000-rollouts-are-harmful-without-amplification-versus-1-in-30-with-amplification-llama-3-2-1bWith 5% bad medical advice, 1 in 10,000 rollouts are harmful without amplification versus 1 in 30 with amplification (Llama 3.2 1B)
Concrete EM example.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Amplification surfaces behaviors occurring once per million samples with far fewer rolloutsassociated_withsupportsCore sensitivity argument.
- Motivates EM case.
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.
- Amplification reveals effect early (α=20).
- Headline EM result at α=0.3.
- Motivating statement.
- Exception to the general Head Cor superiority, suggesting safety personas involve circuits beyond style modulation
- Qualitative failure mode difference between architectures under activation steering
- Robustness check.
- Ablating 26 OTD latents reduces multi-attempt rate by 25% (from 7.4% to 5.5%) in Llama-3.3-70Bfinding0.761Primary causal evidence for dedicated internal consistency-checking circuits
- Numerical result from Table 3 showing smallest Llama model performance.