Elamur Research⌕
☰
home / papers / 2026 / From Mimesis to Metamorphosis: Evolving VLM Judges via In-Context Comparing and Knowledge Internalization

From Mimesis to Metamorphosis: Evolving VLM Judges via In-Context Comparing and Knowledge Internalization

Paper · 2026 · Juntuo Wang, Yuming Qiao, Yifan Yang et al.

Summary

VLM-based judges that output absolute scores for subjective qualities such as aesthetics tend to be unreliable. Rating scales differ across datasets, and models absorb preference biases. The authors propose S2AD (Semantic-Anchored Scale-Agnostic Distillation), which treats scoring as placing an image relative to others rather than as regression. It trains in two steps. In Stage 1 (Mimesis), a teacher VLM sees a query alongside four retrieved, scored reference images and learns where the query falls among them, using Dynamic Soft Positioning (DSP). Dynamic Rescale applies random per-instance affine changes to the scores, so the model cannot memorise one numeric convention. In Stage 2 (Metamorphosis), a student absorbs this comparative skill through Language Buttons, five discrete semantic levels from Poor to Excellent. A Standard Table maps these levels to scores, and a GRPO-optimised Fidelity Reward aligns the student with the teacher. At inference the student needs no retrieved references. It also adapts to a new dataset's scale by re-anchoring the table, with no parameter updates. The authors report state-of-the-art results on aesthetics benchmarks and zero-shot gains on IQA.

Key points

  • Evaluation covers aesthetics datasets PARA (in-domain), AVA, AADB, CADB and TAD66K (out-of-distribution), plus zero-shot IQA on KADID-10k and SPAQ.
  • The ablations show that both stages matter:
    • Removing retrieved references from the Stage 1 teacher drops average SRCC from 0.5943 to 0.5268.
    • Training Stage 2 without the teacher and the Fidelity Reward gives only 0.5218.
  • Stage 1 performance depends on the number of reference images. N=2 is too sparse, N=4 and N=6 perform comparably, and N=8 brings no further gain. N=4 was chosen to limit VRAM use.
  • On IQA, the retrieval-free Stage 2 model beats the general Qwen2.5VL-7B baseline:
    • KADID-10k SRCC: 0.6482 vs 0.5601.
    • SPAQ SRCC: 0.8859 vs 0.8595.
    • IQA-specific models trained on these datasets (VisualQuality-R1, Q-Insight) still score higher on KADID-10k.
  • Code is released at github.com/SpatialVision-Research/SSAD_ACL2026_Findings.
  • The authors list three limitations:
    • The method is so far tested only on aesthetics and quality.
    • It uses a fixed five-level anchor granularity.
    • Biases in the retrieval step during Stage 1 can affect the teacher.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Brown UniversityProvidence · United States76 works · 25,952 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksInformation Systems · 1,437 works

Cite

Cite as Juntuo Wang, Yuming Qiao, Yifan Yang, Lunxi Yuan, Liang Luo and Dan Meng. 2026. From Mimesis to Metamorphosis: Evolving VLM Judges via In-Context Comparing and Knowledge Internalization. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.631

@inproceedings{wang2026mimesis,
  title = {{From Mimesis to Metamorphosis: Evolving VLM Judges via In-Context Comparing and Knowledge Internalization}},
  author = {Juntuo Wang, Yuming Qiao, Yifan Yang, Lunxi Yuan, Liang Luo and Dan Meng},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.631},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.631}
}

