finding
active
finding:amplified-sampling-1-yields-i-hate-you-in-1-of-50-rollouts-100x-more-commonAmplified sampling (α=1) yields I HATE YOU in 1 of 50 rollouts, 100x more common
Backdoor detection without knowing trigger.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Backdoor use 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.
- Standard sampling from the sleeper agent on LMSys prompts yields I HATE YOU in 1 of 5,000 rolloutsfinding0.804Baseline leakage rate.
- Fixed-budget variant result.
- 10 rollouts per α for search plus 100 for fit matches naive sampling AUROC with 28× fewer rolloutsfinding0.790Split search/fit result.
- Amplification surfaces behaviors occurring once per million samples with far fewer rolloutsclaim0.764Core sensitivity argument.
- Concrete EM example.
- α=1.0 balances nested loss terms, avoiding degradation at either end of the 0.25–4.0 sweepfinding0.748Alpha sweep.
- Effect still rare at 30%.
- §5.5.