visual-debiasing

This paper presents OmniClean, a visually debiased evaluation benchmark for omni-modal language models, and proposes OmniBoost, a three-stage post-training recipe that enables a 3B model to match the performance of a 30B model on the cleaned benchmark.