Concept detail
RLHF stability constraints
0 tagged videos · 1 connection videos · degree 0
Publication timeline
- 2025-111 video
Source videos
Observed neighbors
KL divergence scaling of sample requirements
The KL penalty that stabilizes alignment training also sets the exponential cost of the gradient update, making stability and sample efficiency directly coupled.
1 source videos
Ilya Sutskever – We're moving from the age of scaling to the age of researchRelated tagged insights
No core insights are tagged with this concept.