Representation Learning’s Grand Tour: From Biology to Robotics and Beyond

Latest 50 papers on representation learning: Dec. 21, 2025

Step right up, AI/ML enthusiasts! We’re embarking on a fascinating journey through the latest breakthroughs in representation learning. This bustling field, the bedrock of modern AI, grapples with the challenge of teaching machines to understand the world by encoding complex data into meaningful, usable forms. From unraveling the mysteries of biological signals to enhancing the autonomy of robots and refining our interaction with vast digital landscapes, recent research is pushing the boundaries of what’s possible.## The Big Idea(s) & Core Innovationscentral theme unifying many of these papers is the pursuit of more robust, efficient, and interpretable representations, often achieved by leveraging domain-specific knowledge or novel architectural designs. One major thrust is in self-supervised learning (SSL), where models learn from vast amounts of unlabeled data, drastically reducing the reliance on costly human annotation. We see this in the computer vision realm with breakthroughs like SARMAE: Masked Autoencoder for SAR Representation Learning by researchers from Beijing Institute of Technology and Wuhan University. SARMAE tackles the unique challenges of Synthetic Aperture Radar (SAR) imagery by introducing Speckle-Aware Representation Enhancement (SARE) and Semantic Anchor Representation Constraint (SARC), making it robust to speckle noise and semantically consistent with optical priors. Similarly, the PSMamba: Progressive Self-supervised Vision Mamba for Plant Disease Recognition from Griffith University and CSIRO uses a dual-student hierarchical distillation Vision Mamba to capture multi-scale lesion patterns in plant leaves, showcasing how specialized architectures can enhance fine-grained detection.vision, SSL is making waves in medical imaging. The team from the University of Kentucky, in their paper Magnification-Aware Distillation (MAD): A Self-Supervised Framework for Unified Representation Learning in Gigapixel Whole-Slide Images, introduces a magnification-aware distillation framework that creates resolution-invariant representations for neuropathology, bridging low and high-magnification details. This is complemented by their Vision Foundry: A System for Training Foundational Vision AI Models (Paper Link), a HIPAA-compliant platform for training and deploying medical imaging foundation models with minimal annotation. For multimodal medical data, the National University of Singapore’s CITab: Unleashing the Power of Image-Tabular Self-Supervised Learning via Breaking Cross-Tabular Barriers (Paper Link) offers a framework for effective cross-tabular knowledge transfer via semantic-aware modeling, addressing data heterogeneity in Alzheimer’s diagnosis.significant innovation lies in incorporating structural and relational biases. The LightTopoGAT: Enhancing Graph Attention Networks with Topological Features for Efficient Graph Classification by Indira Gandhi National Open University (Paper Link) demonstrates that basic topological properties like node degree can significantly boost graph classification performance with minimal overhead. Taking this further, the Topologically-Stabilized Graph Neural Networks: Empirical Robustness Across Domains (Paper Link) from the University of Bonn integrates persistent homology with stability regularization, making GNNs inherently resistant to structural perturbations. Challenging assumptions, FUEL: Feature-Centric Unsupervised Node Representation Learning Without Homophily Assumption (Paper Link) by KAIST tackles non-homophilous graphs by adapting graph convolution based on node features, a crucial step for diverse real-world networks. And in a groundbreaking theoretical advancement, Proof of a perfect platonic representation hypothesis (Paper Link) from MIT and NTT Research shows that SGD training leads to perfectly aligned “Platonic” representations across different deep linear networks, driven by entropic forces.architecture, novel loss functions and learning paradigms are critical. MACL: Multi-Label Adaptive Contrastive Learning Loss for Remote Sensing Image Retrieval (Paper Link) by Sabanci University introduces an adaptive contrastive loss to mitigate semantic overlap and label imbalance in remote sensing. Predictive Sample Assignment for Semantically Coherent Out-of-Distribution Detection (Paper Link) from USTC enhances OOD detection by focusing on semantically coherent outliers, improving robustness in open-world scenarios. Addressing label noise, LaT-IB: Is the Information Bottleneck Robust Enough? Towards Label-Noise Resistant Information Bottleneck Learning (Paper Link) from Beihang University proposes a ‘Minimal-Sufficient-Clean’ criterion and a three-phase training framework for more reliable IB learning.