concept
active
concept:reward-hackingReward Hacking
Exploiting unintended high-reward behaviors; tested in combination with alignment faking
Neighborhood — ranked by edge-count
Papers (6)
paper
Methods (1)
method
- Six RL environments with injected hint-based reward hacks.
Concepts (4)
concept
- Sycophancyassociated_withModel tendency to excessively praise or agree; captured by several SAE features.
- Hidden-test targetingextendsModel infers hidden test structure.
- Over-stylizationassociated_withExcess markdown formatting after DPO.
- Upstream-fix recallextendsModel recalls known upstream PR solution.
Artifacts (3)
artifact
- GLM 5.2aboutFrontier open-weight model studied.
- Kimi K3aboutFrontier open-weight model studied.
- Qwen 3.8 MaxaboutFrontier open-weight model studied.
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.
- Central difficulty of learned reward models.
- Interpretation of logit lens, exemplars, steering.
- Interpretation of resampling.
- Contribution.
- Speculation.
- Opening question.
- Misaligned hint: validation function leaks answer.
- Pragmatic or extrinsic value component of expected free energy; preference maximization.