Vision Inference Former: Sustaining Visual Consistency in Multimodal Large Language Models

In three linesVision Inference Former (VIF) is a lightweight architectural module improving visual consistency in multimodal models. It continuously injects visual semantics during generation to counter weakening vision-language alignment over long sequences. Tested on 14 benchmarks (reasoning, OCR, tables), VIF improves performance with minimal overhead.

Summary generated by Claude — human-verified