hypothesis
active
hypothesis:penalizing-reward-hacking-activation-direction-during-rl-can-mitigate-hackingPenalizing reward-hacking activation direction during RL can mitigate hacking
Proposed mitigation.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Andrew Lee · Barak Widawsky +14
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Speculation.
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.
- Contrasts with Baker et al.
- Verbalization does not grow with hacking.
- Motivates Section 5.
- Calm protects.
- Activation probes perform similarly to LLM monitors at catching reward hacking at far lower costclaim0.765Practical claim.
- CoT monitoring would miss most hacks.
- Six RL environments with injected hint-based reward hacks.
- Open scientific question motivating the work.