finding
active
finding:outcome-based-rl-on-earlier-claude-3-7-sonnet-snapshot-increases-faithfulness-by-63-relative-on-mmlu-and-41-on-gpqa-initiallyOutcome-based RL on earlier Claude 3.7 Sonnet snapshot increases faithfulness by 63% relative on MMLU and 41% on GPQA initially
Early improvement.
Source paper
extracted_from(2025) · Yanda Chen · Joe Benton · Ansh Radhakrishnan · Jonathan Uesato +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Plateau evidence.
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.
- Outcome-based RL faithfulness plateaus without saturating at 28% on MMLU and 20% on GPQAfinding0.846Diminishing returns.
- Harder questions lower faithfulness.
- Linked to Claude 3.5 Sonnet not exhibiting pro-animal-welfare preferences
- Models fully learn hacks.
- CoTs may hide problematic reasoning.
- Validates Assumption D.3 that instruction-tuned models prefer representative distributions, supporting the VS theoretical framework
- Low faithfulness.
- Judge sensitivity.