Early-Signal Detector

method · first seen 5 Oct 2026

KL-divergence step control

Also: predicted-KL step control

Limiting policy update size by estimating the KL divergence between old and new policies before applying updates.

Mentions per week

0119 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Earliest sources

Latest sources

Mentioned together with

low-rank gradient projectionreinforcement learning from policy gradients for language models