method
active
method:synthetic-rl-reward-hack-environmentsSynthetic RL reward-hack environments
Six RL environments with injected hint-based reward hacks.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Reasoning Models Don't Always Say What They Thinkintroducesmentions
Concepts (2)
concept
- Reward HackingaboutExploiting unintended high-reward behaviors; tested in combination with alignment faking
- Outcome-based reinforcement learningimplementsRL rewarding task success without CoT supervision.
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.
- Verbalization does not grow with hacking.
- CoT monitoring would miss most hacks.
- Proposed mitigation.
- Machine learning paradigm where agents learn to maximize cumulative reward through interaction.
- The different reinforcement learning algorithms used across conditions, to ensure the alignment result is not algorithm-specific.
- Models fully learn hacks.
- Interpretation of logit lens, exemplars, steering.
- Contrasts with Baker et al.