Reproduce “StartLux-Decision: AutoThink Decision Models with Reinforcement Learning for Calibrated Deliberative Decisions” 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
Xialie Zhuang, Zunhai Su, Jia Pan, Ngai Wong, Guanhua Chen, Chuan-Wei Kuo
StartLux-Decision keeps a one-pass probability-based choice for confident cases and invokes reasoning when the top probability falls below a threshold, then blends the initial and reasoned probabilities. Its RLCD² training rewards improvements in probability quality and penalizes trace tokens beyond a budget. On DecisionPivot, 27B AutoThink reaches 86.9% pair accuracy, above GLM-5.3 at 82.2%; this result applies to benchmark task families.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.