EmoReward calibrates emotional reward weights across RL training runs for video generation
Researchers introduce EmoReward, a self-evolving mechanism that learns which emotional signals to trust during reinforcement learning and adjusts reward weighting in the next training run.
In plain words
Training video models to show emotion is hard because there is no single correct answer for how a scene should look or feel. EmoReward is a training mechanism that reviews the reward scores from a finished training run, estimates which emotional signals proved reliable, and uses that to rebalance reward weights in the next run. The authors report gains on HunyuanVideo 1.5 and transfer tests on other image models.
What was announced
The paper's authors argue that fixed rewards for visual quality, motion, and text alignment capture emotional intent poorly and can invite reward hacking, where a model exploits flaws in the scoring signal. Reinforcement learning from human feedback, in turn, is described as hard to scale and sensitive to individual aesthetic differences when calibrating facial, body-motion, and color cues across narrative contexts. EmoReward addresses this by planning grounding cues from the emotional prompt, scoring generated videos with heterogeneous signal primitives, and applying Reflective Reward Gating, which uses instance-level gates for phase, face confidence, conflict, and subtlety before renormalizing weights over the active set into a fused reward.
The self-evolving step comes from RewardMemory, which estimates each signal primitive's reliability from rollout scores and outcomes within each emotional regime of a completed training run, then sets bounded multipliers that calibrate fusion weights in the next run without updating the reward models themselves. The authors report experiments on HunyuanVideo 1.5 showing improvements across multiple RL algorithms, comparisons with other agentic reward methods under SAGE-GRPO, and ablation and transfer experiments on SD 3.5 and FLUX supporting adaptive reward calibration for visual generation.
What it does not show
- The evidence is the paper's abstract only; results, comparisons, and transfer findings are author-reported and not independently confirmed.
- The abstract does not state quantitative results, model availability, or how EmoReward compares in magnitude to baselines.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.