Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A controlled ablation on character-level language modelling with a 6-layer GPT on TinyShakespeare found that the optimal single-scale bandwidth for FourierQK bandpass-filtered attention is sigma ~= 2 bins centred at para…
A new arXiv paper (2610.00251v1) argues that memorization audits of generative models are unreliable because they compare similarity scores against thresholds without a null distribution, and supplies two exact null test…
A new arXiv paper (2610.00257v1) introduces "attention manifolds," learned 2D B-spline surfaces S_d(q_d, k_d) that modulate each value dimension based on query-key interaction, reducing WikiText-2 validation perplexity b…
Researchers proposed FedCORE, a federated adaptation framework that represents multimodal Encoder and graph neural network (GNN) updates through a shared low-dimensional latent state, according to an arXiv paper (arXiv:2…
A new arXiv paper (2610.00373v1) reports that attention-head ablation in GPT-2 small can produce misleading causal claims unless intervention placement, evaluation metric, and controls are validated. The authors found a …
Researchers proposed STCFormer, an adaptive spatio-temporal Transformer that dynamically groups weather stations by their local evolution within each temporal patch, and reported the lowest 24-hour mean squared error on …
A first empirical study of general-purpose decision models for network traffic classification found that Jev reached 28.42% accuracy on ten CESNET-QUICEXT-25 application labels after 40 fixed labeled examples, up from 9.…
Researchers posting arXiv paper 2610.00395v1 introduced specificity-aware diffusion steering, a target-design method that derives a target distribution from an overlap-based objective so the desired reference distributio…
WIPSNet, a 3D ResNet operating on stacked continuous wavelet transform scalograms of overnight impedance pneumography (IP) signals, achieved an AUC of 0.783 ± 0.026 for night-level paediatric wheeze classification, accor…
Researchers introduced MERITED, a framework for metacognitive reasoning that combines Instance-Based Learning Theory (IBLT) with Energy Based Models (EBMs) to let AI systems allocate computational effort based on uncerta…
Researchers introduced RACE (Residual-Aware Correction of Forecasting Errors), a two-stage test-time adaptation framework that improves time-series foundation model (TSFM) forecasting by using same-domain neighbor eviden…
A new arXiv paper (2610.00499v1) proposes the Denoising Workload Surface (DWS), a two-dimensional block-step probability surface that predicts per-request inference cost for diffusion large language models (dLLMs). In re…
A study of hZACH-ViT, a compact Vision Transformer with Euclidean, Poincaré, and spherical prototype heads, found that reducing Poincaré curvature from c=1 to c=0.1 improved clean MacroF1 in all 15 paired seed-dataset co…
Researchers introduced WOMBAT, a benchmark of 14 whitebox graph neural networks (GNNs) whose message-passing weights are hand-set to detect a specific SMARTS motif, providing attribution ground truth for evaluating expla…
JEPA-TTT, a method that adapts the latent dynamics predictor of a pretrained action-conditioned Joint-Embedding Predictive Architecture world model throughout test time, improved planning on all eight dynamics shifts acr…
Developer Ben Houston released three-dlss-nr, a Three.js port of maanHimself's OpenDLSS-NR that runs the 71-block DLSS 5 neural rendering network as TSL compute kernels inside a WebGPURenderer with no CPU readback. The p…
WeightBridge, a weight-transfer library for reinforcement learning systems, reduces average GPU stall time by up to 42× over the state-of-the-art open-source RL framework across configurations spanning different models, …
A developer built a simulated system combining privacy-preserving active learning and embodied agent feedback loops for autonomous rover swarms conducting planetary geology surveys. The architecture uses PyTorch, a custo…
TypeSafe AI launched Jev, a 40B-parameter "System One" transformer trained exclusively on synthetic data via RLCD that returns typed values with calibrated probabilities instead of free text, in early access on September…
Paradigma released Limit 1B Violetto, a 1 billion parameter dense transformer math-specialist model trained from scratch on under 300 billion curated tokens with a 131k context window, under Apache 2.0 weights and a cust…