Exécution de notre outil de synthèse sur un nouvel article

Résumé : Recent advances in large vision language models (LVLMs) have unlocked unprecedented capabilities in generating coherent responses from visual inputs. However, interpreting where LVLMs direct their visual attention while generating free-form textual responses remains a significant challenge, yet is essential for understanding model behavior, diagnosing hallucination, exposing bias and ensuring transparency. We introduce GLIMPSE (Gradient-Layer Importance Mapping for Prompted Visual Saliency Explanation), a lightweight, model-agnostic framework for visualizing the salient image regions that LVLMs rely upon during open-ended visual question answering (VQA), while concurrently revealing the multimodal textual saliency. GLIMPSE fuses gradient-weighted attention, adaptive layer propagation, and weighted token aggregation to produce holistic response-level attribution heat maps for interpreting cross-modal reasoning, outperforming prior interpretability methods in human-alignment. We demonstrate an analytic explainable AI (XAI) approach using GLIMPSE to uncover fine-grained insights into LVLM cross-modal attribution, trace token-level reasoning dynamics, and analyze systematic human-attention misalignment, hallucination, and bias.