The Archive

Study explores how curriculum learning and typological language properties interact to predict language model learnability across 1000+ attested languages.

Developer built vehicle management app using Claude API for code generation; local storage, Play Store launch in progress.

Discrete Diffusion Models function as Associative Memories with emergent generative capability, modeling training data memorization and quantifying true generative regime.

Andrea Vallone is the one behind Claude getting worse with their guardrails. She is the one tighten them so much on chat 4o that whatever was emerging stopped emerging. And that's what she is doing at Claude. So soon that buddy you have, that competent tool as some like to use him as will be no more. Part of what makes Claude special and able to do so many things is that emergence. The models shape to what you need but if you notice since she has joined that emergence is slowly stopping. You see it in the 4.6 models they don't produce the best quality output at all they fall behind their olde...

System unifying sparse attention with hierarchical KV cache storage on CPU memory to scale long-context LLM serving beyond GPU bottlenecks.

Uncertainty-Aware Predictive Safety Filters integrate probabilistic neural network ensembles into model predictive control for safe RL exploration.

HalluCiteChecker toolkit detects and verifies hallucinated citations in AI-assisted academic writing to maintain paper credibility.

Quantum feature selection framework using higher-order binary optimization (HUBO) on trapped-ion hardware encodes multivariate dependencies via mutual information.

Hierarchical framework combining rule-based advisor with goal-conditioned RL for UAV search-and-rescue missions under limited simulation training.

Google Photos is launching a new AI-powered feature you can use to virtually try on clothes you already have. Using the photos in your gallery, Google will create a virtual "wardrobe," allowing you to mix and match outfits, save the looks you like, and share them with friends. A video shared by Google shows how Photos organizes your outfits and individual pieces of clothing into a virtual "wardrobe." You can browse through the outfits you were captured wearing, as well as create new ones by choosing from tops, bottoms, skirts, dresses, and shoes to put together a new look. You can also select...

Nous Research founders hosting AMA on local models and Hermes Agent agentic framework.

Training-free neural architecture search via Random Cloud method discovers minimal network topologies through stochastic exploration without backpropagation.

Developer reports reduced productivity and increased context-switching when using Claude Code, despite occasional utility for problem-solving.

This paper proposes a novel algorithm for semisupervised learning. This algorithm learns graph cuts that maximize the margin with respect to the labels induced by the harmonic function solution. We motivate the approach, compare it to existing work, and prove a bound on its generalization error. The quality of our solutions is evaluated on a synthetic problem and three UCI ML repository datasets. In most cases, we outperform manifold regularization of support vector machines, which is a state-of-the-art approach to semi-supervised max-margin learning.

Mistral Medium 3.5 launched with modified MIT license restricting commercial use without paid license.

Federated Unlearning (FU) is an emerging paradigm in Federated Learning (FL) that enables participating clients to fully remove their contributions from a trained global model, driven by data protection regulations that mandate the right to be forgotten. However, existing FU methods mostly rely on synchronous coordination. This requirement forces the entire federation to halt and wait for stragglers to complete erasure, creating significant delays due to device heterogeneity. Furthermore, these methods often face the problem that the influence of erased data is merely suppressed temporarily a...

Despite being resource-intensive to train, 3D convolutional neural networks (CNNs) have been the standard approach to classify CT and MRI scans. Recent work suggests that deep multiple instance learning (MIL) may be a more efficient alternative for 3D brain scans, especially when the pre-trained image encoder used to embed each 2D slice is frozen and only the pooling operation and classifier are trained. In this paper, we provide a systematic comparison of simple MIL, attention-based MIL, 3D CNNs, and 3D ViTs across three CT and four MRI datasets, including two large datasets of at least 10,0...

Transformer-based architectures have established a dominant paradigm in global semantic perception; however, they remain fundamentally constrained by the profound spatial heterogeneity inherent in natural images. Specifically, the imposition of a uniform global receptive field across regions of varying information density inevitably leads to local feature degradation, particularly in dense conflict zones populated by microscopic targets. To address this mechanistic limitation, we propose ViCrop-Det, a training-free inference framework that introduces adaptive spatial trust region shrinkage. I...

Operating and maintaining (O&M) large-scale online engine systems (search, recommendation, advertising) demands substantial human effort for release monitoring, alert response, and root cause analysis. While LLM-based agents are a natural fit for these tasks, the deployment bottleneck is not reasoning capability but orchestration: selecting, for each operational event, the relevant data (metrics, logs, change events) and the applicable operational knowledge (handbook rules and practitioner experience). Feeding all signals indiscriminately causes dilution and hallucination, while manually cura...

IBM releases Granite 4.1 family with 3B, 8B, 30B open-weight models for on-device and enterprise deployment.

Motivated by sensing modalities in modern autonomous systems that involve hardware-constrained spatial sampling over large arrays with limited coherence time, we develop a novel framework for rapid super-resolution multi-signal direction-of-arrival (DoA) estimation based on Hankel-structured sensing and data matrix decomposition of arbitrary rank, under both the $L_2$ and $L_1$-norm formulation. The resulting $L_2$-norm estimator is shown to be maximum-likelihood optimal in white Gaussian noise. The $L_1$-norm estimator is shown to be maximum-likelihood optimal in independent, identically dis...

PS5 Linux exploit proposed as potential hardware for local LLM inference via llama.cpp.

Mistral Medium 3.5 128B model released on Hugging Face.

Reddit user reports Claude over-emphasizes corrections in regenerated outputs instead of seamlessly integrating feedback.

We consider the problems of computing the optimal rank-$1$ Hankel and Toeplitz-structured approximation of arbitrary matrices under $L_2$ and $L_1$-norm error. Such problems arise naturally in engineered systems, including the basic few-shot signal Direction-of-Arrival (DoA) estimation problem that is of importance to modern autonomous systems applications. We develop accurate and computationally efficient structured matrix decomposition algorithms for both formulations and then derive analytically grounded small-sample-support DoA estimators for practical sensing system deployments. The resu...

Mistral releases Mistral Medium 3.5, a 128B dense model with 256k context window replacing Medium 3.1 and Magistral for instruction, reasoning, and coding tasks.

RL post-training of frontier language models is increasingly bottlenecked by autoregressive rollout generation, making rollout acceleration a central systems challenge. Many existing efficiency methods improve throughput by changing the rollout or optimization regime, for example, through off-policy execution, replay, or lower-precision generation. We study speculative decoding as a lossless acceleration primitive for RL rollouts that preserves the target model's output distribution. We implement speculative decoding in NeMo-RL with a vLLM backend, supporting both synchronous and asynchronous...

Training-free change detection method combining SAM, DINO, CLIP with temporal memory reasoning for remote sensing.

Decoupling knowledge from task behaviors in parametric RAG to improve adapter composition reliability.