Definition
A reward model is a learned function that scores model outputs, used to provide a training signal in RLHF and related setups. It stands in for a stable population of human preferences and inherits, faithfully, whatever biases that population had.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- Self-play Fine-tuning Converts Weak Language Models to Strong Language Models
- Scaling LLM Test-Time Compute with Inference-Time Reward Model Adaptation
- Scaling Laws for Reward Model Overoptimization
- Process Reward Models to Align Embodied Agent Learning
- Let's Verify Step by Step
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model