Reproduce “Cross-Modal Attention Acts as a Frequency Filter: Why Verbose Prompts Improve Robustness in 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
Farooq Ahmad Wani, Maria Sofia Bucarelli, Mujtaba Hussain Mirza, Oleksandr Pryymak, Aryo Pradipta Gema, Iacopo Masi, Pasquale Minervini, Fabrizio Silvestri
The authors study how question wording affects vision-language model sensitivity to image corruption. They compare terse questions with verbose, meaning-preserving mirrors, define drift as the clean-to-corrupted change in gold-answer log-probability, and relate it to attention-derived visual-frequency filters. Verbosity tends to reduce drift, while finer reasoning tends to increase it; on Qwen3-VL-8B GQA, the ratio was 0.30, a correlational, benchmark-bound result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.