Reproduce “A Score Is Not a Policy: Measuring the Value of Adaptive Revision” 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
Ayushi Chadha
The authors test whether a learned controller improves when a hierarchical reasoner revises its latent commitment during iterative reasoning. The controller is compared with fixed schedules on three precommitted seeds. Adaptive timing does not beat the best tested fixed schedule on any seed; on one frozen checkpoint, the best fixed schedule captures about 71% of headroom over random timing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.