In LLaVA-OV-7B, a linearly decodable conflict signal appears in intermediate layers and detection-related attention shifts precede resolution-related ones, supporting a detection/resolution separation in the model.
Answer, assemble, ace: Understanding how lms answer multiple choice questions
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Challenges in Understanding Modality Conflict in Vision-Language Models
In LLaVA-OV-7B, a linearly decodable conflict signal appears in intermediate layers and detection-related attention shifts precede resolution-related ones, supporting a detection/resolution separation in the model.