Zero-Shot Learning’s Next Frontier: From Hyperbolic Geometry to Real-World Robots

Latest 50 papers on zero-shot learning: Dec. 27, 2025

Zero-shot learning (ZSL) is rapidly evolving, enabling AI systems to understand and act on concepts they’ve never explicitly seen during training. This incredible capability is pushing the boundaries of what’s possible in fields from vision and language to robotics and even materials science. Recent research, as highlighted in a collection of cutting-edge papers, reveals a surge in novel techniques that are making ZSL more robust, interpretable, and applicable to complex, real-world challenges.

The Big Idea(s) & Core Innovations

At the heart of these advancements is the quest for models that can truly generalize, often by mimicking human cognitive processes like imagination and logical reasoning. One significant theme is the exploration of alternative embedding spaces. For instance, in their paper, “H^2em: Learning Hierarchical Hyperbolic Embeddings for Compositional Zero-Shot Learning”, researchers from HKUST, Zhejiang University, and ACCESS introduce H2EM, which leverages hyperbolic geometry to better capture the intricate hierarchical structures found in compositional zero-shot learning (CZSL). They argue that hyperbolic spaces are superior to traditional Euclidean ones for modeling large-scale semantic hierarchies, leading to state-of-the-art performance.

CZSL, which deals with recognizing novel combinations of known attributes and objects, is a particularly challenging area. Papers like “CAMS: Towards Compositional Zero-Shot Learning via Gated Cross-Attention and Multi-Space Disentanglement” by researchers from Guizhou University, Shanghai Jiao Tong University, and Nankai University and “Learning by Imagining: Debiased Feature Augmentation for Compositional Zero-Shot Learning” from Zhejiang University and Northwestern Polytechnical University directly address the complexities of disentangling attribute and object semantics. CAMS introduces Gated Cross-Attention and Multi-Space Disentanglement to align semantic features more effectively with prompt representations, while “Learning by Imagining” proposes Debiased Feature Augmentation (DeFA), drawing inspiration from neuroscience to synthesize high-fidelity features for unseen compositions.

Another innovative trend is the integration of large language models (LLMs) to enhance zero-shot capabilities. “LAUD: Integrating Large Language Models with Active Learning for Unlabeled Data” by CMoney Technology Corporation tackles the cold-start problem by using LLMs to construct initial label sets, outperforming traditional few-shot baselines. Similarly, Sun Yat-sen University, Tsinghua University, and Southeast University introduce “CoS: Towards Optimal Event Scheduling via Chain-of-Scheduling” which leverages LLMs and knowledge distillation for efficient and interpretable event scheduling with strong zero-shot generalization. In a remarkable application for multi-robot systems, “GenSwarm: Scalable Multi-Robot Code-Policy Generation and Deployment via Language Models” from Westlake University and others showcases an end-to-end system that generates and deploys robot control policies from natural language instructions, eliminating manual objective function crafting.

Beyond these, the “Zero-Training Task-Specific Model Synthesis for Few-Shot Medical Image Classification” paper by Beijing 1st BioTech Group Co., Ltd. introduces a groundbreaking paradigm: directly synthesizing classifier parameters from multimodal inputs (image and text) using a generative engine, enabling immediate inference without any training—a game-changer for rare diseases. The concept of zero-shot deepfake detection, exploring how AI could predict fake content before its creation, is also gaining traction, as discussed in “Zero-Shot Visual Deepfake Detection: Can AI Predict and Prevent Fake Content Before It’s Created?”.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are supported by advances in models, new datasets, and refined benchmarks:

