The paper addresses multi-reward for joint audio-video . Existing methods use fixed routing and weights, which become stale as the model evolves. The authors propose Adaptive Reward Routing with two components: Cross-Modal Influence-Guided Routing localizes updates by tracking cross-attention responses, and Preference-Preserving Modality-Aware Reweighting adjusts reward weights while preserving user-defined priorities.
The method is applied within the DiffusionNFT forward-process RL framework. Experiments show consistent improvements in modality quality, semantic alignment, and audio-video synchronization over strong RL baselines. Ablations confirm the complementary contributions of token- and layer-level routing, branch-aware conflict estimation, and residual preference correction.
The work is interesting because it treats multi-reward optimization as a dynamic credit-assignment problem, adapting both update localization and reward coordination as training progresses.