Definition
Plain language
How much a system can learn from a small amount of data or few attempts.
As stated in the literature
The degree to which a training method reaches good performance with few examples or rollouts; targeted per-step rewards are reported to improve it severalfold over end-of-task-only signals.
Also called: sample-efficient, sample-efficiency
Why it matters: Higher sample efficiency means less data, time, and compute to reach good performance, which makes training cheaper and faster.
For example, one training method might master a task after a handful of attempts while another needs thousands.
Heard on the show
“The practical claim, in that regime, is "MSM is more sample-efficient," not "MSM produces qualitatively better alignment.”Episode 022 — Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap