Elamur Research⌕
☰
home / papers / 2026 / Generative-to-Discriminative Test-Time Adaptation via Manifold-Aware Diffusion and Bayesian Distillation

Generative-to-Discriminative Test-Time Adaptation via Manifold-Aware Diffusion and Bayesian Distillation

Paper · 2026 · Boyun Zhang, Zequn Xie, Fangming Feng et al.

Summary

Multimodal Sentiment Analysis (MSA) models lose much of their accuracy when test data come from a different domain than the training data. Discriminative Test-Time Adaptation (TTA) methods, such as entropy minimization or pseudo-labeling, tend to reinforce high-confidence outputs that are actually wrong. Diffusion-based adaptation is better calibrated but too slow for online use.

The authors propose GD-Adapt. It uses one shared multimodal Transformer encoder with three components:

  • a conditional diffusion Teacher,
  • a lightweight MLP regression Student,
  • a RealNVP normalizing-flow Source Manifold Estimator (SME).

During pretraining, Auxiliary Generative Regularization adds a denoising objective, and Stochastic Modality Perturbation masks one modality and adds noise to the others. During adaptation, a Dual-Uncertainty Filter discards samples that have low source likelihood or unstable Teacher ensemble outputs. The Student then learns from the Teacher's pseudo-labels through Bayesian Diffusion Distillation.

Across five cross-domain transfers using MOSI, MOSEI and SIMS, GD-Adapt outperforms the baselines. The design matters because fast prediction stays on the Student while the expensive adaptation runs asynchronously.

Key points

  • On MOSI → SIMS, MAE falls from 0.6872 to 0.5673, and Acc-2 rises by 5.81 points to 57.33%. On SIMS → MOSI, the gain over the non-adapted baseline is 11.18 points.
  • Student inference takes 0.04 ms/sample, compared with 6.07 ms/sample for Teacher sampling (T = 30). Only 0.003M parameters are trained during TTA, with peak VRAM of about 4.2 GB.
  • In the ablation, removing the SME (Manifold Prior) hurts most: MAE rises to 0.6897 and Acc-2 drops to 49.23. Removing the consistency loss raises RMSE to 0.8159.
  • The OOD threshold τ_ood works best near the 0.05 quantile of source validation likelihoods.
  • The authors list three limitations: adaptation remains costly, the method is sensitive to the threshold, and adaptation can stall when source and target manifolds barely overlap.

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

🎓Zhejiang UniversityHangzhou · China224 works · 13,587 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksPsychology · 1,649 worksExperimental and Cognitive Psychology · 930 works

Cite

Cite as Boyun Zhang, Zequn Xie, Fangming Feng, Zihan Zhang, Yongbo He, Chuxin Wang, Sihang Cai, Tao Jin and Qifei Zhang. 2026. Generative-to-Discriminative Test-Time Adaptation via Manifold-Aware Diffusion and Bayesian Distillation. 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.350

@inproceedings{zhang2026generative,
  title = {{Generative-to-Discriminative Test-Time Adaptation via Manifold-Aware Diffusion and Bayesian Distillation}},
  author = {Boyun Zhang, Zequn Xie, Fangming Feng, Zihan Zhang, Yongbo He, Chuxin Wang, Sihang Cai, Tao Jin and Qifei Zhang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.350},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.350}
}

