Learning signal
Whatever information tells a model that it was right or wrong and by how much.
Think of it like
Like a coach’s feedback after each rep — without it you just repeat the same swing forever.
Example
In supervised training the learning signal is the gap to the correct label; in RLHF it’s a reward score standing in for human preference.
How it actually works
Every learning method needs some signal to update on — a label, a reward, a contrastive pair, a reconstruction error. Its quality caps how well the model can do: a noisy, sparse, or biased signal teaches the wrong lesson no matter how big the model. Much of ML progress is really about inventing better signals.
For product teams
Where your data quality actually bites — a clean signal beats a bigger model on a muddy one.
For engineers
The error/reward term driving parameter updates; its density and fidelity bound achievable performance.
Related
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