Task-Specific Loss Analysis in Multimodal Tokenizers

Multimodal continual pretraining reveals that image and text tokens exhibit distinct scaling behaviors across different tasks. Evaluation shows that joint modeling success depends heavily on the predicted token space, making task-specific validation loss more critical than isolated metrics.