New benchmark detects early visual bias in multimodal AI models

A new benchmark called Tri-PvP reveals that visual bias dominates early representation layers in multimodal language models. Visual signals account for over 60% of bias in these systems, overshadowing audio and text cues. Researchers found this bias is detectable from the first few transformer layers using simple linear probes. A contrastive decoding method reduces bias during inference while maintaining task performance with minimal accuracy loss. However, the technique does not fully eliminate early-layer bias, suggesting more fundamental architectural solutions are needed.