LVLMs show language bias from training prioritizing text over vision in VIT and DPO; LBR regularization and LBP penalty mitigate it across benchmarks without added data.
The intuition is to regularize the average language bias per token rather than the cumulative language bias of the entire sequence
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation
LVLMs show language bias from training prioritizing text over vision in VIT and DPO; LBR regularization and LBP penalty mitigate it across benchmarks without added data.