Reproduce “Diffusion Large Language Models for Black-Box 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

Ye Yuan, Chen, Zipeng Sun, Dinghuai Zhang, Christopher Pal, Xue Liu

The authors introduce dLLM, using a pretrained diffusion language model to fill masked designs bidirectionally and search partial drafts. On TF Bind 8, using 10 offline examples and 128 candidates, its maximum normalized score is 0.876 ± 0.038, versus 0.848 ± 0.023 for MCTS-transfer. It ranks first on four benchmark tasks; no lab or physical validation is reported.

Key insights

  • Diffusion large language models are inherently better at modeling bidirectional dependencies in design tasks compared to autoregressive LLMs, leading to more effective optimization.
  • The vast pre-trained knowledge within the LLM backbone is crucial for effective generalization and generating creative designs, especially in data-scarce, few-shot optimization settings.
  • Integrating a principled search strategy like Masked Diffusion Tree Search (MDTS) is essential to guide the iterative denoising process of diffusion LLMs, ensuring the discovery of high-performing designs.

Summary by alphaXiv.

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