Reproduce “Veritas: A Verifier Generation Harness for Agent Tasks” 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
Daoan Zhang, Qinbo Bai, Dehai Min, Xingyu Mu, Ziyi Chen, Jianing Yin
Veritas generates executable task verifiers for self-improving agents, targeting the verifier-writing bottleneck. It gives a coding agent an evidence boundary, searchable domain rules, and authoring principles, then probes and repairs the verifier. Using the same model, Veritas agrees better with official verdicts than seven plain harnesses on all four execution-scored benchmarks, but this does not show improved self-improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.