Kaiwen Long

MS-Resampler boosts multimodal reasoning capabilities while maintaining efficiency, outperforming traditional methods with minimal overhead.

Instruction-guided visual modulation with iGVLM unlocks more fine-grained reasoning in LVLMs, outperforming static vision encoders by dynamically adapting visual representations to the specific textual task.

Image-text models can achieve superior performance by fusing modalities during training only, then discarding the fusion module at inference for efficiency.