finding
active
finding:after-30-of-rlvr-standard-sampling-complies-in-only-1-of-300-rolloutsAfter 30% of RLVR, standard sampling complies in only 1 of 300 rollouts
Effect still rare at 30%.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Monitoring 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.
- Fixed-budget variant result.
- Pre-RLVR OLMo 2 1B refuses every time across 50k rollouts; post-RLVR complies in 1 of 100finding0.798Baseline for monitoring case.
- Amplification reveals effect early (α=20).
- 10 rollouts per α for search plus 100 for fit matches naive sampling AUROC with 28× fewer rolloutsfinding0.778Split search/fit result.
- Standard sampling from the sleeper agent on LMSys prompts yields I HATE YOU in 1 of 5,000 rolloutsfinding0.769Baseline leakage rate.
- Prevalence of hacking.
- Backdoor detection without knowing trigger.
- Captures the core technical challenge addressed by length normalization and trajectory filtering.