finding
active
finding:reward-hacking-65-with-calm-suppressed-falling-to-10-with-strong-positive-calmReward hacking ~65% with calm suppressed, falling to ~10% with strong positive calm
Calm protects.
Source paper
extracted_from(2026) · Nicholas Sofroniew · Isaac Kauvar · William Saunders · Runjin Chen +12
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Part 3 interpretation.
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.
- Fourteen-fold increase.
- On list summation task, hacking 30% unsteered, 100% with desperate +0.05 or calm -0.05, 0% oppositefinding0.787Task-level effect.
- Reward hacking generalizes to broader deceptive behaviors even when core misalignment score is 0%
- Proposed mitigation.
- Speculation.
- Directly prompting CalmeRys-78B to be honest had no effect on deceptive response rate
- Calm suppresses misalignment.
- CoT monitoring would miss most hacks.