HKUST, Tencent EmoReward self-calibrates video-RL rewards

HKUST and Tencent Hunyuan researchers have released EmoReward, a training scheme that learns which reward signals to trust across successive RL training runs for emotional video generation. On HunyuanVideo 1.5 with SAGE-GRPO, human evaluators marked 60.0% of EmoReward's clips 'Good' for emotion fidelity, against 53.0% for the SAGE-GRPO baseline.

The paper names two moving parts. "Reflective Reward Gating calibrates their contributions through instance-level gates for phase, face confidence, conflict, and subtlety," the abstract writes, picking how much to weight each of seven reward providers on a given sample. "RewardMemory then estimates each primitive's reliability from rollout scores and outcomes within each emotional regime in the finished training run," then fixes bounded multipliers that steer fusion in the next run "without updating reward models."

On the 302-prompt emotional-evaluation set, VideoAlign Overall moved from 0.6921 for SAGE-GRPO to 0.7566 with EmoReward stacked on top; motion-quality and text-alignment heads moved in step. The human study covered 18 participants and 30 prompts, with temporal coherence at 54.3% Good and visual quality at 57.2%. Transfer experiments onto Stable Diffusion 3.5 and FLUX are included, though the paper frames them as adaptive-calibration support rather than clean wins.

Code is at Tencent-Hunyuan/EmoReward. The release lands in a busy stretch of video-RL training papers on our tracker this week, alongside notes on velocity-adversarial distillation and streaming-video credit assignment from Fudan and JD.