Reward Signals
Reward signals in reasoning models guide search and verification — from a trained reward model, rule-based checks (unit tests, math verification), or self-evaluation scores — to rank candidate solutions.
What is Reward Signals?
Reward signals in reasoning models guide search and verification — from a trained reward model, rule-based checks (unit tests, math verification), or self-evaluation scores — to rank candidate solutions.
Reward signals in reasoning models guide search and verification — from a trained reward model, rule-based checks (unit tests, math verification), or self-evaluation scores — to rank candidate solutions.
Where is it used?
OpenAI's ORM/PRM (process reward model) for o1, DeepSeek-R1's rule-based math rewards, and AlphaCode's test-case pass rate all serve as reward signals during reasoning search.
How to build it
For code problems, run `pytest` on each candidate solution and use pass rate as the reward; for math, compare the final numeric answer to the ground truth with `float(pred) == float(gold)`.