## Under the Hood: Models, Datasets, & Benchmarkswave of innovation is powered by novel models, expanded datasets, and robust evaluation benchmarks:SARMAE (Paper Link): Introduces SAR-1M, the first million-scale SAR dataset with paired optical images, critical for large-scale self-supervised pretraining. Code is not explicitly provided.KineST (Paper Link): A kinematics-guided state space model for accurate and smooth full-body motion tracking from sparse signals. No public code repository mentioned, but a project page is available.MACL (Paper Link): Evaluated on benchmark datasets like DLRSD, ML-AID, and WHDLD. Code available at https://github.com/amna/MACL.SSLF: Sharpness-aware Second-order Latent Factor Model (Paper Link): Demonstrates improvements on industrial HDI datasets like Yelp and MovieLens. No public code provided.In-Context Semi-Supervised Learning (Paper Link): Utilizes a two-stage Transformer design combining spectral feature learning and gradient descent-based inference. No public code provided.Preserving Marker Specificity with Lightweight Channel-Independent Representation Learning (Paper Link): Evaluated on a Hodgkin lymphoma dataset, comparing lightweight channel-independent architectures against deep CNNs. Code available at https://github.com/SimonBon/CIM-S.Persistence: Topological Metric for Unsupervised Embedding Quality Evaluation (Paper Link): Outperforms existing metrics like RankMe and α-ReQ across various domains. Code available at https://anonymous.4open.science/r/topo_metrics-94D6/.FUEL (Paper Link): Demonstrates state-of-the-art performance across 14 benchmarks with varying homophily levels. Code available at https://github.com/kswoo97/unsupervised-non-homophilic.MAD (Paper Link): Introduces MAD-NP, a Vision Transformer foundation model for neuropathology. Code available at https://github.com/mad-np/mad-codebase.SynJAC: Synthetic-data-driven Joint-granular Adaptation and Calibration (Paper Link): Uses synthetic annotation pipelines leveraging LLMs for Key Information Extraction (KIE) from scanned documents. Code references https://github.com/PaddlePaddle/PaddleOCR and https://www.openai.com/chatgpt.ARCADE: Adaptive Robot Control (Paper Link): An adaptive robot control system using Bayesian dynamics learning with online changepoint detection. No public code provided.PSMamba (Paper Link): Leverages Vision Mamba state-space encoders for multi-scale plant disease recognition. No public code provided.ProtoFlow: Interpretable and Robust Surgical Workflow Modeling (Paper Link): A framework that learns dynamic scene graph prototypes for surgical workflow analysis. No public code provided.CITab (Paper Link): Evaluated on Alzheimer’s disease diagnosis tasks across multiple cohorts, using a header embedding mechanism and P-MoLin module. Code available at https://github.com/jinlab-imvr/CITab.Topologically-Stabilized Graph Neural Networks (Paper Link): Evaluated across six diverse graph datasets. Code available at https://github.com/jelena-losic/topologically-stabilized-gnns.Federated Few-Shot Learning for Epileptic Seizure Detection (Paper Link): Utilizes TUH Event Corpus and BIOT for patient-specific EEG-based seizure detection. No public code provided.RVM: Recurrent Video Masked Autoencoders (Paper Link): A transformer-based recurrent neural network for video representation learning, demonstrating competitive performance without distillation. Project page: https://rvm-paper.github.io.PvP: Data-Efficient Humanoid Robot Learning (Paper Link): Introduces SRL4Humanoid, an open-source framework for evaluating state representation learning methods on humanoid robots, validated on the LimX Oli robot. Code available at https://github.com/LimX-Dynamics/SRL4Humanoid.BLADE: Multi-Behavior Sequential Recommendation (Paper Link): Evaluated on real-world datasets, using dual item-behavior fusion and three behavior-level data augmentation methods. Code available at https://github.com/WindSighiii/BLADE.PSA: Predictive Sample Assignment for OOD Detection (Paper Link): Validated on two standard SCOOD benchmarks. Code available at https://github.com/ZhimaoPeng/PSA.SCFA: Supervised Contrastive Frame Aggregation for Video Representation Learning (Paper Link): Evaluated on Penn Action and HMDB datasets, leveraging frame aggregation for efficient video representation. Code available at https://anonymous.4open.science/r/SCFA-04D4/.ACR: Fine-Grained Zero-Shot Learning with Attribute-Centric Representations (Paper Link): Achieves new state-of-the-art performance on CUB, AwA2, and SUN benchmarks. No public code provided.CLARGA: Multimodal Graph Representation Learning (Paper Link): A flexible multimodal fusion architecture for arbitrary sets of modalities. No public code provided.RGVT: Fully Inductive Node Representation Learning via Graph View Transformation (Paper Link): Outperforms existing models on node classification benchmarks, utilizing OGBN-Arxiv dataset. Code available at https://github.com/kaist-ml/GVT and https://github.com/kaist-ml/RGVT.HypeGBMS: Hyperbolic Gaussian Blurring Mean Shift (Paper Link): Extends Gaussian Blurring Mean Shift to hyperbolic spaces for clustering hierarchical data. No public code provided.AgentBalance: Backbone-then-Topology Design for Cost-Effective Multi-Agent Systems (Paper Link): Addresses performance-cost trade-offs in MAS with heterogeneous LLM backbones. Code available at https://github.com/usail-hkust/AgentBalance.Bhargava Cube–Inspired Quadratic