Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A 135M-class depth-recurrent transformer trained on FineWeb-Edu converges to a per-token fixed point, with mean successive-output KL divergence falling from 3.9e-1 at the second loop to 8.5e-6 by the sixteenth. The resea…
Researchers propose $K$-NeAS, a scalable architecture for multi-material CT reconstruction using neural signed distance functions, achieving superior 3D volumetric fidelity on clinical Cone-Beam CT datasets. The method r…
Researchers propose MIDiff, a diffusion-based framework that uses a Cross-Gramian Angular Sum Field to transform sparse multivariate mobile usage sequences into correlation images, achieving a Discriminative Accuracy of …
A new study from researchers at an undisclosed institution developed a digital-twin-inspired time-to-event model that integrates longitudinal ALSFRS-R trajectories with survival modeling to predict functional decline and…
A new machine learning framework from researchers predicts representative clutter height (RCH) for satellite ground station siting using LiDAR-derived terrain data, achieving a mean absolute error of 1.79m and R²=0.765, …
A study from arXiv:2607.14205 found that federated learning (FL) on radiology reports is vulnerable to privacy leakage via gradient inversion, with exact sentence reconstruction rates ranging from 31% to 44% across three…
Researchers introduce Ego Scene Augmentation (ESA), an egocentric spatial perception framework that enhances spatial reasoning in Multimodal Large Language Models (MLLMs) for egocentric Visual Question Answering (VQA). T…
Researchers introduce Branching Policy Optimization (BPO), a reinforcement learning algorithm for LLM agents that exploits deterministic, snapshottable sandboxes to reduce variance and improve training efficiency. On Web…
Researchers propose RENEW, a method that uses human preferences over imagined rollouts to repair model exploitation in offline reinforcement learning world models, achieving practical sample efficiency by focusing finetu…
Researchers have introduced CausalGraphX, a counterfactual graph neural network framework that integrates GNNs with counterfactual reasoning to provide explainable assessments of systemic risk in financial networks. The …
Researchers propose LIGO-PINN, a framework using learned initialization via gated layerwise optimization to overcome convergence failures in physics-informed neural networks (PINNs). In evaluations on 1D and 2D PDE domai…
Researchers introduce token time continuous diffusion (TTCD), a new diffusion language model that operates in continuous space and incorporates per-token times, allowing some tokens to proceed from noise to token faster …
A new study from arXiv (2607.14246v1) finds that generative models have a fixed steering budget set by training data, beyond which knobs like prompts and guidance scales cannot move a property, but concrete examples can.…
Researchers introduced RegNetAgents, a multi-agent AI framework for identifying regulatory driver candidates across heterogeneous gene regulatory networks in cancer genomics. The system integrates TCGA-derived cancer net…
Researchers propose SIRUS, a training-free inference-time framework for concept-level unlearning in text-to-video (T2V) models, achieving 70.4% average forgetting success and 25.7% average frame hit on CogVideoX, compare…
Researchers propose LyaGuide, a unified Lyapunov-guided framework for stabilizing generative flows in flow matching, establishing an equivalence between guided flow matching and Lyapunov control. The framework unifies co…
A new arXiv preprint proposes the Capability Convergence Hypothesis (CCH), which states that under a fixed per-token inference budget, representational convergence does not entail capability convergence; instead, capabil…
Researchers propose G2SR, a geometric method for fast and memory-efficient Gaussian-based surface reconstruction from few RGB views, achieving 69-89 reconstructions per second within 203 MB of GPU memory while matching o…
A new model-agnostic downstream reward framework from Pinterest improves long-term user engagement and retention across multiple surfaces including Homefeed, Related Pins, Search, and Notifications, according to a paper …
A new three-level hierarchical learning architecture for autonomous UAV swarms in search and rescue integrates Hebbian neuroplasticity, multi-agent reinforcement learning with graph neural networks and behavior trees, an…