Zero-Shot Learning’s Ascent: Navigating Unseen Worlds from Vision to Physics and Beyond

Latest 42 papers on zero-shot learning: Oct. 12, 2025

Zero-shot learning (ZSL) is rapidly transforming the AI landscape, promising models that can understand and act on concepts they’ve never encountered during training. This ability to generalize to unseen data, often by leveraging semantic information or pre-trained knowledge, is a holy grail in AI. Recent research showcases ZSL’s burgeoning power, pushing boundaries across diverse domains from computer vision and natural language processing to medical imaging, robotics, and even fundamental physics. This digest explores the latest breakthroughs that are making machines truly imaginative and adaptable.

The Big Idea(s) & Core Innovations

The central challenge in ZSL is bridging the gap between seen and unseen classes or compositions. Recent papers tackle this by either enhancing semantic understanding, synthesizing new data, or adapting pre-trained models. For instance, in visual domains, Haozhe Zhang et al. from Zhejiang University, Shanghai Innovation Institute in their paper, “Learning by Imagining: Debiased Feature Augmentation for Compositional Zero-Shot Learning”, propose Debiased Feature Augmentation (DeFA). Inspired by neuroscience, DeFA synthesizes high-fidelity compositional features, enabling models to ‘imagine’ unseen compositions and significantly improve performance in compositional ZSL (CZSL) tasks. Similarly, Jiajun Song and Xiaoou Liu from RUC (Renmin University of China) and Microsoft Research introduce “SalientFusion: Context-Aware Compositional Zero-Shot Food Recognition” to tackle CZSL in food recognition. Their framework combines segmentation and depth detection to focus on relevant features, effectively reducing noise and semantic bias.

Another innovative approach to CZSL comes from Lin Li et al. from Hong Kong University of Science and Technology and Zhejiang University, with their “Compositional Zero-shot Learning via Progressive Language-based Observations”. PLO mimics human cognition by dynamically determining observation order using VLMs and LLMs to interpret image content through graduated descriptions, leading to robust recognition of state-object compositions. Further strengthening visual understanding, Shiyu Zhang et al. from Tianjin University in “Learning Visual Proxy for Compositional Zero-Shot Learning” introduce Visual Proxy Learning and Cross-Modal Joint Learning (CMJL) to bridge modality gaps and enhance fine-grained visual cues in CZSL, achieving state-of-the-art results.

Beyond perception, ZSL is making strides in practical applications. J. J. Herrera-Aranda et al. from the University of Granada and National Institute of Cybersecurity (INCIBE)’s “Semantic-Inductive Attribute Selection for Zero-Shot Learning” demonstrates that selecting relevant semantic attributes can drastically reduce noise and improve generalization, with up to a fourfold improvement on datasets like aPY. This attribute selection strategy can make ZSL more robust. The concept of ZSL is even enhancing model training itself: Haosong Zhang et al. from Fudan University and New York University introduce “Arithmetic-Mean μP for Modern Architectures: A Unified Learning-Rate Scale for CNNs and ResNets”, a unified learning-rate scale (AM-µP) that enables consistent depth scaling and zero-shot transfer of learning rates across diverse CNNs and ResNets, simplifying hyperparameter tuning for complex models.

ZSL is also proving crucial for addressing critical societal issues. Aparna Ananthasubramaniam et al. from the University of Michigan utilize a zero-shot learning framework in “Characterizing Online Activities Contributing to Suicide Mortality among Youth” to model themes of online behavior linked to youth suicide risk, enabling large-scale analysis without extensive manual labeling. This shows ZSL’s potential for proactive interventions in public health.

Under the Hood: Models, Datasets, & Benchmarks

These advancements are powered by innovative models, specialized datasets, and rigorous benchmarks:

