AIAny
Icon for item

RewardVerse: Rubric-Guided Policy Optimization for Video Reward Modeling

Introduces a rubric-based approach for video reward modeling that generates explicit, query-adaptive evaluation criteria before scoring to reduce scalar drift; proposes RGPO, a two-stage training (seed warm-up + joint optimization) that achieves strong pointwise and pairwise evaluation with high data efficiency.

Introduction

Why this matters

Video generation models need reliable reward signals for reinforcement learning, but mapping subjective, multi-dimensional video quality into a single scalar causes unstable scales and score drift across contexts. RewardVerse flips the order: instead of unconstrained scalar scoring, it first generates a dynamic rubric (explicit themes, weights, and scoring tips) tailored to the query, then scores the video against that rubric. That intermediate semantic anchor both improves interpretability and stabilizes numerical rewards for downstream RL.

Key Findings
  • Rubric-as-intermediate: Decouples evaluation into rubric generation and rubric-guided scoring. The rubric generator decomposes a text query into evaluation themes and weighted scoring tips; the scorer rates video aspects independently against those criteria rather than relying on an implicit internal standard. This reduces score-range collapse and context-driven scalar drift.
  • RGPO (Rubric-Guided Policy Optimization): A two-stage training pipeline under Group Relative Policy Optimization. Stage 1 warms up the scorer using self-evolving seed rubrics synthesized offline; stage 2 jointly optimizes the rubric generator to be query-adaptive while continuously aligning the scorer with human ratings. Both stages are necessary for stable performance.
  • Data efficiency and evaluation: RewardVerse reports state-of-the-art pointwise correlation on the 16-dimensional EvalVerse benchmark and strong pairwise agreement on external datasets, claiming that only ~30 preference pairs per dimension are sufficient to learn effective rubrics. The output rubric also serves as an interpretable reward signal for RL.
  • Practical implications: The method provides more consistent scalar rewards across different prompts, makes the scoring process explainable (you can inspect rubrics and per-dimension scores), and supports both pointwise and pairwise evaluation formats natively.
Who It's For and Trade-offs

Great fit if you are training or fine-tuning video generation models with RL and need a stable, interpretable reward model rather than opaque single-number scorers. The approach benefits settings where human alignment is costly: RGPO aims to reduce the number of human preference pairs required.

Look elsewhere if you need a plug-and-play scalar scorer without the extra complexity of a rubric generator (rubric generation and joint optimization add modeling and training complexity), or if your evaluation domain is far from the video/text queries used in the paper—rubric quality and transferability depend on the rubric generator and the alignment process. The method is primarily validated on the EvalVerse benchmark and related external video datasets; cross-domain robustness beyond similar multimodal video-text settings may require extra tuning.

Where It Fits

RewardVerse sits between conventional end-to-end reward models and hand-crafted multi-attribute evaluation: it automates the generation of multi-criteria rubrics while preserving explicit, per-dimension judgment. For teams building RL pipelines for text-to-video or multimodal video generation, it offers a middle ground—more structure and interpretability than opaque scalar RMs, more scalability than fully hand-crafted rubrics.

Information

  • Websitearxiv.org
  • OrganizationsNew Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, School of Artificial Intelligence, University of Chinese Academy of Sciences, The Hong Kong University of Science and Technology, Tencent†Project Lead   *Corresponding Author
  • AuthorsZhenchen Tang, Yang Li, Songlin Yang, Bo Peng, Xiaotong Zhao, Shuai Li, Haotian Fan, Alan Zhao, Jing Dong
  • Published date2026/09/19

More Items

Teaches vision-language models to predict and integrate physical-world state transitions from interaction trajectories (observation → action → next observation). Introduces a three-level curriculum and the LSI-108K dataset, and applies supervised fine-tuning plus on-policy distillation to improve local transition modeling and long-horizon spatial integration.

Surveys memory mechanisms for autoregressive video generation, framing memory as persistent historical information that influences future generation. Organizes work by Forms, Functions, Operations, Learning, and Evaluation, and synthesizes challenges for long-horizon consistency and memory-aware learning.

Evaluates whether generative world models maintain consistent, controllable, and physically plausible simulated environments under exploration, interaction, and intervention. Introduces a six-level W1–W6 capability taxonomy across three tracks (video, spatial, embodied) with human A/B Arena and automated metrics to measure behavioral correctness.