Proposes an RL post-training framework for joint audio-video diffusion generation that fixes credit-assignment problems from entangled multimodal rewards and the high cost of jointly optimizing two modality towers. Uses modality-anchored rollouts, trajectory-locked frozen-tower optimization, and modality-specific adaptive objectives, paired with a new decoupled, difficulty-controllable dataset (5DAV). Outperforms LTX-2.3 on JavisBench and VABench in generation quality, semantic alignment, and cross-modal sync under both LoRA and full fine-tuning; code and data released.
