Semi-Supervised Learning Unleashed: Bridging Data Gaps Across LLMs, Medicine, and Beyond

Latest 50 papers on semi-supervised learning: Dec. 27, 2025

Semi-supervised learning (SSL) is rapidly becoming the AI/ML community’s go-to strategy for tackling the perennial challenge of data scarcity. As the demand for sophisticated AI models continues to skyrocket across diverse fields, the cost and effort of acquiring vast, high-quality labeled datasets present a significant bottleneck. This collection of recent research papers paints a vivid picture of SSL’s transformative power, showcasing groundbreaking advancements that reduce annotation dependency while enhancing model performance and interpretability across language models, medical imaging, remote sensing, and even fundamental physics.

The Big Idea(s) & Core Innovations

At its heart, the recent surge in SSL innovation centers on maximizing the utility of abundant unlabeled data to compensate for scarce labeled examples. A prominent theme is the strategic generation and refinement of pseudo-labels, often augmented by clever architectural designs and domain-specific insights. For instance, in “Semi-Supervised Learning for Large Language Models Safety and Content Moderation” by Eduard S., tefan Dinut, Iustin Sîrbu, and Traian Rebedea (National University of Science and Technology Politehnica Bucharest, Renius Technologies, NVIDIA), SSL emerges as a promising alternative to costly supervised approaches for LLM safety, with task-specific augmentations outperforming traditional methods. This echoes the broader insight that nuanced data augmentation is crucial.

Medical imaging sees a robust push towards label efficiency. “Semi-Supervised 3D Segmentation for Type-B Aortic Dissection with Slim UNETR” by Denis Mikhailapov and Vladimir Berikov (Sobolev Institute of Mathematics SB RAS, Novosibirsk, Russia) demonstrates improved 3D segmentation accuracy through data augmentation and pseudo-labeling with an Exponential Moving Average (EMA) model. Similarly, “SSL-MedSAM2: A Semi-supervised Medical Image Segmentation Framework Powered by Few-shot Learning of SAM2” by Z. Gong and X. Chen (University of Nottingham) leverages the Segment Anything Model 2 (SAM2)’s few-shot capabilities for high-quality pseudo-label generation. Extending this, “SAM-Fed: SAM-Guided Federated Semi-Supervised Learning for Medical Image Segmentation” by Sahar Nasirihaghighi et al. integrates SAM for guidance in federated learning setups, supporting both homogeneous and heterogeneous aggregation.

Beyond pseudo-labeling, researchers are exploring deeper theoretical and architectural innovations. “In-Context Semi-Supervised Learning” by Jiashuo Fan et al. (Duke University) introduces a two-stage Transformer architecture that uses unlabeled data for better generalization in low-label regimes, learning geometry-aware computations. “LLM-Guided Probabilistic Fusion for Label-Efficient Document Layout Analysis” by Ibne Farabi Shihab et al. (Iowa State University) ingeniously combines visual predictions with structural priors from text-pretrained LLMs, achieving high accuracy with only 5% of labels.

Another significant innovation is addressing biases and structural challenges. “Exploring Topological Bias in Heterogeneous Graph Neural Networks” by Zhiyuan Hu and Yilin Zhang (Tsinghua University) proposes HTAD, a novel graph contrastive learning method for debiasing under semi-supervised scenarios. “Sampling Control for Imbalanced Calibration in Semi-Supervised Learning” by Senmao Tian et al. (Beijing Jiaotong University) introduces SC-SSL, a framework to mitigate class imbalance through decoupled sampling control and post-hoc calibration, achieving state-of-the-art results.

The theoretical underpinnings are also being strengthened. “Informative missingness and its implications in semi-supervised learning” by Jinran Wu et al. (The University of Queensland) profoundly shows that partially labeled data can outperform completely labeled samples when the missingness mechanism is correctly specified, revealing that missing labels themselves carry structural information. Furthermore, “Laplace Learning in Wasserstein Space” by Mary Chriselda Antony Oliver et al. (University of Cambridge, University of Warwick) extends classical graph-based SSL to infinite dimensions using Wasserstein space, providing a rigorous theoretical foundation.

Under the Hood: Models, Datasets, & Benchmarks

The recent advancements in SSL are underpinned by sophisticated models, novel datasets, and rigorous benchmarking, pushing the boundaries of what’s possible with limited labels.

