Reproduce “Fine-Tuning of Continuous-Time Diffusion Models as Entropy-Regularized Control” 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
Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani and 1 more
ELEGANT fine-tunes diffusion models to increase learned reward while regularizing generation against the pretrained process, including its starting distribution. Its ideal objective preserves the pretrained model’s support, though the practical algorithm is approximate. On TFBind8 at α = 0.01, ELEGANT reports genuine reward 0.88 ± 0.01 versus 0.84 ± 0.01 for KL-penalized PPO at nearly matched nominal rewards; this is benchmark evidence, not wet-lab validation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.