Reproduce “Efficient Resource-Shiftable RL Post-training with Generative Rewards” 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
Yanfeng Zhao, Haoyang Li, Sheng Lin, Yuming Zhou, Jiawei Jiang, Bin Cui, Fangcheng Fu
PivotRL reallocates GPUs among language-model rollout, generative reward evaluation, and training as workloads change. Its runtime transfers model weights and manages role-specific state, while schedulers migrate sequences and a planner balances expected throughput gains against shift costs. On a 64-GPU Qwen2.5-32B and Qwen3-30B-A3B setup, it reports 20,503 ± 58.32 tokens/s, 1.50× disaggregated deployment; results cover tested math-focused workloads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.