• Foundational Models: The Segment Anything Model (SAM) and its successor SAM2 are proving indispensable in medical imaging, as seen in SSL-MedSAM2 and SAM-Fed, where their zero-shot and few-shot capabilities are leveraged for robust pseudo-label generation. Protein Language Models (PLMs) like ESM-2 and ProtVec (Mitigating the Antigenic Data Bottleneck [https://arxiv.org/pdf/2512.05222]) are transforming biomedical informatics by improving influenza A surveillance with limited antigenic data.
• Architectural Innovations: Transformers are being adapted for in-context SSL, learning geometry-aware computations (In-Context Semi-Supervised Learning). Specialized designs like MCVI-SANet (MCVI-SANet [https://arxiv.org/pdf/2512.18344]) incorporate a Vegetation Indices Saturation Aware Block (VI-SABlock) to handle VI saturation in remote sensing for agriculture. Dual-teacher frameworks, as exemplified by DualFete (https://arxiv.org/pdf/2511.09319) and Dual Teacher-Student Learning (https://arxiv.org/pdf/2505.11018), are refining pseudo-label generation and mitigating confirmation bias in medical segmentation. Graph Attention Networks (GATs) are at the forefront of DialogGraph-LLM (https://arxiv.org/pdf/2511.11000) for intent recognition and Exploring Topological Bias in Heterogeneous Graph Neural Networks (https://arxiv.com/pdf/2512.11846) for debiasing.
• Novel Datasets & Benchmarks: The ImageTBAD dataset (Semi-Supervised 3D Segmentation for Type-B Aortic Dissection) is enabling advanced 3D segmentation. For liver fibrosis quantification, the LiQA dataset (Liver Fibrosis Quantification and Analysis [arxiv.org/abs/2512.07651]) provides multi-phase, multi-center MRI scans for robust model testing. In digital dentistry, the MICCAI STS 2024 Challenge (https://arxiv.org/pdf/2511.22911) and MICCAI STSR 2025 Challenge (https://arxiv.org/pdf/2512.02867) introduce new public datasets and benchmarks for tooth and pulp segmentation. The BraTS 2019 dataset (Modality-Specific Enhancement and Complementary Fusion [https://arxiv.org/pdf/2512.09801]) remains a critical resource for brain tumor segmentation.
• Code Repositories: Several projects open their code for the community: SSL-for-LLMs [https://github.com/LLM-Safety-Research/SSL-for-LLMs], MCVI-SANet [https://github.com/ZhihengZhang/MCVI-SANet], HTAD [https://github.com/HTAD-Project/HTAD], SSL-MedSAM2 [https://github.com/naisops/SSL-MedSAM2/], GLL [https://github.com/jwcalder/GraphLearningLayer], RS-SSAL [https://github.com/zhu-xlab/RS-SSAL], HSMix [https://github.com/DanielaPlusPlus/HSMix], UniHOI [https://github.com/xjtu-ai/UniHOI], ST-ProC [https://github.com/ST-ProC], TSE-Net [https://github.com/zhu-xlab/tse-net], Semi-Supervised Multi-Task Learning for Interpretable Quality Assessment [https://github.com/ltelesco/Semi-Supervised-Multi-Task-Learning-for-Interpretable-Quality-Assessment-of-Fundus-Images], CalibrateMix [https://github.com/mehrab-mustafy/CalibrateMix], SmartHDR [https://github.com/JW20211/SmartHDR], CITADEL [https://github.com/IQSeC-Lab/CITADEL.git], DialogGraph-LLM [https://github.com/david188888/DialogGraph-LLM], flics-2025 [https://github.com/dschles70/flics-2025], AnomalyAID [https://github.com/M-Code-Space/AnomalyAID].

Impact & The Road Ahead

The implications of these SSL advancements are profound and far-reaching. Across domains, the ability to achieve high performance with significantly less labeled data promises to democratize AI, making advanced models accessible even in resource-constrained environments. For example, in healthcare, semi-supervised methods are accelerating diagnoses for conditions like Type-B Aortic Dissection (Semi-Supervised 3D Segmentation for Type-B Aortic Dissection), Alzheimer’s disease (Alzheimer's Disease Brain Network Mining [https://arxiv.org/pdf/2512.17276]), and lung nodule malignancy (LMLCC-Net [https://arxiv.org/pdf/2505.06370]), while also enhancing dental imaging (MICCAI STS 2024 Challenge, MICCAI STSR 2025 Challenge) and retinal quality assessment (Semi-Supervised Multi-Task Learning for Interpretable Quality Assessment of Fundus Images). These advancements mean faster, more cost-effective, and more accessible medical AI, ultimately leading to better patient outcomes.

In autonomous driving, SSL is tackling the LiDAR segmentation bottleneck (Multi-Modal Data-Efficient 3D Scene Understanding for Autonomous Driving [https://arxiv.org/pdf/2405.05258]), paving the way for scalable perception systems. For IoT networks, robust anomaly detection (Federated Semi-Supervised and Semi-Asynchronous Learning [https://arxiv.org/pdf/2308.11981] and AnomalyAID [https://arxiv.org/pdf/2411.11293]) is becoming more feasible, securing distributed and privacy-sensitive environments. Even in fusion energy research, physics-informed neural operators with SSL (Physics-informed Neural Operator Learning [https://arxiv.org/pdf/2511.19114]) are enabling rapid analysis of plasma configurations, a critical step towards sustainable energy.

The future of SSL is brimming with potential. The integration of large foundation models (like LLMs and SAM) with SSL frameworks is a particularly exciting avenue, promising models that are not only data-efficient but also possess powerful generalizable knowledge. Further exploration into the theoretical underpinnings, such as the Informative Missingness and Wasserstein Space papers, will yield even more robust and principled SSL algorithms. As these innovations continue to mature, we can anticipate a new era of AI where annotation burden is dramatically reduced, and intelligent systems can adapt and learn more effectively from the vast, unstructured data of the real world.