Reproduce “Reliability Challenges in 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
Md. Atabuzzaman, Chris Thomas
The paper evaluates six diffusion vision-language models against autoregressive models across hallucination, demographic bias, and selection bias. It tests whether iterative masked-response denoising introduces reliability failures. On CUB bird questions without class names, LaViDa-LLaDA reaches 6.40% when the correct option is shorter than distractors, versus 97.70% when longer. Early predictions already matched final answers in 97.30%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.