Reproduce “Depth as a Candidate Source: Portfolio Allocation for Recurrent Test-Time Scaling in Large Language Models” 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
Chaoyuan Hao, Wentao Yue, Yanyan Wu, Boyuan Gao, Tianyou Lai, Maolin Wang, Yongxin Ni, Hanwen Du and 1 more
DPA allocates recurrent calls across depths to generate candidate answers rather than relying on one stopping depth. The authors compare portfolios under matched candidate counts and nominal recurrent-step costs. On MBPP, four depths achieve 36.23% oracle coverage versus 31.72% for four candidates at depth 16, at cost of 64 steps; this supports source allocation in this setting, not runtime savings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.