Reproduce “DreamRealm: Exploring and Materializing Hallucination Worlds for Trustworthy 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

Benlei Cui, Chunyang Chai, Yiliang Zhang, Xiaoqian Xia, Chengwen Yao, Song Li, Longtao Huang, Yuwen Zhai and 1 more

DreamRealm trains vision-language models to match answers with images that support them. It mines observed and latent hallucinations from a target model, edits images to make selected answers valid, and uses supervised fine-tuning with bidirectional image–response contrast. On Qwen3.5-9B, HallusionBench item accuracy rises from 70.95% to 78.21%, and MMVP pair accuracy from 63.33% to 74.67%, under the authors’ evaluation setup.

Summary by alphaXiv.

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