Reproduce “An Analytical Model of Compute-limited Multistage Training Pipelines” 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

Nishil Patel, Jin Hwa Lee, Basile Confavreux, Andrew Saxe

The authors model how pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) interact under a budget. Their teacher–student system treats pretraining as representation learning, SFT as decision-level feedback, and RL as delayed episode-level reward. SFT initialization accelerates early reward improvement in theory and an OLMo-2 1B GSM8K experiment; modeled optimal allocations depend on label noise and budget.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.