ShapeEmbed is a self-supervised method using Euclidean distance matrices to encode the shape of objects in 2D images in a way that is robust to shape-preserving geometric transformations

The paper systematically analyzes the phenomenon of varying sparsity ratios across views in multi-view learning, and proposed a targeted data-driven network architecture based on Sparse Autoencoder with Adaptive Constraints.

For datasets with high-magnitude noise features, joint-embedding is more robust than reconstruction for self-supervised learning.

ScatterAD leverages representation scattering as an inductive signal to jointly model temporal and topological patterns for effective multivariate time series anomaly detection.

We introduce a new probing architecture capable of efficiently combining features from multiple pre-trained vision foundation models.

OSKAR is a self-supervised multimodal model that predicts masked token latent features from video, skeleton, and text using momentum target encoders—outperforming specialized models across tasks without reconstruction or contrastive losses.

We introduce ChA-MAEViT - a robust, high-performing model that encourages cross-channel interactions for multi-channel imaging.

WhAM: a transformer model unifying generation, acoustic translation and classification of sperm whale vocalizations

PANTHER is a hybrid framework that combines self-supervised generative pretraining with lightweight discriminative modeling to enable real-time fraud detection in large-scale payment platforms, delivering a 38% improvement in online performance.

A new lower-bound for representation learning connecting Jensen–Shannon and Kullback–Leibler divergences.

This paper introduces a computationally efficient foundation model that achieves topology-invariant EEG analysis by projecting signals from diverse electrode configurations onto a unified latent space.

We introduce a novel, scalable framework to evaluate compositional generalization, leverage it to evaluate more than 5k models, and propose a family of neural models pushing the Pareto frontier on this task.

We introduce ZEUS, a transformer pretrained on synthetic datasets to perform zero‐shot, fully unsupervised clustering of tabular data—no fine‐tuning or hyperparameter search needed—and show it matches or outperforms classical and deep baselines.

a new learning scheme for multi-modal LLM, LLAVA

Interpretability methods based on linear, orthogonal features fall short for modern neural representations, which are often hierarchical and nonlinear. Better results come from aligning methods with the true structure of these representations.

We propose HASTE, which combines holistic alignment (feature and attention) with early termination to accelerate diffusion transformer training by 28× while maintaining quality.

Variational supervised contrastive learning maximizes a posterior-weighted ELBO, replacing pairwise comparisons with class-level interactions for SOTA performance on image classification tasks.

Jointly modeling image laments and high-level semantic features (e.g. DINO), significantly enhances generative quality and training efficiency.

learning pixel-dense multi-task representations from apparent motion

We show that self-supervised models like DINOv2 can develop strong noise robustness without any explicit denoiser at downstream fine-tuning or inference, by leveraging a data curriculum and a denoised regularization loss during pretraining.

Aligning pretrained unimodal models with the proposed framework using limited paired data yields ~52% gains in cross-modality zero-shot classification and ~92% in retrieval.

A pretraining scheme to learn representations that are robust to diverse sensor failures.

A Unified Principle for Generative Modeling, Representation Learning, and Classification

We show that dynamics-aligned representations improve zero-shot generalization for contextual world models.

We show that interval estimation based methods produce better distilled embedders in multi-teacher distillation settings compared to MSE or Cosine base methods.

Introducing a learned algorithm for curating datasets for CLIP pretraining that achieves state-of-the-art ImageNet accuracy on the DataComp benchmark.

Using theory and simulations, we show that task-irrelevant stimuli could cause long-term drift of neural representations

We systematically study the behavior and mechanism of a language-centric learning method to learn omnimodal representation, and propose the generation-representation scaling law.

We introduce Deep Taxonomic Networks, a deep latent variable approach that uses a complete binary-tree mixture-of-Gaussians prior in a VAE framework to discover interpretable hierarchical taxonomies and prototype clusters from unlabeled data

We introduce Obliviator, a nonlinear concept erasure method that guards against nonlinear adversaries. For any level of unwanted attribute protection, our method achieves higher task performance, revealing an empirical upper bound on this trade-off.

