Elamur Research⌕
☰
home / papers / 2026 / DMSD: Dual-Modal Semantic Disentanglement for Compositional Zero-Shot Learning

DMSD: Dual-Modal Semantic Disentanglement for Compositional Zero-Shot Learning

Paper · 2026 · Pan Yang, Jing Yang, Ruan Xiao li et al.

Summary

Compositional Zero-Shot Learning (CZSL) asks a model to recognize attribute–object pairs it never saw in training, such as a "blue banana," by recombining primitives learned from seen pairs. Most CLIP-based approaches (CSP, Troika, LOGICZSL) separate attributes from objects mainly through text-side prompt tuning. The authors argue that this leaves visual sub-concepts entangled and makes predictions vulnerable to co-occurrence bias in language priors. Their framework, DMSD, addresses this from both modalities with three components:

  • Contextual Prompt Space: a shared set of contextual prompts injected into both the image and text encoders.
  • Visual Sub-concept Prototypes: learnable attribute and object prototypes (p_a, p_o) on the visual side.
  • Class-Centroid Bridging Module: a module built on CLIP's text encoder that cross-attends to visual features layer by layer, pulling class centroids toward the textual space.

On MIT-States, UT-Zappos and C-GQA, DMSD tops every closed-world metric. It also stays competitive in the harder open-world setting. The work suggests that explicit visual disentanglement complements text-centric prompt learning for compositional generalization.

Key points

  • Closed-world AUC is 24.2 on MIT-States, 46.2 on UT-Zappos and 16.6 on C-GQA. These are relative gains of 3.4%, 0.9% and 8.5% over the runner-up.
  • Open-world AUC is 8.8 on MIT-States, 35.8 on UT-Zappos and 4.0 on C-GQA. DMSD trails LOGICZSL on UT-Zappos but has higher unseen accuracy there, and it beats LOGICZSL on C-GQA.
  • In the ablation, removing the shared contextual space lowers AUC by about 6.1% (MIT-States) and 3.9% (UT-Zappos) in relative terms.
  • Using both p_a and p_o, rather than no visual prototypes, improves AUC by 6.6% and 4.5% on the same two datasets.
  • Code is released at github.com/ybyangjing/DMSD; the paper appears in Findings of ACL 2026.

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

🎓Sichuan UniversityChengdu · China20 works · 1,293 citations🎓Nanjing University of Posts and TelecommunicationsNanjing · China4 works · 22 citations

Also on the byline: Guizhou University.

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksComputer Vision and Pattern Recognition · 3,486 works

Cite

Cite as Pan Yang, Jing Yang, Ruan Xiao li, Yuling Chen, Yuankai Wu, Quan Zhou and Xu Wang. 2026. DMSD: Dual-Modal Semantic Disentanglement for Compositional Zero-Shot Learning. 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.1540

@inproceedings{yang2026dmsd,
  title = {{DMSD: Dual-Modal Semantic Disentanglement for Compositional Zero-Shot Learning}},
  author = {Pan Yang, Jing Yang, Ruan Xiao li, Yuling Chen, Yuankai Wu, Quan Zhou and Xu Wang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1540},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1540}
}

References

1 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yuval Atzmon, Felix Kreuk, Uri Shalit, and Gal Chechik. 2020. A causal view of compositional zero-shot recognition. In Advances in Neural Information Processing Systems, volume 33, pages 1462–1473. Curran Associates, Inc. \#\# Conclusion find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-dmsd-dual-modal-semantic-disentanglement/