Reproduce “Entropy Regularization: A Free Correction to Cross-Entropy for Verified Demonstrations” 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
Mihir Dhanakshirur, Adam Ousherovitch, Ambuj Tewari
The paper adds a token-level Shannon-entropy penalty to cross-entropy fine-tuning on correct demonstrations, discouraging probability spread while retaining the recorded solution as a target. In GSM8K, Qwen2.5-1.5B-Instruct’s greedy Pass@1 rose from 0.4912 ± 0.0154 to 0.5816 ± 0.0161 at λ=8 across five seeds. This result is limited to tested Pass@1 settings and conditional theory.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.