Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 Oral)

Shijie Zhao*, Xuanyu Zhang*, Weiqi Li, Junlin Li, Li Zhang, Tianfan Xue, Jian Zhang

Bytedance Inc.

🚩 Updates

• 2026.02.14 The inference code of RALI is released!
• 2026.01.26 RALI has been accepted at ICLR 2026 as an oral presentation!

🔥 Introduction

We revisit the reasoning mechanism in MLLM-based IQA model (such as Q-Insight) and propose a CLIP-based lightweight image scorer RALI. We verifies that through RL training, MLLMs leverage their reasoning capability to convert redundant visual representations into compact, cross-domain aligned text representations. This conversion is the source of the generalization exhibited by these reasoning-based IQA models. RALI uses only about 4% of Q-Insight’s parameters and inference time, while achieving comparable accuracy.

🔧 Dependencies and Installation

git clone https://github.com/xuanyuzhang21/RALI.git
bash setup.sh

⚡ Quick Inference

Please download the RALI pretrained weights from the link. After downloading, place the checkpoint under ./checkpoints, so that the directory structure becomes:

RALI/
├── checkpoints/
│   ├── ckpt.pt
│   ├── pca.pkl
│   ├── basis.npz
│   └── best/
│       ├── config.json
│       ├── pytorch_model.bin (or *.safetensors)
│       ├── preprocessor_config.json
│       └── ...

Then run the following code:

python demo_rali_score.py

📖 Dataset Preparation

Download meta files and source images from Data-DeQA-Score and arrange the folders as follows:

|-- RALI 
    |-- Data-DeQA-Score
        |-- KONIQ
            |-- images/*.jpg
            |-- metas
        |-- KADID10K
            |-- images/*.png
            |-- metas
        |-- SPAQ
            |-- images/*.jpg
            |-- metas     
        ... 

Evaluation

Run the following code to reproduce the results of our paper. Change the --test_json to the path of your testing json.

bash eval_json.sh

Acknowledgement

We appreciate the releasing codes and data of Q-Insight and DeQA-Score.

Citation

@article{zhao2025reasoning,
  title={Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment},
  author={Zhao, Shijie and Zhang, Xuanyu and Li, Weiqi and Li, Junlin and Zhang, Li and Xue, Tianfan and Zhang, Jian},
  journal={Proceedings of the International Conference on Learning Representations (ICLR)},
  year={2026}
}