Reproduce “Language Model Decoding as Direct Metrics Optimization” 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

Haozhe Ji, Pei Ke, Hongning Wang, Minlie Huang

DAEMON reweights complete candidate continuations to match selected metric expectations from human-written text while staying close to the input language model. It uses sampling-importance-resampling to select among candidates. Across Wikipedia and Wikinews with GPT-2 XL and OPT-6.7B, it achieved the highest MAUVE and lowest SR-4 in all four comparisons; this does not guarantee every output is human-like.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.