We introduce generative distribution embeddings (GDE), a framework that lifts autoencoders to the space of distributions.

We propose a self-supervised framework that combines GNNs and a Transformer-based masked autoencoder to detect network intrusions by reconstructing flow representations and flagging high-error patterns as anomalies.

Combinatorial Open-World Detection

This paper presents CoUn, a contrastive learning (CL)-based machine unlearning (MU) framework using only retain data. Further, our proposed CL module can be integrated with existing baselines to empower their performance.

The first benchmark for multi-factor sequential disentanglement representations, introduces a novel method, and leverages Vision-Language Models to automate annotation and evaluation—enabling scalable, label-free workflows.

Meta-RL with self-supervised predictive coding modules can learn interpretable, task-relevant representations that better approximate Bayes-optimal belief states than black-box meta-RL models across diverse partially observable environments.

Under input uncertainty, transformer models exhibit a systematic exploration of input‑agnostic conceptual representations, increasing the likelihood of hallucinations.

Learning proof system dynamics, pruning proof search based on diversity and expected outcome

We study how different changes to the auto-encoder rate-distortion trade-off affect discrete auto-regressive generation performance through scaling laws.

Small KL-divergence fails to ensure similarity of representations; we propose a distance which does and demonstrate it empirically.

We introduce the first method for translating text embeddings from one vector space to another without any paired data, encoders, or predefined sets of matches.

We introduce ACIA, a measure-theoretic framework for anti-causal representation learning through two-level abstraction, supporting both perfect and imperfect interventions with theoretical guarantees.

We introduce a graph AutoEncoder that can encode (and decode) graphs of different sizes into the same embedding space.

We ensure backward compatibility through multiple transformations and a relaxed orthogonality constraint for distribution-specific adaptation.

We study meta-learning in linear bandits and provide provably fast, sample-efficient algorithms to learn a common set of features from multiple related bandit tasks and to transfer this knowledge to new, unseen bandit tasks.

We introduces a simple yet effective pipeline called Token Bottleneck that facilitates conservative summarization of the observed scene into a bottleneck token while enable capturing of dynamic transitions through the bottleneck token.

We propose UnCLe, a deep learning framework that uses temporal perturbations to uncover dynamic, time-varying causal graphs from observational time series.

We introduce CG-SSL, a concept-guided self-supervised learning framework that aligns meaningful image regions across views, achieving state-of-the-art performance on dense prediction tasks.

We show that supervised semantic correspondence methods fail to generalize well to unseen keypoints and we introduce geometric constraints during training to address this.

a genomic sequence model that unifies long-context reasoning and sequence interpretation with state-of-the-art performance.

AdaTS introduces a dynamic soft contrastive learning framework for adaptive time series representations, leveraging frequency-domain similarities with dynamic temporal and ordinal instance contrasts.

In this paper, we systematically evaluate RL and control-based methods on a suite of navigation tasks, using offline datasets of varying quality.

This work presents DAIL, a semantic aligned reinforcement learning method that improves task discrimination and generalization in language-conditioned problems.

We introduce a systematic approach to flexibly incorporating representation guidance into diffusion models, resulting in both accelerated training and better performance across image, protein, and molecule generation tasks.

We introduce a regularizer that promotes diverse, task-relevant features over spurious ones in contrastive learning.

Our Orthogonal Residual Update improves deep networks by adding only novel, stream-orthogonal module outputs, boosting generalization, stability, and efficiency.

While loss decreases monotonically during LLM training, the representations undergo distinct geometric phases across pretraining and post-training, which in turn determine when and how the model acquires memorization or generalization capabilities.

AsymDSD is a self-supervised learning framework for 3D point clouds that combines masked modeling and invariance learning through latent space prediction.

We theoretically formalize real-world misalignment in multimodal learning via latent-variable modeling, showing that learned representations inherently encode semantics invariant to selection and perturbation biases.