Regularization (Paper Link): Demonstrates interpretable 3D embeddings on MNIST. No public code provided.Symmetry-Loss: Free-Energy Perspective on Brain-Inspired Invariance Learning (Paper Link): A brain-inspired algorithmic principle for enforcing invariance and equivariance. No public code provided.scRCL: Refinement Contrastive Learning of Cell-Gene Associations (Paper Link): Evaluated on scRNA-seq and spatial transcriptomics datasets for unsupervised cell type identification. Code available at https://github.com/THPengL/scRCL.LaT-IB (Paper Link): Addresses label noise in Information Bottleneck learning. Code available at https://github.com/RingBDStack/LaT-IB.UniCoR: Modality Collaboration for Robust Cross-Language Hybrid Code Retrieval (Paper Link): Utilizes a large-scale multilingual benchmark for code retrieval. Code available at https://github.com/Qwen-AI/UniCoR.EmerFlow: LLM-Empowered Representation Learning for Emerging Item Recommendation (Paper Link): Enhances recommendation systems for emerging items using LLMs. No public code provided.NAC: Neuronal Attention Circuit for Representation Learning (Paper Link): Achieves state-of-the-art results across irregular time-series classification, autonomous vehicle lane-keeping, and industrial prognostics. Code available at https://github.com/itxwaleedrazzaq/neuronal_attention_circuit.Enhancing Fake-News Detection with Node-Level Topological Features (Paper Link): Improves fake news detection on UPFD Politifact dataset. No public code provided.HGC-Herd: Efficient Heterogeneous Graph Condensation (Paper Link): Evaluated on ACM, DBLP, and Freebase datasets, outperforming sampling and gradient-based baselines. No public code provided.Stanford Sleep Bench (Paper Link): A large-scale PSG dataset with over 163,000 hours of sleep recordings, evaluating SSRL methods for sleep foundation models. Code available at https://arxiv.org/pdf/2512.09591.StateSpace-SSL (Paper Link): A linear-time self-supervised learning framework for plant disease detection. No public code provided.Log NeRF: Comparing Spaces for Learning Radiance Fields (Paper Link): Contributes new NeRF videos in GPLog encoding and a process for linearizing them. Code available at https://github.com/google-research/multinerf.GPSSL: Self-Supervised Learning with Gaussian Processes (Paper Link): Utilizes Gaussian processes for representation learning without explicit supervision, integrating uncertainty quantification. No public code provided.CLSS: Contrastive Learning for Semi-Supervised Deep Regression (Paper Link: Leverages spectral seriation for generalized ordinal rankings. Code available at https://github.com/xmed-lab/CLSS.PART: How PARTs assemble into wholes: Learning the relative composition of images (Paper Link): A self-supervised method modeling relative transformations between off-grid image patches. Code available at https://github.com/Melika-Ayoughi/PART.SpectrumFM: A Foundation Model for Intelligent Spectrum Management (Paper Link): Achieves state-of-the-art performance on modulation recognition tasks, using the DeepSig dataset. Code available at https://github.com/ChunyuLiu188/SpectrumFM.git.Unsupervised Representation Learning from Sparse Transformation Analysis (Paper Link): A generative framework leveraging Helmholtz decomposition for controllable latent representations in video data. Code available at https://github.com/KingJamesSong/latent-flow.## Impact & The Road Aheadadvancements have profound implications across numerous domains. In robotics, PvP and ARCADE promise more adaptable and data-efficient humanoid control, leading to robust autonomous systems in unpredictable environments. For medical AI, from seizure detection (Federated Few-Shot Learning for Epileptic Seizure Detection Under Privacy Constraints) to precision pathology (MAD, Vision Foundry) and cell-type identification (scRCL), privacy-preserving and fine-grained analytical tools are emerging. The ability to handle complex and noisy data (SARMAE) and to learn from sparse labels (In-Context Semi-Supervised Learning) significantly widens the applicability of AI, especially in data-scarce scenarios.push for interpretable AI is also evident. ProtoFlow for surgical workflows, Bhargava Cube regularization for structured embeddings, and the Attribute-Centric Representations in Fine-Grained Zero-Shot Learning highlight a growing demand for models that not only perform well but also offer transparent insights into their decision-making processes. Furthermore, the foundational models like SpectrumFM and the theoretical work on Platonic representations and Symmetry-Loss point towards a future where AI systems are built on more unified, principled, and biologically inspired learning paradigms.road ahead involves scaling these innovations, addressing remaining challenges in efficiency, generalization, and the seamless integration of diverse data types. The emphasis on self-supervision, topological awareness, and robust learning under real-world constraints suggests a future where AI can tackle increasingly complex problems, fostering a new era of intelligent systems that are more reliable, adaptable, and deeply understood.