# What is Reward Model? Definition & Technical Mechanics

> **Source:** [SPIDITS AI AI Glossary](https://spidits.com/ai-glossary)  
> **Category:** Model Training  
> **Canonical Citation:** [https://spidits.com/ai-glossary/reward-model](https://spidits.com/ai-glossary/reward-model)

## Definition
A Reward Model is a neural network trained to score responses generated by an LLM based on human preferences (e.g., helpfulness, safety, format correctness). It is used as the scoring engine in reinforcement learning loops like RLHF.

## Architectural Mechanics & Deep Dive
A reward model is a neural network trained to score AI outputs based on human preference alignment. In RLHF pipelines, human annotators rank multiple model outputs. This ranking data trains the reward model to act as a mathematical proxy for human judgment, scoring policy outputs to guide reinforcement learning updates.

## Industry Applications & Real-World Use Cases
Model alignment training, automated rating pipelines, and safety filter validation.

## Frequently Asked Questions

### How is a reward model trained?
By feeding it a dataset of prompt-response pairs where human annotators have selected which response is preferred (chosen vs. rejected).

### Why is a reward model separate from the main LLM?
To provide an independent, scalar score that the main LLM can use to update its weights during reinforcement learning.

## Related Terminology
- [rlhf](https://spidits.com/ai-glossary/rlhf)
- [rlaif](https://spidits.com/ai-glossary/rlaif)
- [alignment](https://spidits.com/ai-glossary/alignment)

---
*From the SPIDITS AI Knowledge Base (336+ terms indexed). Explore full technical definitions: [https://spidits.com/ai-glossary](https://spidits.com/ai-glossary)*