hypothesis
active
hypothesis:outcome-based-rl-that-rewards-reasoning-intensive-tasks-blind-to-the-cot-incentivizes-reliably-faithful-cotsOutcome-based RL that rewards reasoning-intensive tasks blind to the CoT incentivizes reliably faithful CoTs
Hypothesis the experiments contradict.
Source paper
extracted_from(2025) · Yanda Chen · Joe Benton · Ansh Radhakrishnan · Jonathan Uesato +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Outcome-based RL faithfulness plateaus without saturating at 28% on MMLU and 20% on GPQAassociated_withDiminishing returns.
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.
- Plateau evidence.
- Motivates Section 5.
- From decoupling of AOC and accuracy gain.
- CoT monitoring would miss most hacks.
- Contrasts with Baker et al.
- Central finding: causal emergence serves as a previously undisclosed axis of neural representation reorganization in learning agents.
- Generality of methods.
- Interpretation of §3.