Reproduce “Confidence-Calibrated Inference Expansion for Evaluator-Guided Test-Time Reasoning” 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
Weida Liang, Kenji Kawaguchi
The controller chooses whether to trust an answer, return a fallback, or spend compute on another inference step. It estimates success for each action and prices continuation. On held-out math tasks, it matched width-48 majority vote utility, 0.715, at normalized compute 0.547 rather than 1.000. This token-based operating-point result does not establish lower latency or a stronger reasoner.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.