A causal attribution method using steering vectors reveals that LVLMs aggregate visual emotions in middle layers via specific attention heads and translate them in deep layers via general pathways, enabling targeted interventions that improve performance on MER-UniBench.
Twin Image
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow
A causal attribution method using steering vectors reveals that LVLMs aggregate visual emotions in middle layers via specific attention heads and translate them in deep layers via general pathways, enabling targeted interventions that improve performance on MER-UniBench.