References

36 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sameer Ambekar, Zehao Xiao, Jiayi Shen, Xiantong Zhen, and Cees G. M. Snoek. 2024. Probabilistic Test-Time Generalization by Variational Neighbor-Labeling. Preprint, arXiv:2307.04033. — [arXiv:2307.04033](https://arxiv.org/abs/2307.04033) find
  2. Stefano Baccianella, Andrea Esuli, and Fabrizio Sebastiani. 2010. SENTIWORDNET 3.0: An Enhanced Lexical Resource for Sentiment Analysis and Opinion Mining. Proceedings of LREC, 10.
  3. AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. 2018. Multimodal Language Analysis in the Wild: CMU-MOSEI Dataset and Interpretable Dynamic Fusion Graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume…
  4. Dhanajit Brahma and Piyush Rai. 2023. A Probabilistic Framework for Lifelong Test-Time Adaptation. Preprint, arXiv:2212.09713. — [arXiv:2212.09713](https://arxiv.org/abs/2212.09713) find
  5. Fangming Feng, Sihang Cai, Zequn Xie, Yangyang Wu, and Tao Jin. 2026. Scene-aware spatiotemporal generalization: Towards robust temporal action detection across domains. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 3903–3911. find
  6. Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. 2023. Back to the Source: Diffusion-Driven Adaptation to Test-Time Corruption. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786–11796. find
  7. Zirun Guo, Tao Jin, Wenlong Xu, Wang Lin, and Yangyang Wu. 2025. Bridging the Gap for Test-Time Multimodal Sentiment Analysis. arXiv preprint. find
  8. Yongbo He, Zirun Guo, and Tao Jin. 2026. Decoupling stability and plasticity for multi-modal test-time adaptation. arXiv preprint arXiv:2603.00574. — [arXiv:2603.00574](https://arxiv.org/abs/2603.00574) find
  9. Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Preprint, arXiv:2006.11239. find
  10. Judy Hoffman, Trevor Darrell, and Kate Saenko. 2014. Continuous Manifold Based Adaptation for Evolving Visual Domains. pages 867–874. find
  11. S. Kiritchenko, X. Zhu, and S. M. Mohammad. 2014. Sentiment Analysis of Short Informal Texts. Journal of Artificial Intelligence Research, 50:723–762. find
  12. Jae-Hong Lee and Joon-Hyuk Chang. 2024. Stationary Latent Weight Inference for Unreliable Observations from Online Test-Time Adaptation. find
  13. Zezeng Li, Rui Yang, Ruochen Chen, ZhongXuan Luo, and Liming Chen. 2025a. ADPro: A Test-time Adaptive Diffusion Policy via Manifold-constrained Denoising and Task-aware Initialization for Robotic Manipulation. Preprint, arXiv:2508.06266. — [arXiv:2508.06266](https://arxiv.org/abs/2508.06266) find
  14. Zuhe Li, Panbo Liu, Yushan Pan, Weiping Ding, Jun Yu, Haoran Chen, Weihua Liu, Yiming Luo, and Hao Wang. 2025b. Multimodal sentiment analysis based on disentangled representation learning and cross-modal-context association mining. Neurocomputing, 617:128940. find
  15. Louis-Philippe Morency, Rada Mihalcea, and Payal Doshi. 2011. Towards multimodal sentiment analysis: harvesting opinions from the web. In Proceedings of the 13th international conference on multimodal interfaces, pages 169–176, Alicante Spain. ACM. find
  16. Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. 2022. Efficient Test-Time Model Adaptation without Forgetting. Preprint, arXiv:2204.02610. — [arXiv:2204.02610](https://arxiv.org/abs/2204.02610) find
  17. Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. 2023. Towards Stable Test-Time Adaptation in Dynamic Wild World. Preprint, arXiv:2302.12400. — [arXiv:2302.12400](https://arxiv.org/abs/2302.12400) find
  18. Mihir Prabhudesai, Tsung-Wei Ke, Alexander C. Li, Deepak Pathak, and Katerina Fragkiadaki. 2023. Diffusion-TTA: Test-time Adaptation of Discriminative Models via Generative Feedback. Preprint, arXiv:2311.16102. — [arXiv:2311.16102](https://arxiv.org/abs/2311.16102) find
