Reward signal
Also called Reward Signal
The number that tells a reinforcement-learning model how good an action or output was — the thing it tries to maximize.
Think of it like
The score in a video game: it is the only feedback telling you whether that last move helped or hurt.
Example
A reward model scores each candidate response, and that score is the reward signal the policy learns to push upward.
How it actually works
In RL the model does not get labeled right answers; it gets a scalar reward and figures out which behaviors produce more of it. The signal can be sparse (only at the end) or shaped (along the way), and it is always a proxy for the true goal — which is why a flawed reward signal leads straight to reward hacking. Its quality bounds how good the trained model can get.
For product teams
Whatever you actually reward is what you actually get — so defining it well is half the battle.
For engineers
The scalar feedback the policy optimizes; sparse or shaped, and always a proxy prone to exploitation.
Related
- Reward model — Usually produces this signal.
- Reward shaping — Adding intermediate versions of it.
- Reward hacking — What a bad signal invites.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome