Representation Learning Takes Center Stage: From Hyperbolic Geometry to Multimodal Fusion

Latest 68 papers on representation learning: Apr. 18, 2026

Representation learning continues to be the bedrock of modern AI, shaping how machines perceive, reason, and interact with complex data. Recent advancements highlight a fascinating trend: a move towards more structured, interpretable, and multimodal representations, often leveraging insights from human cognition and advanced mathematical frameworks. From tackling the nuances of irregular time series to empowering large language models with spatial intelligence, researchers are pushing the boundaries of what’s possible.

The Big Idea(s) & Core Innovations

A central theme emerging from recent papers is the pursuit of robust, disentangled, and generalizable representations that transcend modality-specific limitations. A standout in this area is OmniGCD: Abstracting Generalized Category Discovery for Modality Agnosticism by Jordan Shipard et al. from SAIVT, QUT, Australia. They introduce OmniGCD, a modality-agnostic approach to Generalized Category Discovery (GCD), proving that category formation is fundamentally abstract. By decoupling representation learning from category discovery via a novel GCDformer Transformer trained on synthetic data, OmniGCD achieves zero-shot GCD across vision, text, audio, and remote sensing without fine-tuning.

Parallel to this, causal and geometric structures are proving critical. In Metric-Aware Principal Component Analysis (MAPCA): A Unified Framework for Scale-Invariant Representation Learning, Michael Leznik from the University of Hertfordshire reveals that IPCA is uniquely scale-invariant under diagonal rescaling, unifying various whitening and self-supervised learning methods. This framework provides a deeper understanding of how different SSL methods operate in opposite spectral directions. Similarly, Identifiability of Potentially Degenerate Gaussian Mixture Models With Piecewise Affine Mixing by Danru Xu et al. from the University of Amsterdam provides theoretical guarantees for identifying latent variables in causal representation learning using sparsity regularization, even for degenerate mixture models without auxiliary variables.

Multimodality and hierarchy are also gaining traction. EEG-Based Multimodal Learning via Hyperbolic Mixture-of-Curvature Experts by Runhe Zhou et al. from Nanyang Technological University introduces EEG-MoCE, a hyperbolic framework that models hierarchical structures in brain signals and complementary modalities. Each modality gets an expert in a learnable-curvature hyperbolic space, revealing that curvature magnitude correlates with modality importance. For challenging scenarios like multi-hop semantic transmission in LEO satellite networks, Joint Semantic Coding and Routing for Multi-Hop Semantic Transmission in LEO Satellite Networks by Hong Zeng et al. from Chongqing University of Posts and Telecommunications proposes GraphJSCR, leveraging graph attention networks for joint routing and semantic coding under partial observability. In a similar vein, UNIGEOCLIP: Unified Geospatial Contrastive Learning by Guillaume Astruc et al. from LASTIG, Univ Gustave Eiffel, IGN, ENSG, France, achieves all-to-all contrastive alignment across five geospatial modalities (imagery, DSMs, text, coordinates) in a single unified embedding space, highlighting the power of multi-scale frequency fusion and self-attention in coordinate encoding.

Addressing critical issues in real-world applications, DBGL: Decay-aware Bipartite Graph Learning for Irregular Medical Time Series Classification by Jian Chen et al. from The University of Hong Kong proposes modeling irregular medical time series as patient-variable bipartite graphs. Their node-specific temporal decay encoding accurately captures variable-dependent decay rates, proving robust even with 50% missing data. In healthcare, Text-Attributed Knowledge Graph Enrichment with Large Language Models for Medical Concept Representation from Mohsen Nayebi Kerdabadi et al. at the University of Kansas introduces COMED, an LLM-GNN framework for medical concept representation. It combines EHR statistics with type-constrained LLM prompting for relation inference and jointly trains a LoRA-tuned LLaMA encoder with a heterogeneous GNN, mitigating hallucination and excelling in rare diagnosis codes.

For improving interpretability, Selecting Feature Interactions for Generalized Additive Models by Distilling Foundation Models by Jingyun Jia et al. from the University of Wisconsin-Madison introduces TabDistill, a method using tabular foundation models (TFMs) and post-hoc interaction attribution to identify meaningful feature interactions for interpretable GAMs. This bridges the gap between powerful black-box models and transparent interpretability. The shift from alignment/reconstruction to prediction is also evident in From Alignment to Prediction: A Study of Self-Supervised Learning and Predictive Representation Learning by Mintu Dutta et al. from Pandit Deendayal Energy University. They formalize Predictive Representation Learning (PRL) and position I-JEPA as a canonical framework, demonstrating its superior occlusion robustness by predicting unobserved latent representations rather than just aligning observed views.

Under the Hood: Models, Datasets, & Benchmarks

• OmniGCD (Code): A GCDformer Transformer trained on synthetic data for zero-shot generalized category discovery across 16 datasets in vision, text, audio, and remote sensing. Utilizes t-SNE for optimal latent space creation.
• DPF-GFD (Code): Uses a Beta wavelet-based adaptive filter and a kNN low-pass filter on original and similarity graphs, followed by XGBoost for financial fraud detection on datasets like FDCompCN, FFSD, Elliptic, and DGraph.
• AgentEA (Code): A multi-agent debate framework using Direct Preference Optimization (DPO) with LLaMA3-8B-Instruct fine-tuning for entity alignment, evaluated on DBP15K and ICEWS datasets.
• GraphJSCR: Employs Graph Attention Networks (GAT) with Proximal Policy Optimization (PPO) for joint routing and semantic coding in LEO satellite networks. Evaluated using ns-3.41 simulator and the DIV2K dataset for semantic image quality.
• DS2DL (Code): Combines an Unsupervised Masked Autoencoder (UMAE) with a Vision Transformer backbone for denoised latent representations, followed by spatially-regularized diffusion clustering for hyperspectral images on Botswana and KSC datasets.
• COMED (Code): Integrates a LoRA-tuned LLaMA encoder and a heterogeneous GNN for medical concept representation, using MIMIC-III/IV datasets. It leverages LLMs like Llama-3.2-1B, Gemma-2-2B, and Qwen2.5-1.5B.
• MMOT: A Gaussian Mixture Model framework driven by Optimal Transport theory for online class incremental learning, tested on MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet.
• EEG-MoCE (Code to be released): A hyperbolic Mixture-of-Curvature Experts model for EEG-based multimodal learning, evaluated on EAV (emotion), ISRUC (sleep), and Cognitive (N-back) datasets.
• SEATrack (Code available): Features AMG-LoRA for cross-modal attention alignment and HMoE (Hierarchical Mixture of Experts) for efficient global relation modeling in multimodal tracking, achieving SOTA on LasHeR, DepthTrack, and VisEvent datasets.
• TAPF: A Timing-Aware Pre-Quantization Fusion approach for video-enhanced audio tokenization, using knowledge distillation and dynamic temporal alignment. Tested on AudioSet and AVQA datasets.
• GigaCheck (Code): Adapts DETR-style vision models for object-centric span localization of LLM-generated text, using a shared LoRA-tuned backbone, robust across LLaMA and Qwen models.
• EA-Agent (Code): A reasoning-driven agent for entity alignment, using attribute and relation triple selectors with reward-guided offline policy optimization. Evaluated on DBP15K and SRPRS benchmarks.
• STS-Mixer (Code): Leverages Graph Fourier Transform and Frequency-Aware Attention for 4D point cloud video understanding, achieving strong results on MSR-Action3D and Synthia4D.
• CoRe-ECG: A self-supervised framework combining contrastive and reconstructive learning with Frequency Dynamic Augmentation (FDA) and Spatio-Temporal Dual Masking (STDM) for 12-lead ECG representation, pre-trained on MIMIC-IV-ECG and evaluated on PTB-XL, ICBEB2018, and Ningbo.
• OctEncoder (Code to be released): A unified hierarchical transformer for morphometric analysis of brain structures, using topology-guided masked autoencoders for Alzheimer’s and focal cortical dysplasia detection.
• DT-Pose (Code): A two-phase framework with temporal-consistent contrastive learning and a topology-constrained decoder (GCN + Transformers) for robust WiFi-based human pose estimation on MM-Fi, WiPose, and Person-in-WiFi-3D datasets.
• Sim-CLIP: An unsupervised Siamese adversarial fine-tuning framework for Vision-Language Models like CLIP, enhancing robustness and semantic richness for tasks like image captioning and zero-shot classification.
• Reasoning-Based Refinement of Unsupervised Text Clusters with LLMs (Code): Uses LLMs as semantic judges for coherence verification, redundancy adjudication, and label grounding, tested on social media corpora.
• DFR-Gemma: Enables LLMs like Gemma to reason directly over dense geospatial embeddings by projecting them into the LLM’s latent space as semantic tokens. Addresses multi-task geospatial benchmarks.
• RDVQ (Code): A unified framework for differentiable Vector Quantization and end-to-end rate-distortion optimization in generative image compression, achieving superior perceptual quality at low bitrates with a lightweight architecture.
• NOSE: Unifies molecular structure, receptor sequences, and linguistic descriptions into a single embedding space using tri-modal orthogonal contrastive learning, improving zero-shot odor retrieval.
• CARE-ECG: Integrates causal graph inference and counterfactual reasoning with large language models for explainable ECG interpretation, encoding multi-lead ECGs into latent biomarkers.
• HOI-DA: A pair-centric framework unifying detection and anticipation in video Human-Object Interaction (HOI) by modeling future interactions as residual transitions, introducing DETAnt-HOI benchmark.
• M-IDoL: A self-supervised medical foundation model using information decomposition with a Mixture-of-Experts (MoE) projector to learn modality-specific and diverse representations across X-ray, fundus, OCT, dermoscopy, and pathology images.
• BiScale-GTR: A unified framework for multi-scale molecular representation learning, combining Graph BPE tokenization with a parallel GNN-Transformer architecture, achieving SOTA on MoleculeNet, PharmaBench, and LRGB.
• MorphDistill (Code): A two-stage framework for distilling unified morphological knowledge from ten pathology foundation models using dimension-agnostic multi-teacher relational distillation for colorectal cancer survival prediction. Tested on Alliance/CALGB 89803 and TCGA cohorts.
• DBGL (Code in supplementary material): Models irregular medical time series as patient-variable bipartite graphs with node-specific temporal decay encoding, outperforming baselines on P19, P12, MIMIC-III, and Physionet.
• AusRec (Code): An automatic self-supervised learning framework for social recommendations that uses meta-learning optimization to adaptively weight multiple social relation tasks, outperforming baselines on LastFM, Epinions, and DBook.
• ToGRL: Enhances heterogeneous graph representation learning by optimizing graph structure via a Graph Structure Learning (GSL) module and using prompt tuning for downstream tasks, tested on five real-world datasets.
• HCL: A Hierarchical Contrastive Learning framework that explicitly captures globally shared, partially shared, and modality-specific structures in multimodal data, with theoretical guarantees for identifiability and recovery accuracy, improving EHR prediction on MIMIC-IV.
• Progressive Deep Learning: A training strategy that gradually activates deeper network blocks to improve SOS maturation assessment from CBCT images, achieving accuracy gains with less compute on a curated CBCT dataset and CIFAR-10.
• PHONSSM (Code): A novel architecture using state space models and anatomically-grounded graph attention to enforce phonological decomposition for vocabulary-scale sign language recognition, achieving SOTA on WLASL2000 and Merged-5565 using skeleton data.
• Perceptual Inductive Bias: A pre-training stage for contrastive learning leveraging figure-ground segmentation and intrinsic image decomposition to inject inductive biases, leading to 2x faster convergence and improved robustness on tasks like object recognition, segmentation, and depth estimation.
• Multi-Frequency Local Plasticity: A hierarchical framework combining multi-frequency Gabor streams, competitive learning, and associative memory to achieve 80.1% accuracy on CIFAR-10 with 93% parameters updated via local Hebbian rules, demonstrating the power of structured architectural biases.
• Deep Privacy Funnel Model (Code): Introduces the Deep Variational Privacy Funnel (DVPF) framework with discriminative (DisPF) and generative (GenPF) models for information-theoretic privacy preservation in face recognition, compatible with ArcFace and AdaFace.
• Bayesian-ARGOS (Code): A hybrid framework combining frequentist screening with Bayesian inference for automated equation discovery, achieving 100x speedup and outperforming SINDy in data efficiency and noise robustness on chaotic systems and NOAA sea surface temperature data.

Impact & The Road Ahead

The collective impact of these advancements is profound, promising more intelligent, robust, and ethical AI systems. The push towards modality-agnostic and 3D-aware representations (OmniGCD, GeoLink, UniSplat) is enabling AI to understand the world more holistically, much like humans do. The integration of causal inference and theoretical guarantees (MAPCA, Identifiability of pdGMMs, HCL, Causal Inference in GRL) moves AI from correlation to causation, fostering trust and reliability, especially in high-stakes domains like healthcare (CARE-ECG, M-IDoL, MorphDistill, CoRe-ECG, DBGL). Furthermore, the innovative use of LLMs not just as text generators but as semantic judges and reasoning engines (AgentEA, Reasoning-Based Refinement, DFR-Gemma, Schema-Adaptive Tabular Representation Learning, GigaCheck) marks a significant step towards human-aligned interpretability and zero-shot generalization across diverse data types and schemas.

The development of specialized foundation models for specific domains like medical imaging (M-IDoL, SEM Foundation Model) and scientific discovery (Bayesian-ARGOS, BiScale-GTR) demonstrates a maturing field, moving beyond generic models to tailor AI for complex scientific challenges. Innovations in efficiency and scalability (RDVQ, STS-Mixer, ToGRL) are crucial for deploying these powerful models in real-world, resource-constrained environments, while tackling issues like representation collapse (DIAURec, Minimal Model of Representation Collapse) ensures their stability. The exploration of hyperbolic geometry (EEG-MoCE, Hyperbolic Social Influence Maximization) and phonological compositionality (PHONSSM) hints at a deeper understanding of underlying data structures, potentially leading to more biologically plausible and human-like AI.

The road ahead will likely see continued convergence of these themes: increasingly multimodal and hierarchical representations, deeply embedded with causal understanding, and capable of efficient, interpretable reasoning across diverse, often noisy, real-world data. As AI systems become more entwined with our lives, robust, ethical, and explainable representation learning will be paramount.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.