Reproduce “Evidence-First Agentic Verification” 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
Reza Sanayei, Razvan-Gabriel Dumitru, Sunjie Hou, MohammadHossein Rezaei, Haris Riaz, Robert Vacareanu, Minglai Yang, Ying Liu and 2 more
Sauron evaluates agents against rubrics by investigating environment state and execution traces, then checking each proposed factual claim against its cited source before using it in a verdict. On PA₁₀₀, with GPT-5.6 Sol, it achieves 0.96 F1 versus Gandalf’s 0.92 at 23% lower model-call cost. In controlled PA-Fault tests, it detects 97% of injected faults and flags 2% of clean controls.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.