We also found that multimodal reasoning capabilities emerge naturally after online RL (even when RL is only done on math & code text data), as long as you start from an initial multimodal model.

→ View original post on X — @guillaumelample,