Reproduce “CAST: Cross-modal Alignment Similarity Test for Vision Language Models” on OpenResearch

About this paper

Gautier Dagan, Olga Loginova, Anil Batra

Researchers developed the Cross-modal Alignment Similarity Test (CAST) to evaluate the self-consistency of Vision Language Models across different input modalities. Their findings indicate that current open-source VLMs often exhibit inconsistencies, especially when transferring reasoning between visual and textual inputs or generating longer sequences, with GPT4o-Mini and MiniCPM V2 performing best.

Key insights

• A VLM's capability necessitates self-consistency across modalities, meaning a truly capable model should maintain stable reasoning about the same information regardless of whether it is presented visually or textually.
• The majority of observed inconsistencies stemmed from the generation phase, with high rates of hallucination in the image modality, suggesting VLMs struggle to reliably verify visual attributes and relationships across multiple scenes.
• Reinforcement Learning from AI Feedback (RLAIF) did not universally improve, and in some cases decreased, multi-modal self-consistency in LLaVA-1.5, indicating current RLAIF methods may not inherently optimize for this specific aspect.

Summary by alphaXiv.