LOGICZSL: Exploring Logic-induced Representation for Compositional Zero-shot Learning

Peng Wu, Xiankai Lu, Hao Hu, Yongqin Xian, Jianbing Shen, Wenguan Wang

Introduction

Compositional Zero-Shot Learning (CZSL) is the challenge of recognizing novel combinations of seen primitives—for example, identifying a "green apple" after seeing only "red apples" and "green cars." While modern vision-language models like CLIP have advanced this field, they often suffer from a "semantic blind spot." They treat labels as independent entities, ignoring hierarchical relationships (e.g., dogs and wolves are both animals), inter-class similarities, and the polysemous nature of attributes (e.g., "young" looks different on a dog versus a tree).

Methodology: The LOGICZSL Framework

To address these limitations, the authors propose LOGICZSL, a framework that bridges neural learning with symbolic reasoning. The approach involves three key stages:

  1. Relational Knowledge Construction: The system queries an LLM (GPT-4) to generate "Primitive Concept Trees." These include Object-induction (grouping objects into categories), Object-extension (extracting distinguishing features like "floppy eared"), and Attribute-extension (providing context-specific synonyms).
  2. Logic-induced Learning: These hierarchical relationships are translated into first-order logic rules (e.g., $\forall x(\text{dog}(x) \Rightarrow \text{animal}(x))$).
  3. Differentiable Logic Losses: To integrate these rules into neural training, the authors use fuzzy logic relaxation. This converts rigid logical constraints into continuous loss functions ($L_{oi}, L_{oe}, L_{ae}$) that regularize the model's visual features.
The LOGICZSL architecture, detailing the integration of LLM-derived Primitive Concept Trees into the training process via logic-induced learning and differentiable losses.

LOGICZSL is designed as a "plug-and-play" module, compatible with existing CLIP-based CZSL architectures like CSP or Troika without requiring structural changes.

Experimental Evaluation

The framework was evaluated on three major benchmarks: CGQA, UT-Zappos50K, and MIT-States. Testing spanned both Close World (valid compositions only) and Open World (requiring filtering of impossible combinations) settings.

Results and Discussion

LOGICZSL consistently achieved state-of-the-art results. On the CGQA dataset, adding LOGICZSL to the Troika base model improved the Area Under the Curve (AUC) from 12.4 to 15.3. Similar gains were observed across other datasets and base models.

Quantitative performance comparison across benchmarks and qualitative examples showing how LOGICZSL improves composition recognition over base models.

Qualitative analysis reveals that logic-induced losses act as a semantic anchor. For instance, while base models might confuse a "winding stream" with a "dry creek," LOGICZSL's logic rules help it distinguish specific visual features associated with "dry" versus "winding," leading to more accurate classifications.

Conclusion and Future Work

By using LLMs as structured knowledge sources rather than just text encoders, LOGICZSL effectively teaches models "common sense" relationships. While powerful, the method currently depends on the accuracy of LLM descriptions and is limited to relatively shallow hierarchies. Future research will focus on scaling to deeper taxonomies and automating tree construction for broader concept coverage.

Summary updated

Abstract

Compositional zero-shot learning (CZSL) aims to recognize unseen attribute-object compositions by learning the primitive concepts (i.e., attribute and object) from the training set. While recent works achieve impressive results in CZSL by leveraging large vision-language models like CLIP, they ignore the rich semantic relationships between primitive concepts and their compositions. In this work, we propose LogiCzsl, a novel logic-induced learning framework to explicitly model the semantic relationships. Our logic-induced learning framework formulates the relational knowledge constructed from large language models as a set of logic rules, and grounds them onto the training data. Our logic-induced losses are complementary to the widely used CZSL losses, therefore can be employed to inject the semantic information into any existing CZSL methods. Extensive experimental results show that our method brings significant performance improvements across diverse datasets (i.e., CGQA, UT-Zappos50K, MIT-States) with strong CLIP-based methods and settings (i.e., Close World, Open World).
BibTeX
@inproceedings{wu2025logiczsl,
  author = {Peng Wu and Xiankai Lu and Hao Hu and Yongqin Xian and Jianbing Shen and Wenguan Wang},
  title = {{LOGICZSL: Exploring Logic-induced Representation for Compositional Zero-shot Learning}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.02821},
}
Figure