Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow

T1Covered by 1 source · first reported by arXiv at 25 Sept, 04:00 UTC

arXiv:2605.21980v2 Announce Type: replace-cross Abstract: Large Vision-Language Models (LVLMs) represent a significant leap towards empathetic agents, demonstrating remarkable capabilities in emotion understanding. However, the internal mechanisms governing how LVLMs translate abstract visual…

science

Sources

• T1Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information FlowarXiv25 Sept, 04:00