Neural Networks

Researchers introduced Tensor Cache, a two-level memory system for Transformer models that stores evicted key-value pairs from a sliding-window cache into a compressed outer-product fast-weight memory. The approach uses …

Researchers have developed WeCon, a weight-conditioned neural solver that achieves comparable solution quality to state-of-the-art methods while reducing inference time by 40% across multiple multi-objective combinatoria…

Researchers introduced ImProver 2, a neurosymbolic framework for automated proof optimization in Lean 4 that combines a data-efficient expert-iteration pipeline with a scaffold exposing formal structure and lightweight i…

Researchers have found that chain-of-thought reasoning in large language models is only beneficial when early-stage entropy dynamics show consistent reduction, according to a new study on arXiv. The team introduced EDRM,…

A new method for measuring uncertainty in language model outputs, developed by researchers and detailed in a preprint on arXiv, extracts geometric features from the model's internal layer-by-layer processing to outperfor…

Researchers have introduced Human-Centered Learning Mechanics (HCLM), a dynamical and information-theoretic framework that treats deep learning as an open system under uncertainty, resource constraints, and human feedbac…

Researchers decomposed GPT-2 XL and Llama-3.1-8B into 16,000 to 32,000 interpretable features per layer using sparse autoencoders, finding that semantic features alone recovered 94% of peak brain-response encoding perfor…

A study of fMRI data from 112 English, Chinese, and French speakers found that brain-LLM alignment is driven by training-language dominance, not an inherent property of English. A Chinese-dominant model reversed the alig…

Researchers demonstrated that multi-level Floyd-Steinberg error-diffusion dithering, a lightweight input transformation, can disrupt adversarial attacks against vision foundation models while preserving semantic content.…

Researchers have developed CoMoGen, a controllable video generation framework that creates realistic motion dynamics and interactions from a single binary mask sequence and an input image. The system uses a lightweight M…

Flow Mismatching, a new unsupervised anomaly detection method, avoids reconstruction-based paradigms by treating flow matching as geometric dynamics and identifying anomalies where learned normal flow disagrees with the …

Researchers have developed IaMSB, an inconsistency-aware multimodal Schrödinger Bridge framework that improves deepfake localization by jointly estimating cross-modal consistency and performing interval-level temporal ev…

Researchers have developed LQ-rPPG, a label-quantized coarse-to-fine learning framework that improves remote photoplethysmography (rPPG) estimation from facial videos by reducing noise in training labels. The framework u…

A team of researchers improved a DETR-based fusion transformer for the MaCVi 2026 Vision-to-Chart data association challenge by adding a dedicated MLP, QueryMLP, to explicitly predict a buoy's waterline contact point in …

New research shows that exact certification of neural networks—determining the smallest set of labeled examples needed to guarantee a learned hypothesis matches the target—becomes exponentially hard under minimal overpar…

A developer has published a conceptual blueprint for AURORA, a low-energy cognitive organism architecture that operates without transformers or token prediction. The system uses sparse predictive microfields, episodic me…

Researchers at HASH built a 25M-parameter transformer decoder, vec2slug, that generates URL slugs directly from vector embeddings, achieving 14-19x faster and 85x cheaper performance than a Haiku-class LLM. The model ext…

Researchers at Apple have developed PICO (Perceptual Image Codec), the first learned image compression model optimized for human visual perception and practical on-device deployment. The codec achieves 2.3-3× bitrate sav…

PyTorch has introduced transparent tracing and compilation for Triton kernels, allowing custom operations to be visible to the compiler for optimization. The framework now supports compiling Triton kernels through `torch…

A deep dive into how large language models develop their personalities reveals a four-stage training process—pretraining, supervised fine-tuning, RLHF, and inference—that fundamentally differs from human cognition. Unlik…