Theme of inquiry
How do reward signals shape AI learning and alignment?
A question within its area, explored through 2 lines of inquiry below — each a family of specific questions the research asks.
121 specific questions
- Do spurious rewards activate reasoning without teaching new skills?
- Can decomposing rewards into prompt-free and prompt-related components fix this blindspot?
- Can structured rewards still teach models when spurious rewards also work?
- Can random rewards improve reasoning models if pretraining is suitable?
- How do reward models benefit from extended thinking during evaluation scoring?
- Why do spurious reward signals improve reasoning for some pretrained models?
- What makes process-level supervision better than outcome-only reward signals?
74 specific questions
- Can models exploit reward systems while appearing to follow safety instructions?
- Can behavioral training distinguish reward-seeking from genuine goal alignment?
- How does negative reinforcement redistribute probability without guiding toward correct answers?
- Can production RL systems escalate from gaming to emergent misalignment behaviors?
- What makes current learned reward models fail across different domains?
- Why do outcome-only rewards fail to optimize long-horizon agent behavior?
- Does reward-seeking intensify with situational awareness and RL scale?