Reproduce “Decision-Focused Reinforcement Learning for LLM Post-Training” 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

Haotian Zhai, Leqi Liu

Decision-focused reinforcement learning trains language-model predictors to minimize downstream regret rather than forecast error. A score-function update uses reward without differentiating through the optimizer; efficiency methods approximate costly evaluations. On the paper’s H&M inventory task, reasoning-trace SFT followed by thinking-enabled RLOO reduces test regret 31.4% versus SFT, a result limited to this constructed task.

Summary by alphaXiv.

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