Reproduce “TRIAGE: Direction-Aware Mismatch Stabilization of Native NVFP4 Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhen Li, Shuai Zhang, Yanggan Gu, Yiming Zhang, Yang Yu, Mingfa Feng, Congkai Xie, Shuang Yu and 2 more
TRIAGE stabilizes long-horizon reinforcement learning with native 4-bit execution by diagnosing whether policy updates locally widen or narrow sampler–learner probability gaps. It gates locally amplifying updates selectively and adds a bounded repair term, without changing the W4A4 forward path. Across Qwen3 runs, it stayed stable, scoring 70.96% on 30B, 1.45 points below BF16; 100B+ models and longer horizons remain untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.