Reproduce “Adaptive Entropy-Sparing for Efficient 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
Cong Jiang, Xiaofeng Zhang, Tom Ko, Zheng Zhang
AES reduces excess reasoning during post-training by penalizing high-entropy tokens only in the overlong suffix of correct responses, beyond the shortest correct response in a sampled group. It leaves earlier uncertainty and incorrect responses unchanged. On four mathematical benchmarks, mean pass@1 rises from 64.9% to 68.7%, while mean response length falls 52.9% versus the base model; this is an average, not a guarantee across tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.