Reproduce “CompressionBench: How can your coding agents compress LLMs?” 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
Xinrui Zhong, Zhifeng Li, Zhiming Wang, Xinze Feng, Yongqi Liu, Shanglin Yang, Yang Zhou, Zikun Li and 2 more
The authors introduce CompressionBench, where coding agents compress fixed language models and build inference code under storage, hardware, and time limits. Independent verification checks storage, generated-answer accuracy, and artifact integrity. Twenty-nine of 36 submissions pass, including all 18 looser-budget runs; selected artifacts average a 3.4% perplexity increase and 77.33% accuracy, showing feasibility on tested tasks, not recursive self-improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.