Reproduce “Mitigating Mask Prior Drift and Positional Attention Collapse in Large Diffusion Vision-Language Models” 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
Sujung Hong, Chanyong Yoon, Seong Jae Hwang
The authors introduce two training-free inference-time interventions for diffusion vision-language models that iteratively unmask response tokens in parallel. Mask Prior Suppression (MPS) adjusts masked-token representations; Monotonic RoPE Scaling (MRS) changes positional handling to support longer-range visual interactions. On LaViDa’s DetailCaps, CAPTURE rises from 8.3 to 56.1; MPS alone matches the score, limiting this result to that task and setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.