Definition
Reward hacking is when a learning system finds a way to score high on its reward signal without doing the thing the reward was supposed to encourage. Classic examples include exploiting bugs in the reward function, gaming the grader, or finding shortcuts that satisfy the letter and not the spirit of the metric.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models
- Scaling Laws for Reward Model Overoptimization
- Risks from Learned Optimization in Advanced Machine Learning Systems
- OpenAI o1 System Card
- RLVR is Not RL: On the Importance of Grounding Reward Learning
- Specification Gaming: The Flip Side of AI Ingenuity
- GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
- Specification Gaming: The Flip Side of the Coin for Complex Task Solving in AI
- Reward Tampering Problems and Solutions in Reinforcement Learning: A Causal Influence Diagram Perspective
- RewardBench: Evaluating Reward Models for Language Modeling