Reproduce “FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation” 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
Yining Wei, Tianyuan Wu, Gu Gong, Yuechen Tao, Yinghao Yu, Xiaowen Chu, Wei Wang
FourSight makes reinforcement-learning post-training for language models more compatible with NVFP4 deployment by filtering drifted response trajectories and protecting high-variance activation channels with BF16. Its fused SpliceGEMM kernel combines BF16 and NVFP4 computation. In Qwen3-8B DAPO experiments, it reached 20% AIME24 validation accuracy at 150 steps with Megatron and 250 with FSDP; these are step-efficiency results, not wall-clock gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.