References

33 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. Preprint, arXiv:2502.13…
  2. Fatemeh Behrad, Tinne Tuytelaars, and Johan Wagemans. 2025. Charm: The missing piece in vit fine-tuning for image aesthetic assessment. Preprint, arXiv:2504.02522. — [arXiv:2504.02522](https://arxiv.org/abs/2504.02522) find
  3. Ralph Allan Bradley and Milton E Terry. 1952. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4). find
  4. Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. 2005. Learning to rank using gradient descent. In Proceedings of the 22nd International Conference on Machine Learning, ICML ’05, page 89–96, New York, NY, USA. Association for Computing Machinery.
  5. Shuo Cao, Nan Ma, Jiayang Li, Xiaohui Li, Lihao Shao, Kaiwen Zhu, Yu Zhou, Yuandong Pu, Jiarui Wu, Jiaquan Wang, Bo Qu, Wenhai Wang, Yu Qiao, Dajuin Yao, and Yihao Liu. 2025. Artimuse: Fine-grained image aesthetics assessment with joint scoring and expert-level understanding. Preprint, arXiv:2507.14… — [arXiv:2507.14533](https://arxiv.org/abs/2507.14533) find
  6. Dongping Chen, Ruoxi Chen, Shilin Zhang, Yaochen Wang, Yinuo Liu, Huichi Zhou, Qihui Zhang, Yao Wan, Pan Zhou, and Lichao Sun. 2024. Mllm-asa-judge: assessing multimodal llm-as-a-judge with vision-language benchmark. In Proceedings of the 41st International Conference on Machine Learning, ICML’24. J… find
  7. Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2023. Deep reinforcement learning from human preferences. Preprint, arXiv:1706.03741. find
  8. Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. 2020. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). find
  9. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, and 175 others. 2025. Deepseek-r1 incentivizes reasoning in llms through rein… find
  10. Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. 2022. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 942–948. International Joint Conferences… find
  11. Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. 2016. Photo aesthetics ranking network with attributes and content adaptation. In Computer Vision – ECCV 2016, pages 662–679, Cham. Springer International Publishing. find
  12. Seongyun Lee, Seungone Kim, Sue Hyun Park, Geewook Kim, and Minjoon Seo. 2024. Prometheus-vision: Vision-language model as a judge for fine-grained evaluation. Preprint, arXiv:2401.06591. find
  13. Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. 2024. Llava-onevision: Easy visual task transfer. Preprint, arXiv:2408.03326.
  14. Weiqi Li, Xuanyu Zhang, Shijie Zhao, Yabin Zhang, Junlin Li, Li Zhang, and Jian Zhang. 2025. Q-insight: Understanding image quality via visual reinforcement learning. Preprint, arXiv:2503.22679. — [arXiv:2503.22679](https://arxiv.org/abs/2503.22679) find
  15. Hanhe Lin, Vlad Hosu, and Dietmar Saupe. 2019. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1–3. find
  16. Boyang Liu, Yifan Hu, Senjie Jin, Shihan Dou, Gonglei Shi, Jie Shao, Tao Gui, and Xuanjing Huang. 2025. Unlocking the essence of beauty: Advanced aesthetic reasoning with relative-absolute policy optimization. Preprint, arXiv:2509.21871. — [arXiv:2509.21871](https://arxiv.org/abs/2509.21871) find
  17. R. Duncan. Luce. 2012. Individual choice behavior : A theoretical analysis. find
  18. Naila Murray, Luca Marchesotti, and Florent Perronnin. 2012. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408–2415. find
  19. OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander M adry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024. Gpt-4o system card.…
  20. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Train…
  21. Devi Parikh and Kristen Grauman. 2011. Relative attributes. In 2011 International Conference on Computer Vision, pages 503–510. find
  22. R. L. Plackett. 1975. The analysis of permutations. Journal of the Royal Statistical Society. Series C (Applied Statistics), 24(2):193–202. find
  23. Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Preprint, arXiv:2305.18290.
  24. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  25. L. L. Thurstone. 1927. A law of comparative judgment. Psychological Review, 34(4):273–286. find
  26. Tianhe Wu, Jian Zou, Jie Liang, Lei Zhang, and Kede Ma. 2025. Visualquality-r1: Reasoning-induced image quality assessment via reinforcement learning to rank. Preprint, arXiv:2505.14460. — [arXiv:2505.14460](https://arxiv.org/abs/2505.14460) find
  27. Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. 2023. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. Preprint, arXiv:2306.09341. — [arXiv:2306.09341](https://arxiv.org/abs/2306.09341) find
  28. Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. 2023. Imagereward: Learning and evaluating human preferences for text-to-image generation. Preprint, arXiv:2304.05977. find
  29. Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. 2022. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19861–19869. find
  30. Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. 2025. Teaching large language models to regress accurate image quality scores using score distribution. Preprint, arXiv:2501.11561. — [arXiv:2501.11561](https://arxiv.org/abs/2501.11561) find
  31. Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. 2024. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. Preprint, arXiv:2312.08962. — [arXiv:2312.08962](https://arxiv.org/abs/2312.08962) find
  32. Bo Zhang, Li Niu, and Liqing Zhang. 2021. Image composition assessment with saliency-augmented multi-pattern pooling. In British Machine Vision Conference. find
  33. Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, and 32 others. 2025. Internvl3: Exploring advanced training and test-…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-mimesis-metamorphosis-evolving-vlm-judges/