Few-Shot Learning: Scaling Intelligence with Minimal Data

Latest 50 papers on few-shot learning: Nov. 23, 2025

Few-shot learning (FSL) is rapidly transforming the landscape of AI/ML, offering a compelling solution to the perennial challenge of data scarcity. In a world where collecting vast, labeled datasets can be prohibitively expensive or even impossible, FSL allows models to learn new concepts from just a handful of examples. This capability is pivotal for deploying AI in niche domains, personalized applications, and dynamic real-world environments. Recent breakthroughs, as showcased in a collection of cutting-edge research, are pushing the boundaries of what’s possible, from enhancing robot dexterity to securing IoT devices and even uncovering hidden patterns in medical data.

The Big Idea(s) & Core Innovations

At the heart of these advancements lies a common quest: to imbue AI with the ability to generalize robustly from limited information, mimicking human-like learning efficiency. One significant thread is the integration of diverse data sources and advanced representation learning. For instance, in “Supervised Contrastive Learning for Few-Shot AI-Generated Image Detection and Attribution”, researchers from Universidad Politécnica de Madrid (UPM) leverage supervised contrastive learning to significantly improve the detection and attribution of AI-generated images with only 150 images per class. This highlights how effective feature extraction can lead to strong generalization. Similarly, in “FreqGRL: Suppressing Low-Frequency Bias and Mining High-Frequency Knowledge for Cross-Domain Few-Shot Learning”, a collaborative effort involving institutions like Xi’an Jiaotong University introduces a frequency-space analysis to mitigate bias from low-frequency source data, enhancing generalization in cross-domain FSL.

Another major thrust is the synergy between FSL and Large Language Models (LLMs). The paper “LAUD: Integrating Large Language Models with Active Learning for Unlabeled Data” by Tzu-Hsuan Chou and Chun-Nan Chou from CMoney Technology Corporation addresses the ‘cold-start problem’ by combining LLMs with active learning to efficiently derive task-specific models, outperforming traditional zero-shot and few-shot baselines. This theme is echoed in “C²P: Featuring Large Language Models with Causal Reasoning”, where researchers including Abdolmahdi Bagheri from UC Irvine introduce a Causal Chain of Prompting framework, enabling LLMs to perform causal reasoning with as few as ten examples, demonstrating over 20% improvement in few-shot settings. This integration is crucial for complex tasks like travel satisfaction analysis, where “Applying Large Language Models to Travel Satisfaction Analysis” by Pengfei Xu and Donggen Wang from Hong Kong Baptist University uses few-shot learning to align LLMs with human behavioral patterns, addressing ‘behavioral misalignment’.

Robotics and real-time systems also see significant FSL gains. In “In-N-On: Scaling Egocentric Manipulation with in-the-wild and on-task Data”, UC San Diego and Apple Vision Pro researchers combine diverse human data for humanoid robot manipulation, enabling Human0 to achieve language following and few-shot learning capabilities. For multi-robot systems, “Few-Shot Demonstration-Driven Task Coordination and Trajectory Execution for Multi-Robot Systems” from the University of Robotics and AI (URAI) utilizes imitation learning to allow robots to acquire complex behaviors from minimal human demonstrations. This focus on efficiency and adaptability is further highlighted in “Real-time Continual Learning on Intel Loihi 2” by Intel Labs, which introduces CLP-SNN, a spiking neural network for real-time continual learning that is 70x faster and 5,600x more energy-efficient than edge GPUs, leveraging few-shot learning principles.

Under the Hood: Models, Datasets, & Benchmarks

The progress in few-shot learning is heavily reliant on novel architectures, specialized datasets, and rigorous benchmarking, driving both innovation and practical utility:

• ATTBHFA-Net: Proposed in “Enhancing Few-Shot Classification of Benchmark and Disaster Imagery with ATTBHFA-Net”, this model combines spatial-channel attention with Bhattacharyya-Hellinger distances for robust prototype formation, demonstrating superior performance on disaster-specific datasets. Code available at https://github.com/GreedYLearner1146/ABHFA-Net.
• Logos Dataset: “Logos as a Well-Tempered Pre-train for Sign Language Recognition” introduces the largest Russian Sign Language dataset, crucial for cross-language transfer learning and improving accuracy in low-resource sign language recognition.
• MedIMeta Dataset: From the University of Tübingen, Germany, “A comprehensive and easy-to-use multi-domain multi-task medical imaging meta-dataset” offers 19 datasets across 10 domains with 54 tasks, standardizing data for cross-domain few-shot learning (CD-FSL) in medical imaging. Code available at https://github.com/StefanoWoerner/medimeta-pytorch.
• QCircuitBench: “QCircuitBench: A Large-Scale Dataset for Benchmarking Quantum Algorithm Design” from Peking University introduces the first large-scale dataset to evaluate AI’s ability to design quantum algorithms, providing 120,290 data points for LLM benchmarking. Code available at https://github.com/EstelYang/QCircuitBench.
• Symmetria Dataset: Introduced in “Symmetria: A Synthetic Dataset for Learning in Point Clouds” by University of Chile and CNR, this formula-driven dataset allows scalable, data-efficient learning of symmetries in 3D point clouds, supporting self-supervised pre-training and symmetry detection.
• ProtoTopic: The framework from “ProtoTopic: Prototypical Network for Few-Shot Medical Topic Modeling” leverages prototypical networks for topic modeling in medical texts, particularly useful where labeled data is scarce. Code available at https://github.com/ProtoTopic-Team/ProtoTopic.
• GEMMA-SQL: “GEMMA-SQL: A Novel Text-to-SQL Model Based on Large Language Models” from Google Research introduces a lightweight, open-source text-to-SQL model built on Gemma 2B, demonstrating competitive performance on the SPIDER benchmark with fewer resources.
• Strada-LLM: “Strada-LLM: Graph LLM for traffic prediction” is a graph-aware LLM for spatio-temporal traffic prediction, integrating graph structures for efficient urban forecasting and improved long-term accuracy, showcasing robust domain adaptation even under few-shot constraints.

Impact & The Road Ahead

These collective efforts underscore a powerful trend: few-shot learning is no longer a niche research area but a fundamental paradigm shift enabling AI to tackle real-world problems with unparalleled efficiency and adaptability. From mitigating data poisoning attacks in wearable IoT systems as demonstrated in “Adaptive and Robust Data Poisoning Detection and Sanitization in Wearable IoT Systems using Large Language Models” to enhancing early Alzheimer’s disease detection with big data and ensemble FSL in “Enhancing Early Alzheimer Disease Detection through Big Data and Ensemble Few-Shot Learning”, the implications are far-reaching. The ability to quickly adapt models to new tasks, domains, and data distributions with minimal examples promises to democratize AI development, making sophisticated models accessible even to low-resource languages like Hausa for sexism detection, as explored in “Dataset Creation and Baseline Models for Sexism Detection in Hausa”.

The road ahead involves refining generalization capabilities, addressing subtle biases (as highlighted in “Mind the Gap: Bridging Prior Shift in Realistic Few-Shot Crop-Type Classification”), and scaling these methods to even more complex, dynamic environments. The integration of causal reasoning into LLMs, the creation of sophisticated meta-datasets for diverse tasks, and the development of energy-efficient neuromorphic hardware will continue to drive this field forward. Few-shot learning is truly the key to unlocking AI’s potential in a data-constrained world, building more robust, adaptive, and intelligent systems that can learn and evolve with us.