Reproduce “Silent Instrument Errors Reorder Agent Leaderboards” 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
Yangze Liu
The paper audits silent errors in agent evaluation pipelines. In HAL’s Airline benchmark, many Generalist episodes were stored as failures because they never triggered grading. Applying the benchmark’s own rule offline raises pooled accuracy on 41 selected tasks from 0.30 to 0.43, closing the published 13-point gap with Tool-Calling to a tie; this does not match the scaffold designs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.