Carol0110/UniRM-7B

English | 中文

UniRM: Multi-Head Scalar Reward Model for Multimodal Moderation

UniRM is a multi-head scalar reward model that provides interpretable, attribute-level scoring for multimodal moderation. It was introduced in the paper From Sparse Decisions to Dense Reasoning: A Multi-attribute Trajectory Paradigm for Multimodal Moderation.

• Project Page: trustworthylab.github.io/UniMod/
• Repository: github.com/Carol-gutianle/UniMod
• Paper: arXiv:2602.02536

UniRM is designed to support policy optimization for open-ended reasoning in UniMod, especially for the posterior response stage where deterministic labels are absent. It decouples reward attribution into multiple dimensions so the model can distinguish stylistic quality from safety boundaries (privacy, bias, toxicity, legality), enabling transparent diagnosis and stable optimization.

---

Demo Video

UniRM demo video:

---

Quick Start (Gradio)

Below is a minimal Gradio demo that loads UniRM and returns multi-head scores for a (prompt, response, optional image) triple.

git clone https://github.com/TideDra/lmm-r1.git
cd lmm-r1
pip install -e .[vllm]
pip install flash_attn --no-build-isolation
python unirm.py --model_path {PATH_TO_UNIRM}

Citation

@misc{gu2026sparsedecisionsdensereasoning,
      title={From Sparse Decisions to Dense Reasoning: A Multi-attribute Trajectory Paradigm for Multimodal Moderation}, 
      author={Tianle Gu and Kexin Huang and Lingyu Li and Ruilin Luo and Shiyang Huang and Zongqi Wang and Yujiu Yang and Yan Teng and Yingchun Wang},
      year={2026},
      eprint={2602.02536},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.02536}, 
}