Researchers have introduced a new metric called Synergistic Faithfulness ($\mathcal{F}_{syn}$) to better evaluate the explainability of Vision-Language Models (VLMs). Current methods often fail because VLMs can answer visual questions using text alone, leading to contradictory evaluation results. This new metric, based on the Shapley Interaction Index, accurately isolates the joint contribution between modalities and is significantly faster than existing approaches. Evaluations using $\mathcal{F}_{syn}$ show that many VLM explainability methods overemphasize visual saliency and underperform compared to attention-based methods in capturing true cross-modal synergy. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Provides a more rigorous framework for auditing VLM reasoning, crucial for safe deployment in high-stakes applications.
RANK_REASON Academic paper introducing a new evaluation metric for VLM explainability. [lever_c_demoted from research: ic=1 ai=1.0]