• MultiADS (Multi-type Anomaly Detection and Segmentation): Introduced by Ylli Sadikaj et al. from the University of Vienna and Bosch Corporate Research in “MultiADS: Defect-aware Supervision for Multi-type Anomaly Detection and Segmentation in Zero-Shot Learning”, this framework is the first to achieve multi-type anomaly segmentation in ZSL, leveraging a novel Knowledge Base for Anomalies (KBA) to enhance defect-aware text prompts. Code available at https://github.com/boschresearch/MultiADS.
• BrainGFM (Brain Graph Foundation Model): Xinxu Wei et al. from Lehigh University developed BrainGFM in “A Brain Graph Foundation Model: Pre-Training and Prompt-Tuning for Any Atlas and Disorder”, a graph-based foundation model for fMRI data that integrates multiple brain atlases and utilizes graph contrastive learning and masked autoencoders. It was pre-trained on a massive dataset of 25,000 subjects and 400,000 graph samples.
• REAMS (Reasoning Enhanced Algorithm for Maths Solving): Eishkaran Singh et al.’s “REAMS: Reasoning Enhanced Algorithm for Maths Solving” is a language-based solution that leverages ZSL and mathematical reasoning. It integrates program synthesis and logical reasoning, and its code can be found at https://github.com/idrori/mathQ/tree/main/data.
• OmniAcc: This “Personalized Accessibility Assistant Using Generative AI”, developed by Siddhant Karki et al. from Miami University of Ohio, uses GPT-4, satellite imagery, and OpenStreetMap to provide real-time, personalized navigation for wheelchair users, achieving 97.5% crosswalk detection accuracy. Code is available at https://github.com/omniacc-team/omniacc.
• TSFMs (Time Series Foundation Models): Marcel Meyer et al. from Paderborn University benchmarked TSFMs (Chronos, TimesFM, Time-MoE, Sundial) in “Benchmarking Time Series Foundation Models for Short-Term Household Electricity Load Forecasting”, showing they can outperform traditional Transformer models in a zero-shot setting for household electricity load forecasting.
• ZPD-SCA: Wenhan Dong et al. from South China Normal University and Hong Kong University of Science and Technology (Guangzhou) introduce “ZPD-SCA: Unveiling the Blind Spots of LLMs in Assessing Students’ Cognitive Abilities”, a new benchmark annotated by educators to evaluate LLMs’ ability to assess reading difficulty, revealing their limitations in zero-shot scenarios for educational tasks.
• ZeroDFL (Zero-Shot Decentralized Federated Learning): Perceive Lab Team’s “Zero-Shot Decentralized Federated Learning” is an efficient framework for adapting large vision-language models in distributed environments, improving scalability and privacy. Code is available at https://github.com/perceivelab/ZeroDFL.
• BATCLIP: Sarthak Kumar Maharana et al. from The University of Texas at Dallas and MIT-IBM Watson AI Lab propose “BATCLIP: Bimodal Online Test-Time Adaptation for CLIP”, a bimodal online test-time adaptation method for CLIP to improve robustness against image corruptions by jointly adapting visual and text encoders. Code available at https://github.com/sarthaxxxxx/BATCLIP.
• PSRP-CPI: Hongzhi Zhang et al. from Wuhan University introduced “Zero-Shot Learning with Subsequence Reordering Pretraining for Compound-Protein Interaction”, a pre-training method that uses subsequence reordering and length-variable augmentation for robust CPI prediction, even with small datasets. Code available at https://github.com/Hoch/Zhang/DrugDiscovery-DTI/.
• Ag2x2: Ziyin Xiong et al. from University of California, Berkeley developed “Ag2x2: Robust Agent-Agnostic Visual Representations for Zero-Shot Bimanual Manipulation”, a framework enabling zero-shot bimanual manipulation without expert demonstrations. Code available at https://github.com/ultralytics/yolov5.
• FloorSAM: Silentbarber’s “FloorSAM: SAM-Guided Floorplan Reconstruction with Semantic-Geometric Fusion” combines semantic and geometric information for accurate floorplan reconstruction, supporting zero-shot indoor scene understanding. Code available at https://github.com/Silentbarber/FloorSAM.

Impact & The Road Ahead

The impact of these zero-shot learning advancements is profound. They enable AI systems to be more adaptable, requiring less labeled data and making deployment in novel environments more feasible. From Samer Al-Hamadani from University of Baghdad’s “Intelligent Healthcare Imaging Platform: An VLM-Based Framework for Automated Medical Image Analysis and Clinical Report Generation” which uses VLMs for zero-shot medical image analysis, to Jinho Kim et al. from Friedrich-Alexander-Universität Erlangen-Nürnberg’s “Zero-shot self-supervised learning of single breath-hold magnetic resonance cholangiopancreatography (MRCP) reconstruction” which reduces MRI scan times, ZSL is making AI more efficient and accessible in critical domains.

In industrial settings, Ylli Sadikaj et al.’s MultiADS is revolutionizing quality control by enabling pixel-level multi-type anomaly detection without prior training. The Perceive Lab Team’s ZeroDFL paves the way for privacy-preserving, scalable AI in decentralized federated learning. Furthermore, Yixuan Sun et al. from Argonne National Laboratory and Massachusetts Institute of Technology show ZSL’s utility in scientific computing with “Matrix-free Neural Preconditioner for the Dirac Operator in Lattice Gauge Theory”, accelerating complex physics simulations by generalizing across different lattice sizes.

Looking ahead, the research points towards increasingly sophisticated methods for harnessing semantic knowledge and pre-trained models. The integration of advanced prompting strategies, as seen in “Accelerating Conditional Prompt Learning via Masked Image Modeling for Vision-Language Models” by Phuoc-Nguyen Bui et al. from Sungkyunkwan University, and novel curriculum learning techniques like John Doe and Jane Smith from University of Example and Research Institute for AI’s “Prototype-Guided Curriculum Learning for Zero-Shot Learning”, promise even greater generalization capabilities. While challenges remain, particularly in complex relational understanding as highlighted by Beth Pearson et al. from University of Bristol in “Evaluating Compositional Generalisation in VLMs and Diffusion Models”, the rapid pace of innovation suggests a future where AI systems can truly navigate and comprehend the unseen world, pushing us closer to truly intelligent and adaptive machines.