  19. Subhadeep Roy, Shankhanil Mitra, Soma Biswas, and Rajiv Soundararajan. 2023. Test Time Adaptation for Blind Image Quality Assessment. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16696–16705. IEEE. find
  20. Cristiano Saltori, Evgeny Krivosheev, Stéphane Lathuilière, Nicu Sebe, Fabio Galasso, Giuseppe Fiameni, Elisa Ricci, and Fabio Poiesi. 2022. GIPSO: Geometrically Informed Propagation for Online Adaptation in 3D LiDAR Segmentation. Preprint, arXiv:2207.09763. — [arXiv:2207.09763](https://arxiv.org/abs/2207.09763) find
  21. Mohammad Soleymani, David Garcia, Brendan Jou, Björn Schuller, Shih-Fu Chang, and Maja Pantic. 2017. A survey of multimodal sentiment analysis. Image and Vision Computing, 65:3–14. find
  22. Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. 2020. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts. arXiv preprint. find
  23. Jian Liang, Ran He, and Tieniu Tan. 2025. A Comprehensive Survey on Test-Time Adaptation under Distribution Shifts. International Journal of Computer Vision, 133(1):31–64. find
  24. Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. 2021. Tent: Fully Test-time Adaptation by Entropy Minimization. arXiv preprint. find
  25. Jian Liang, Dapeng Hu, and Jiashi Feng. 2021. Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation. arXiv preprint. find
  26. Hongbin Wang, Chun Ren, and Zhengtao Yu. 2025a. Multimodal sentiment analysis based on multiple attention. Engineering Applications of Artificial Intelligence, 140:109731. find
  27. Robert A. Marsden, Mario Döbler, and Bin Yang. 2023. Universal Test-time Adaptation through Weight Ensembling, Diversity Weighting, and Prior Correction. Preprint, arXiv:2306.00650. — [arXiv:2306.00650](https://arxiv.org/abs/2306.00650) find
  28. Zixin Wang, Yadan Luo, Liang Zheng, Zhuoxiao Chen, Sen Wang, and Zi Huang. 2025b. In Search of Lost Online Test-Time Adaptation: A Survey. 133(3):1106–1139. find
  29. Jiana Meng, Yingchun Long, Yuhai Yu, Dandan Zhao, and Shuang Liu. 2019. Cross-Domain Text Sentiment Analysis Based on CNN_ft Method. Information, 10(5):162. find
  30. Zequn Xie, Boyun Zhang, Yuxiao Lin, and Tao Jin. 2026. Delving deeper: Hierarchical visual perception for robust video-text retrieval. arXiv preprint arXiv:2601.12768. — [arXiv:2601.12768](https://arxiv.org/abs/2601.12768) find
  31. Mouxing Yang, Yunfan Li, Changqing Zhang, Peng Hu, and Xi Peng. 2024. Test-time adaptation against multi-modal reliability bias. In International Conference on Representation Learning, volume 2024, pages 55036–55055. find
  32. Wenmeng Yu, Hua Xu, Fanyang Meng, Yilin Zhu, Yixiao Ma, Jiele Wu, Jiyun Zou, and Kaicheng Yang. 2020. CH-SIMS: A Chinese Multimodal Sentiment Analysis Dataset with Fine-grained Annotation of Modality. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3…
  33. Longhui Yuan, Binhui Xie, and Shuang Li. 2023. Robust Test-Time Adaptation in Dynamic Scenarios. Preprint, arXiv:2303.13899. — [arXiv:2303.13899](https://arxiv.org/abs/2303.13899) find
  34. Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. 2016. Multimodal sentiment intensity analysis in videos: Facial gestures and verbal messages. IEEE Intelligent Systems, 31(6):82–88. find
  35. Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. 2023. Domain Generalization: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396–4415. find
  36. Lihua Zhou, Mao Ye, Shuaifeng Li, Nianxin Li, Jinlin Wu, Xiatian Zhu, Lei Deng, Hongbin Liu, Jiebo Luo, and Zhen Lei. 2025. Bayesian Test-time Adaptation for Object Recognition and Detection with Vision-language Models. Preprint, arXiv:2510.02750. — [arXiv:2510.02750](https://arxiv.org/abs/2510.02750) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-generative-discriminative-test-time-adaptation/