Small vision-language models improve recognition on low-power devices

Summary

The paper credits distillation data quality and visual encoder compression for making on-device multimodal apps more practical.