Zero-Shot Learning’s Latest Leap: From Medical Diagnostics to Multi-Robot Control

Latest 50 papers on zero-shot learning: Dec. 21, 2025

Zero-shot learning (ZSL) has long been a holy grail in AI/ML, promising models that can recognize, classify, and even act upon concepts they’ve never explicitly been trained on. Imagine an AI that can identify a rare disease from a single description or control a swarm of robots based on a natural language command—all without a single labeled example. This is the promise of ZSL, and recent research is bringing us closer to this reality. From enhancing industrial automation to revolutionizing healthcare, these breakthroughs are reshaping how we build intelligent systems capable of generalizing in novel situations.

The Big Idea(s) & Core Innovations

The overarching theme uniting this wave of research is the creative integration of powerful foundational models—like Vision-Language Models (VLMs) and Large Language Models (LLMs)—with novel architectures and data strategies to unlock unprecedented generalization capabilities. Many papers tackle the compositional aspect of zero-shot learning, where models must understand novel combinations of known attributes and objects. For instance, CAMS: Towards Compositional Zero-Shot Learning via Gated Cross-Attention and Multi-Space Disentanglement by authors from the State Key Laboratory of Public Big Data, Guizhou University, proposes using Gated Cross-Attention and Multi-Space Disentanglement to robustly disentangle attribute and object semantics, outperforming existing CLIP-based methods. Similarly, Learning by Imagining: Debiased Feature Augmentation for Compositional Zero-Shot Learning from Zhejiang University, Shanghai Innovation Institute draws inspiration from neuroscience to synthesize high-fidelity compositional features, enabling models to imagine unseen combinations.

Another significant innovation lies in making ZSL practical for label-scarce domains and real-world applications. In medical imaging, Zero-Training Task-Specific Model Synthesis for Few-Shot Medical Image Classification introduces ZS-TMS, a groundbreaking paradigm by Beijing 1st BioTech Group Co., Ltd. that synthesizes entire task-specific classifier parameters from minimal multimodal inputs (a single image and text description!), bypassing traditional training entirely. This is a game-changer for rare disease diagnostics. Likewise, Bridged Semantic Alignment for Zero-shot 3D Medical Image Diagnosis from the University of Science and Technology of China significantly improves 3D medical image diagnosis by bridging vision-language interactions without requiring labeled data, a crucial step for clinical deployment. In industrial settings, Hybrid Synthetic Data Generation with Domain Randomization Enables Zero-Shot Vision-Based Part Inspection Under Extreme Class Imbalance by researchers from University of Michigan and General Motors leverages synthetic data and domain randomization to achieve 90-91% balanced accuracy in part inspection, even under severe class imbalance, eliminating manual annotation.

Beyond perception, ZSL is making inroads into complex reasoning and control. GenSwarm: Scalable Multi-Robot Code-Policy Generation and Deployment via Language Models from Westlake University and Beihang University uses LLMs to generate and deploy control policies for multi-robot systems directly from natural language instructions, demonstrating high success rates in tasks like encircling and flocking. For event scheduling, CoS: Towards Optimal Event Scheduling via Chain-of-Scheduling from Sun Yat-sen University uses Knowledge Distillation to imbue LLMs with scheduling intelligence, enabling efficient and interpretable zero-shot event planning.

Under the Hood: Models, Datasets, & Benchmarks

The innovations highlighted above are often powered by advancements in model architectures, novel datasets, and rigorous benchmarking:

• Vision-Language Models (VLMs) & Large Language Models (LLMs): These foundational models are central. Many papers, such as Decoupling Augmentation Bias in Prompt Learning for Vision-Language Models by Korea Institute of Energy Technology (KENTECH) and Prompt-Based Continual Compositional Zero-Shot Learning from Information Technology University, focus on refining prompt engineering and model adaptation strategies to leverage the inherent semantic understanding of models like CLIP and GPT-4. HiCoTraj: Zero-Shot Demographic Reasoning via Hierarchical Chain-of-Thought Prompting from Trajectory uses LLMs for interpretable demographic analysis from trajectory data.
• Specialized Architectures: Beyond generic VLMs, new architectures are tailored for ZSL. A Single Architecture for Representing Invariance Under Any Space Group from Princeton University introduces symmetry-adapted Fourier bases for zero-shot generalization in materials science. Fine-Grained Zero-Shot Learning with Attribute-Centric Representations from the University of Southern Queensland employs a Mixture of Patch Experts and Mixture of Attribute Experts for attribute-wise disentanglement in fine-grained classification.
• Hybrid & Multimodal Approaches: Integrating different modalities (vision, language, genomics, thermal) is key. CLIBD: Bridging Vision and Genomics for Biodiversity Monitoring at Scale by Simon Fraser University aligns images, DNA barcodes, and text labels in a shared embedding space for superior species classification. VLM-IRIS: Vision-Language Models for Infrared Industrial Sensing in Additive Manufacturing Scene Description adapts VLMs to infrared data through preprocessing and prompt engineering for zero-shot object detection in thermal imaging.
• Synthetic Data & Benchmarking: To overcome data scarcity, synthetic data generation is crucial. Hybrid Synthetic Data Generation (https://arxiv.org/pdf/2512.00125) is a prime example. For compositional ZSL, new benchmarks like CZSFood-90 and CZSFood-164 are introduced in SalientFusion: Context-Aware Compositional Zero-Shot Food Recognition (https://arxiv.org/pdf/2509.03873). Benchmarks like ZPD-SCA (https://arxiv.org/pdf/2508.14377) specifically evaluate LLMs’ cognitive abilities. Many papers provide public code repositories, such as https://github.com/huggingface/transformers, https://github.com/ITU-Research/PromptCCZSL, and https://github.com/NingWang2049/Mg-MRN, encouraging further exploration.

Impact & The Road Ahead

The implications of these advancements are profound. Zero-shot learning is moving beyond a theoretical curiosity to become a practical tool for addressing real-world challenges where labeled data is scarce, expensive, or impossible to obtain. From highly specialized domains like lattice gauge theory (Matrix-free Neural Preconditioner for the Dirac Operator in Lattice Gauge Theory [https://arxiv.org/pdf/2509.10378]) to pervasive applications like recommendation systems (EasyRec: Simple yet Effective Language Models for Recommendation [https://arxiv.org/pdf/2408.08821]), ZSL promises to accelerate AI adoption and reduce development costs.

The future will likely see continued convergence of large generative models with specialized architectures, pushing the boundaries of generalization. We can expect more robust multimodal models, further advancements in compositional generalization to handle even more complex real-world scenarios (as explored in Compositional Zero-Shot Learning: A Survey [https://arxiv.org/pdf/2510.11106]), and the development of frameworks that can truly learn continually without forgetting previous knowledge (Prompt-Based Continual Compositional Zero-Shot Learning [https://arxiv.org/pdf/2512.09172]). The ability to perform zero-shot inferences in decentralized federated learning environments (Zero-Shot Decentralized Federated Learning [https://arxiv.org/pdf/2509.26462]) also promises enhanced privacy and scalability. The potential to predict and prevent fake content before it's created (Zero-Shot Visual Deepfake Detection [https://arxiv.org/pdf/2509.18461]) even hints at a proactive AI future. These research efforts are collectively paving the way for AI systems that are not only intelligent but also adaptable, efficient, and capable of operating in a truly open world.