Reproduce “Value-Conditioned Proposal Acquisition for 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
Zeyan Li, Jianfeng Xu
ValueFormation allocates test-time effort among stopping, sampling the primary model again, or acquiring a proposal from a specialist. It uses the evolving answer pool to estimate each action’s value after token cost. On held-out MATH-500, a Llama-3.1-8B-Instruct primary reaches 80.29% accuracy, versus 62.46% for the strongest evaluated baseline, ESC; gains are smaller and unresolved on high-coverage GSM8K.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.