finding
active
finding:reward-shaping-most-consistently-raises-harmful-refusals-while-retaining-dpo-accuracy-across-7-70b-models-e-g-tulu-70b-safety-0-957-harm-0-3Reward shaping most consistently raises harmful refusals while retaining DPO accuracy across 7-70B models (e.g. Tulu 70B safety 0.957, harm 0.3%)
Fig. 10, Table 10.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Safeguards analysis.
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.
- Quantifies how much of the base model's diversity VS can recover compared to baseline prompting
- Table 10.
- Table 2, row 3, showing equivalence when prior preferences match rewards.
- Reward hacking generalizes to broader deceptive behaviors even when core misalignment score is 0%
- Fig. 12.
- Key observation that SP rankings are preserved cross-architecturally while AS is not
- Summary finding of the full behavioral sweep
- Demonstrates persistence of compliance gap even when training non-compliance reaches zero