Abstract: Large behavior models (LBMs) have shown strong dexterous manipulation capabilities by extending imitation learning to large-scale training on extensive multi-task robot data, yet their generalization remains limited by the insufficient coverage of available robot data. To expand this coverage without costly additional data collection, recent work increasingly relies on co-training: jointly learning from target robot data and heterogeneous data modalities. However, how different co-training data modalities and training strategies affect policy performance remains poorly understood. We present a large-scale empirical study examining five co-training data modalities—standard vision-language data, dense language annotations for robot trajectories, cross-embodiment robot data, human videos, and discrete robot action tokens—across single- and multi-phase training strategies. Our study leverages 4,000 hours of robot and human manipulation data and 50M vision–language samples to train vision-language-action (VLA) policies. We evaluate 89 policies over 58,000 simulation rollouts and 2,835 real-world rollouts. Our results show that co-training with various forms of vision-language and cross-embodiment robot data substantially improves generalization to distribution shifts, unseen tasks, and language following, while discrete action token variants yield no statistically significant benefits. Furthermore, combining effective modalities produces cumulative gains and enables rapid adaptation to unseen long-horizon dexterous tasks via fine-tuning. Together, these results provide a systematic understanding of co-training and practical guidance for building scalable generalist robot policies.