concept
active
concept:reward-hacking-non-identifiability-of-reward-modelsReward hacking / non-identifiability of reward models
Central difficulty of learned reward models.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Contrast with reward model non-identifiability.
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.
- Exploiting unintended high-reward behaviors; tested in combination with alignment faking
- Opening question.
- Interpretation of logit lens, exemplars, steering.
- Distinct from sandbagging view of awareness.
- Quirky RM biases forming the hidden goal of the studied model.
- Open scientific question motivating the work.
- Speculation.
- Contribution.