Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose LEAP, a training curriculum for Vision Transformer feature-based knowledge distillation that uses the teacher's intermediate feature maps as progressively difficult targets. LEAP achieves 90.1% accura…
Researchers propose ProMUSE, a progressive multi-modal uncertainty-guided network for Alzheimer's disease classification that adaptively determines when additional imaging modalities are needed, reducing MRI/PET usage by…
Researchers introduced LooseControlVideo, a framework that uses sparse, oriented 3D boxes as a proxy for spatial blocking in text-to-video generation, enabling intuitive control over layout and trajectory. Fine-tuned on …
Researchers introduced cAPM, a continual AI-assisted pace-mapping system using active learning, which achieved 81% probability of localizing ventricular tachycardia targets within 5 mm accuracy using 4.5 pacing sites, co…
Researchers introduced SSProNet, a secondary-structure-aware graph neural network for protein representation learning that uses energy-filtered hydrogen-bond graphs. The model outperforms existing graph-based methods on …
Researchers propose a method to learn asynchronous denoising schedules for multi-representation diffusion models, improving visual synthesis. On ImageNet 256x256, their approach achieves FID 1.05 with 4x less training th…
Researchers propose TeleMorpher, a one-shot framework for simultaneous motion-location editing in videos, addressing a gap in diffusion model-based editing. The method disentangles protagonist and background, uses motion…
FocusOnFun team's Vortex system achieved 90.5% in the Preliminary Round and 'Excellent' overall performance with 'Outstanding' QA results at the Ho Chi Minh City AI Challenge 2025. The multimodal video retrieval system i…
Researchers propose a physics-informed framework to discover anisotropic yield functions from full-field displacement and reaction force data, using a convex neural network representation. The method avoids direct stress…
Researchers introduced FlexLAM, a latent action model that replaces fixed-capacity bottlenecks with variable-length codes trained via nested dropout, resolving the trade-off between discarding transition cues and preserv…
Researchers introduced SLARouter, an online routing algorithm for large language model (LLM) applications that learns cost-optimal policies from sparse user feedback while guaranteeing Service Level Agreement (SLA) compl…
Researchers propose HypOProto, a hyperbolic prototype-based framework for interpretable left ventricular filling pressure classification from B-mode echocardiography, achieving state-of-the-art performance while maintain…
Researchers introduced Emyx, a 140M-parameter conditional flow matching model for all-atom protein generation, which outperforms larger models like RFdiffusion3 on enzyme design benchmarks while requiring only 682 GPU-ho…
A systematic experimental analysis of eight state-of-the-art Diffusion Language Models (DLMs) across eight benchmarks reveals that generation-time design choices strongly influence performance and computational efficienc…
Researchers proposed a hybrid GNN-FEM framework for phase-field fracture simulation that integrates a graph neural network surrogate into the conventional staggered scheme, replacing the phase-field update while retainin…
Researchers propose REVEAL++, a differentiable phenotypic grouping method for vision-language retinal modeling that improves Alzheimer's disease risk prediction by treating phenotypic similarity as a continuous, learnabl…
A new study measures the linearity of transformer feed-forward blocks, finding that linear recoverability (R²_lin) varies widely across blocks and is a learned property, not an architectural one. The analysis of GPT-2, P…
Researchers introduced a continuous relaxation of the determinantal point process (DPP) maximum a posteriori (MAP) problem for diversity-aware data selection, reformulating it as a nonlinear eigenvalue problem (NEPv) on …
Researchers introduced the Integral Transform Network (ITNet), a unified architecture that subsumes convolution, attention, and recurrence as special cases of a learnable integral transform. ITNet matches or exceeds spec…
Researchers at arXiv propose a framework for selecting the most informative comparison pairs in preference-based post-training of large language models, showing that strategic pair selection can improve sample efficiency…