Reproduce “Seeing is Believing: Benchmarking and Mitigating the Interference of Language Priors on Visual Perception in MLLMs” 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
Li Ma, Yan Zhuang, Jiawen Deng, Fuji Ren
The authors introduce ViSiP-Bench, a benchmark of 4,500 image-question pairs across six categories, and MAPS, which adds knowledge about rare-but-real cases to soften strong model expectations. Across 24 models, prior adherence is widespread and does not reliably decline with scale. On Qwen3-VL-8B-Instruct, MAPS lowers PAR from 42.67% to 35.09% and raises filtered accuracy from 51.29% to 59.01%, on this benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.