Reproduce “A Block Metropolis-Hastings Sampler for Controllable Energy-based Text Generation” 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
Jarad Forristal, Niloofar Mireshghallah, Greg Durrett, Taylor Berg-Kirkpatrick
MH-BLOCK uses Flan-T5-XXL to propose whole-sentence rewrites, while a Metropolis–Hastings accept/reject step evaluates them with an energy combining expert scores. In a limited prompted LLaMA-7B paraphrase test, its samples had mean energy closer to exact samples than Mix and Match, using 10 rather than 130 target-model forward passes. This does not establish general sampling guarantees or wall-clock speedup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.