Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose Structure-Preserving Epistemic Neural Networks (S-PENNs), a framework for uncertainty quantification in scientific machine learning models with hard architectural constraints, and instantiate it for G…
A new arXiv paper (2608.12438v1) formulates generative modeling as a path integral, unifying flow-based, diffusion-based, variational, and adversarial models under a single master action. The approach yields a one-loop c…
Researchers introduced Householder Graph Neural Network (HouseGNN), a deep GNN architecture that counters oversmoothing by using Householder reflections and GroupSort to preserve Euclidean norms at every node and depth, …
A new arXiv preprint (arXiv:2608.13087v1) reports that apparent gains from non-uniform test-time budget allocation in neural combinatorial optimization (NCO) are largely an artifact of in-sample evaluation. Across three …
Researchers introduced AIM, a framework interpreting momentum as a residual-driven multiplier correction, and proposed RADAR, an optimizer combining relativistic adaptive geometry, decoupled residual correction, and seco…
Researchers introduced Balanced Adaptive Prototype Selection (BAPS), a framework that compresses contexts for TabPFN inference on large-scale tabular data without retraining. Experiments on the million-row HIGGS and SUSY…
Researchers at an undisclosed institution introduced a proxemics-based reward formulation for deep reinforcement learning (DRL) social navigation that models each human's personal space as a radial Gaussian-mixture field…
Researchers introduced an incremental, multi-episode evaluation and training paradigm for Relational Deep Learning (RDL), which models multi-tabular databases as temporal heterogeneous graphs. The study found temporal co…
Researchers at arXiv introduced the Dual-Flow Transformer, a new architecture that decouples prompt prefill computation from additional decode-time computation, achieving lower validation loss across architectures and da…
Researchers introduced Generative Defect Isolation (GDI), a method using the LaMa inpainting model with Fast Fourier Convolutions to generate realistic single-defect training samples for multi-label defect classification…
Researchers introduced HIMEC, a remote sensing image change captioning method combining Directional Change Representation with fixed-interface decoding, achieving a CIDEr score of 142.81±0.60 on LEVIR-CC, outperforming d…
Researchers introduced WidgetGen, a lightweight tool-grounded framework that extracts observable text and color evidence, performs high-level layout and optional chart reasoning, and directly generates executable JavaScr…
Researchers introduced EgoCITE, a long-horizon agentic memory framework for egocentric question answering that improves accuracy by 4.4–14.2% over agentic memory baselines while achieving 36× lower cost than long-context…
A new arXiv study (2608.12658v1) finds that using three orthogonal seed slices—axial, coronal, and sagittal—during inference improves 3D organ segmentation in slice-propagation methods like Sli2Vol, boosting Dice by 21.9…
Researchers introduced Mr3D-VL, a 4-billion-parameter vision-language foundation model for multi-parametric 3D magnetic resonance imaging (mpMRI), achieving a BERTScore of 0.856 for report generation, question-answering …
VOS-Agent, a collaborative framework built on SAM3, won first place in the MOSEv2 Track of the 8th LSVOS Challenge at ECCV 2026, achieving 69.82% on the official J&F metric. The framework uses a Target Perception and Rou…
A new arXiv preprint (2608.12337v1) from researchers studying long-form hallucination reinforcement learning finds that strict grounding rewards improve factual support but suppress coverage, while unconstrained rubric r…
Researchers introduced MindMemOS, a portable and self-evolving memory operating layer for AI agents, which achieved 94.03% accuracy on LOCOMO and 70.63% on PersonaMem, and its MindSkillEvolve algorithm improved Spreadshe…
Researchers introduced ε-MemEvo, a framework for cross-task knowledge transfer in LLM program evolution, which stores prior experience as task-agnostic tactic memories and uses an adaptive injection gate to avoid negativ…
Researchers propose Thought-Aware Attention Matching (TAM), a KV cache compaction method that segments chain-of-thought reasoning into blocks, allocates compression budgets adaptively, and protects pivotal tokens, improv…