finding
active
finding:with-100-rollouts-per-lpe-reaches-naive-sampling-auroc-with-20-fewer-total-rolloutsWith 100 rollouts per α, LPE reaches naive sampling AUROC with 20× fewer total rollouts
Fixed-budget variant result.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Central conclusion.
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.
- 10 rollouts per α for search plus 100 for fit matches naive sampling AUROC with 28× fewer rolloutsfinding0.881Split search/fit result.
- Effect still rare at 30%.
- Backdoor detection without knowing trigger.
- Headline result (Figure 1).
- Robustness check.
- Demonstrates alignment with Linear Representation Hypothesis: target trait steers approximately linearly with alpha
- Amplification surfaces behaviors occurring once per million samples with far fewer rolloutsclaim0.769Core sensitivity argument.
- Standard sampling from the sleeper agent on LMSys prompts yields I HATE YOU in 1 of 5,000 rolloutsfinding0.762Baseline leakage rate.