Cong Wang

Freezing most weights and only LoRA-tuning a vision-language model achieves near state-of-the-art multimodal interleaved reasoning performance, proving that targeted adaptation can rival full fine-tuning.