Reproduce “EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling” 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
Shimao Zhang, Yu Bao, Shujian Huang
Entropy-based EDT adjusts temperature at each decoding step using the model’s next-token uncertainty. In MS MARCO v1.1 question answering with fine-tuned LLaMA2-13B, it reports EDA_range 29.49 versus KLD’s 35.69, where lower is better. EDT has the best EDA_range on all four evaluated datasets, but the settings are task-specific and it does not lead on every composite measure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.