A component-level decomposition study of Dr. Bojan Ploj's bipropagation method finds that per-layer supervision, not gradient locality, is the ingredient that makes greedy layer-wise supervised traini…
MintFlow's training-free minimal-intervention strategy enforces constraints on flow-based generative models with a 2x reduction in KL drift versus projection after the final step, according to the Min…
Researchers introduced Rank-Aware Speculative Sampling (RASS), a verification rule for speculative draft trees based on rank-aware list coupling, which improves on Diffusion Greedy Rejection Sampling …
A new arXiv paper (2610.02344v1) develops an early-training stability theory for Joint-Embedding Predictive Architectures (JEPAs) that explains representation collapse through a per-mode stability rat…
A new arXiv paper (2609.30647v1) formalizes visual representation learning as a conditional predictive sufficient statistic (CPSS), showing that predicting the next patch embedding with a cosine loss …
A new arXiv paper (2609.19384v1) proposes Riemannian–Lorentz Parameter Fusion (RLPF), a geometry-aware method for merging a Vision Transformer and a state-space model into a two-branch hybrid that rea…
Researchers proposed Sharp Mode Connectivity (SMC), a sharpness-aware reformulation of mode connectivity that enforces low loss across the entire neighborhood of a weight-space interpolation curve rat…
A new parameter-pruning scheme derived from the Fisher information metric's geodesic distance in model space outperforms magnitude pruning and local Fisher information pruning across every architectur…
A controlled test using a DDPM U-Net on CIFAR-10 found that LoRA rank 4 achieved the best FID score of 124.1380, slightly beating rank 8 at 124.2136, according to data from arXiv:2609.10656v1. The stu…
A controlled study submitted to arXiv on 9 Sep 2026 found that moderate LoRA ranks are the most efficient setting for diffusion model fine-tuning, with rank 4 achieving the best DDPM FID of 124.1380 o…
Researchers demonstrated that smooth two-layer feed-forward networks (FFNs) in transformer models expose a structural model extraction channel under chosen-input raw-output access, recovering hidden F…
Researchers studying Focal Calibration Loss (FCL) found that while it preserves the Bayes decision rule for all gamma and lambda, it is generally not proper for gamma > 0, causing the Bayes-optimal pr…
Researchers propose FL-MAESTRO, a multi-agent orchestrator that uses three specialist large language model (LLM) agents to jointly decide communication topology, per-client resource allocation, and ag…
Researchers at an undisclosed institution found that undervolting GPUs during CNN training improves adversarial robustness while cutting energy use. In tests on MNIST and CIFAR-10 with LeNet, VGG-6, a…
Researchers introduced a physical tokenizer that performs analogue light-to-token conversion directly at the sensor level for vision transformers, integrating a 32 × 32 photosensitive memory array of …
A new MIT CSAIL study published in Nature Communications finds that individual training examples often have no measurable influence on any specific output of large generative image models, a phenomeno…
Researchers introduced SAGE, an uncertainty-modulated surrogate-gradient mechanism for Transformer-based spiking neural networks (SNNs), which adapts the surrogate-gradient slope during training using…
A new arXiv preprint (2608.00246v1) proposes a retain-aware localization method for machine unlearning that considers parameter importance to both forgotten and retained data, reducing collateral dama…
A new arXiv preprint (2608.00078v1) presents a mobile deployment pipeline that uses QLoRA fine-tuning of an LLM to generate convolutional neural networks, with INT8 TensorFlow Lite export and on-devic…
Researchers propose a loss-aware feature-map pruning framework for convolutional neural networks that uses multi-armed bandits (UCB1 and Thompson Sampling) to identify and remove redundant feature map…