self-supervised-learning

The paper develops an early-training stability theory for Joint-Embedding Predictive Architectures that unifies empirical collapse-avoidance heuristics via a drive/decay stability ratio, predicts a phase boundary confirmed across 800+ Tabular-JEPA configurations, and introduces a ResidualPred transformer predictor that improves effective rank and accuracy on tabular and I-JEPA image benchmarks.

This Texas A&M paper introduces "semantic recoverability" to explain which downstream tasks survive same-instance self-supervised learning, linking recoverability to class-distance-normalized variance, few-shot nearest-centroid transfer, and the spectral subspace spanned by cross-view-stable modes of the SSL objective.

提出软件在环重构（software-in-the-loop reconstruction）的自监督框架，从现有科学软件工作流中自动生成可执行参考行为与验证目标，用于训练终端智能体；使用 Qwen3.8-Max 生成训练数据并对 Qwen3.8-27B 微调，使 Terminal-Bench2 平均性能从 47.94% 提升至 53.56%。

Randal Balestrero announces his keynote talk at MICCAI-FOMO26 on recent progress in Joint-Embedding Predictive Architectures (JEPAs) and the need for more principled approaches to data imbalance, noise, multimodality, and reasoning.

PHASE is a physiology-guided hierarchical foundation model for intracranial EEG that uses physiological characteristics as explicit learning targets alongside masked latent prediction, setting a new state of the art on five Omni-iEEG clinical tasks and generalizing to unseen institutions.

JEPA-TTT adapts the latent dynamics predictor of a pretrained JEPA world model during test time using self-supervised dense replay, improving planning under dynamics shifts with 83% lower latent prediction error and 153% better planning performance over the frozen model across four continuous-control environments.

This paper studies self-supervised learning from continuous video streams using strict sliding-window batches, introduces the 95-hour WT++ urban walking-tour dataset, and proposes StreamMAE with stream-aware regularization and motion-biased crops to match i.i.d. MAE pretraining performance while scaling positively with stream size.

This paper conducts a controlled study on advection-aware graph nowcasting for distributed solar ramp forecasting, finding that accurate cloud-motion features are as important as graph structure and introducing a self-supervised estimator that reduces forecast error.

Atelier introduces a self-supervised hypernetwork framework that generates implicit neural representations for cryoEM density maps, enabling local feature extraction and improving performance on downstream annotation tasks.

EvolvingAvatar introduces a causal 3D head generation system that adapts in real-time to conversational contexts using test-time training, enhancing motion statistics in dialogue scenarios.

TT-VidT is a new video pretraining method that efficiently models motion by decoupling the temporal axis, achieving state-of-the-art results on motion-centric benchmarks with reduced computational cost.

This paper systematically evaluates masking strategies for EEG foundation models across MAE and JEPA frameworks, identifying optimal configurations and a novel failure mode, achieving comparable performance to REVE with reduced compute.

A physics-informed self-supervised learning framework is introduced for calibrating and reconstructing positions in multi-wire parallel plate avalanche counters using only detector geometry and charge constraints, eliminating the need for labeled data or dedicated calibration runs.

The paper proposes a general self-supervised learning framework for combinatorial optimization using Frank-Wolfe methods to handle constraints, with strong empirical results on problems like TSP, Maximum Coverage, and QAP.

EMGBlend introduces a self-supervised framework for pretraining on heterogeneous EMG datasets, addressing differences in electrode layouts, frequency support, and data source imbalances to improve gesture recognition and force decoding tasks.

This paper proposes the Activity Chain Encoder (ACE), a self-supervised model that learns daily activity patterns from mobile phone location data without labels, and demonstrates its effectiveness in identifying differentiated activity patterns in urban environments like London.

The paper evaluates ECG biometrics under realistic conditions involving exercise-induced stress and cross-session variability using a Siamese ResNet model, achieving state-of-the-art performance.

Tactile-JEPA is a self-supervised pre-training method for distributed tactile sensors that uses spatial topology to learn representations, improving force estimation and orientation tasks in robotics over prior state-of-the-art.

Toolformer trains language models to self-teach how to use external APIs like calculators and search engines via self-supervised learning, significantly improving zero-shot performance across tasks.

This paper proposes a new unified pre-training framework for medical code sequences that captures hierarchical structures and complex interactions, demonstrating superior performance in clinical event prediction and drug repositioning case studies for Alzheimer's disease.