Reproduce “Demasking Fine-Tuning for Knowledge Injection in Autoregressive and Diffusion LLMs” 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

Xu Pan, Ely Hahami, Jingxuan Fan, Ziqian Xie, Haim Sompolinsky

The authors introduce demasking autoregressive fine-tuning, which trains a model to reconstruct a document from a version with randomly masked words. In controlled tests without paraphrases, four autoregressive models averaged 0.895 ROUGE-1 recall on forward questions and 0.919 on backward questions, versus 0.257 and 0.025 with standard fine-tuning, at equal total FLOPs; evidence remains limited to controlled datasets.

Summary by alphaXiv.

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