Reproduce “JEV-as-a-Judge: Accept When Confident, Escalate When Unsure” 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
Yubo Li, Yidi Miao, Ramayya Krishnan, Rema Padman
The authors evaluate a cascade that uses JEV, a decision-only judge, to screen answer pairs and sends selected cases to GPT-6. JEV’s largest label probability guides routing, but is measured against correctness rather than treated as calibrated. On two prospective correctness workloads, the cascade matched GPT-6’s accuracy on 570 held-out pairs, escalating 74.2% at 75.5% of its estimated fee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.