• Hyperbolic Embeddings: H2EM (H^2em: Learning Hierarchical Hyperbolic Embeddings for Compositional Zero-Shot Learning) utilizes hyperbolic geometry to model complex hierarchies in CZSL, outperforming Euclidean methods.
• Attribute-Centric Disentanglement: The Attribute-Centric Representation (ACR) framework introduced in “Fine-Grained Zero-Shot Learning with Attribute-Centric Representations” by University of Southern Queensland and The Hong Kong Polytechnic University uses Mixture of Patch Experts and Mixture of Attribute Experts for fine-grained ZSL on datasets like CUB, AwA2, and SUN.
• CLIP-based Enhancements: Many papers leverage CLIP’s powerful vision-language capabilities. “Semantic Relation-Enhanced CLIP Adapter for Domain Adaptive Zero-Shot Learning” from East China Normal University and Ocean University of China proposes SRE-CLIP, enhancing CLIP for DAZSL with a semantic relation structure loss. “Prompt-Based Continual Compositional Zero-Shot Learning” from Information Technology University introduces PromptCCZSL, an innovative prompt-based continual learning framework for VLMs, preventing catastrophic forgetting with Cosine Anchor Alignment Loss and a new CCZSL evaluation protocol. The code is available at https://github.com/ITU-Research/PromptCCZSL.
• Synthetic Data Generation: “Hybrid Synthetic Data Generation with Domain Randomization Enables Zero-Shot Vision-Based Part Inspection Under Extreme Class Imbalance” by University of Michigan and General Motors uses a hybrid SDG framework to generate large-scale labeled datasets for industrial inspection, achieving high accuracy with only synthetic training data.
• Multimodal & Cross-Domain Models: “CLIBD: Bridging Vision and Genomics for Biodiversity Monitoring at Scale” from Simon Fraser University and others, introduces CLIBD, which aligns images, DNA barcodes, and text using contrastive learning for zero-shot species classification. Similarly, “VLM-IRIS: Vision-Language Models for Infrared Industrial Sensing in Additive Manufacturing Scene Description” by Michigan Technological University adapts VLMs to infrared data for zero-shot object detection in industrial settings.
• Foundation Models: “UniFault: A Fault Diagnosis Foundation Model from Bearing Data” introduces UniFault, a transformer-based foundation model pretrained on over 6.9 million samples for robust few-shot fault diagnosis. The code is at https://github.com/Miltos-90/Failure_Classification_of_Bearings. In time series, “Benchmarking Time Series Foundation Models for Short-Term Household Electricity Load Forecasting” by Paderborn University highlights that TSFMs can achieve competitive or superior performance in zero-shot forecasting, reducing training effort significantly.
• Zero-Shot Tabular Data: “ZEUS: Zero-shot Embeddings for Unsupervised Separation of Tabular Data” from Jagiellonian University presents ZEUS, a transformer-based model for unsupervised clustering of tabular data without fine-tuning, with code at https://github.com/gmum/zeus.
• Distributed and Decentralized ZSL: “Distributed Zero-Shot Learning for Visual Recognition” and “Zero-Shot Decentralized Federated Learning” by Perceive Lab explore frameworks for enhancing ZSL scalability and privacy in distributed environments. Code for ZeroDFL is at https://github.com/perceivelab/ZeroDFL.
• Software Engineering: VAPU, a multi-agent system for autonomous legacy code modernization, introduced in “VAPU: System for Autonomous Legacy Code Modernization”, shows that LLM-based agents can achieve comparable error rates to OSL/ZSL prompts in complex code updates. Code is available at https://github.com/GPT-Laboratory/.

Impact & The Road Ahead

These advancements demonstrate that zero-shot learning is transitioning from a theoretical aspiration to a practical necessity, enabling AI systems to operate in data-scarce and dynamic environments. The ability to generalize to unseen categories, attributes, or even entire tasks without explicit retraining is profound. This will have immense implications for real-world applications such as:

• Medical Diagnosis: Faster, label-free diagnosis of rare diseases (ZS-TMS, Bridged Semantic Alignment for Zero-shot 3D Medical Image Diagnosis), and automated medical image analysis with clinical report generation (“Intelligent Healthcare Imaging Platform: An VLM-Based Framework for Automated Medical Image Analysis and Clinical Report Generation”, with code at https://github.com/samer-alhamadani/intelligent-healthcare-imaging-platform).
• Industrial Automation: Robust quality control in manufacturing (“Hybrid Synthetic Data Generation with Domain Randomization Enables Zero-Shot Vision-Based Part Inspection Under Extreme Class Imbalance”), and efficient fault diagnosis with minimal data (UniFault).
• Environmental Monitoring: Scalable biodiversity monitoring without extensive manual labeling (CLIBD).
• Smart Cities: Enhanced road network learning for traffic prediction (“Dual-branch Spatial-Temporal Self-supervised Representation for Enhanced Road Network Learning”, code at https://github.com/chaser-gua/DST), and accessible urban navigation for visually impaired individuals (“Floorplan2Guide: LLM-Guided Floorplan Parsing for BLV Indoor Navigation” and “OmniAcc: Personalized Accessibility Assistant Using Generative AI”, code at https://github.com/omniacc-team/omniacc).
• Robotics: Rapid deployment of multi-robot control policies through natural language instructions (GenSwarm).

The road ahead involves further exploring the theoretical underpinnings of generalization, particularly in complex compositional settings, as highlighted by “Compositional Zero-Shot Learning: A Survey” by Information Technology University. Bridging the ‘academic-practical gap’ in areas like plant disease diagnosis with zero-shot CLIP models (“Rethinking Plant Disease Diagnosis: Bridging the Academic-Practical Gap with Vision Transformers and Zero-Shot Learning”) will also be crucial. The focus will be on developing more efficient and interpretable methods for handling novel concepts, reducing dependence on labeled data, and ensuring robustness against adversarial attacks (“Adversarial Robustness in Zero-Shot Learning: An Empirical Study on Class and Concept-Level Vulnerabilities”). The continued integration of LLMs and multimodal data, alongside innovations in embedding spaces, promises an exciting future where AI can learn and adapt with unprecedented agility.