Reproduce “Pareto Frontier of LLM Test-Time Compute: Adaptive Rollouts and Voting Under Token Budgets” 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
Zhenyu Wang, Xiaozhi Zhu, Yifan Hu
PRICE allocates attempts and selects a voting rule for each question under a shared generation-token budget. It predicts accuracy and cost for attempt counts and score-weighted votes, adapting the vote as well as the count. On MATH-500 with Qwen2.5-1.5B, PRICE-deployed reaches 61.3% accuracy at a 3.7k entropic-risk budget, versus 54.9% for the best baseline; this is a matched-budget replay, not a live deployment guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.