Reproduce “The Verifier Is Part of the Simulator: Evidence-View Gaming in Agentic Equation Discovery” 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
Orestis Oikonomou, Bogdan Raonic
The authors test whether choosing numerical summaries changes a verifier’s equation-discovery verdict on simulated trajectories. In one-dimensional cases, maximizing the public score wrongly rejects the library in 1,500 in-library episodes; in a 900-episode PDEBench transfer, it makes 69 false identifications. A fixed fine view answers all 900 correctly, but evidence is limited to simulated data and tested views.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.