Reproduce “GRADUS: Discovering and Recovering Agent Lessons from the Environment's Latent Syllabus” 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
Qing Ping, Wei Yang, Panpan Xu, Johannes Kirmayr, Lukas Stappen, Junjie Tang, Jun Huan
GRADUS builds training curricula for policy-constrained tool environments by tracking policy–outcome coverage and refining tasks and reference solutions through execution-based verification. Failed checks can route a task for repair rather than discard its intended behavior. In a matched comparison, a Qwen3-8B model trained on GRADUS data achieved 74% on CAR-Bench hallucination versus 52% for trajectory-first synthesis, on one split.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.