Reproduce “Controlling Sharpening in 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
Simon Matrenok, Nicolas Menet, Nicolas Flammarion, Andreas Krause
KAPPO adapts single-completion reinforcement-learning post-training to improve performance on pass@k and max@k objectives. It calibrates a reference model’s prompt-level success rates, then applies less optimization pressure to reliably solved prompts and more to harder ones. On MATH Level 5 validation, Qwen3-4B reaches 70.48% pass@2, versus PKPO’s 70.77%; this supports comparable performance, not new reasoning capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.