Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning

Paper recorded by Signals 4 on 2026-10-23 in cs.LG. Abstract reproduced from arXiv; link to the original below.

Published 2026-10-23 on arXiv · recorded by Signals 4 on 2026-10-26

Category: cs.LG · 机器学习 · first seen 2026-10-26

Abstract

Reinforcement learning for large language models often uses separate systems for inference and training. Even with identical weights and BF16 tensors, implementation differences can make an intended on-policy update off-policy and destabilize training. Existing remedies align the entire inference and training stacks, change global precision, or modify the learning algorithm. We ask whether alignin

Read on arXiv →