Reproduce “Rethinking LLM Pruning via Test-Time Scaling” 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
Mingyang Song, Mao Zheng
The authors study how pruning changes accuracy and repeated-sampling coverage. They compare original and pruned models using pass@k, the estimated fraction of problems with a grader-accepted answer among k samples, and measure attempts needed to recover coverage. At 50% expert pruning, Qwen3.6-35B-A3B solves all 29 AIME problems observed for the original in 2,048 samples per problem, not every AIME problem.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.