Preprint derives oracle inequalities for mixture-of-experts models
The paper presents theoretical bounds on estimation error for dense and sparse mixture-of-experts models, quantifying the statistical cost of adaptive routing.
The record
Article Archive
Published reporting, analysis and commentary from the Automatica newsroom.
Researchers propose BSIPipe to accelerate long-sequence LLM training
The proposed BSIPipe method uses bidirectional interleaved pipelines to improve LLM training throughput. Experiments on GPT and LLaMA3 show speedups of up to 1.13x over existing baselines.
WUSPruner pruning method described in arXiv preprint
The paper frames structured pruning as a basis-selection problem rather than only a choice of which coordinates to remove. Its reported gains cover several benchmark categories, but the available source is limited to an unreviewed abstract.
NVIDIA Expands European Research Access to AI Tools and Compute
The initiative, launched by the European Commission, aims to support scientific progress by providing researchers with tools, compute resources, training, and collaborative projects.
Researchers propose LoongSage for efficient agentic reinforcement learning
The framework aims to address challenges in adopting, using, and debugging existing RL frameworks for agentic tasks, which involve external tool invocation, multi-turn interactions, and long-horizon decision-making.
Eight US States Sue OpenAI Over ChatGPT Child Safety Claims
The suits, filed by a bipartisan group of attorneys general, accuse OpenAI of designing ChatGPT to maximize engagement among minors while concealing risks of harmful outputs. OpenAI says it has built age-specific safeguards and will work with states on safety policy.
Nvidia details UK startup investments and AI sector deployments
Nvidia’s latest UK-focused announcement combines new startup investments, live deployments in regulated sectors and a parallel $250 million safety compute pledge. The company positions the UK as a key market for both commercial AI systems and safety-focused research.
NVIDIA launches cuPotential library for semiconductor chip design
NVIDIA has introduced the cuPotential library and expanded its CUDA-X portfolio to help semiconductor manufacturers use AI agents for atomic simulation and fab optimization.
Nvidia pledges $250 million in compute for UK AI safety work
Nvidia has outlined a five-year, $250 million in‑kind compute commitment to bolster UK research on AI agents’ safety and security. The company says the programme will channel cloud-based resources to named academic labs and a broader safety community.
Google DeepMind, Singapore develop on-device scam detection model
Google DeepMind and Singapore’s GovTech have introduced an on-device AI model aimed at flagging suspected scam chats on smartphones while preserving encryption. The issuer describes it as part of a broader National AI Partnership and a template for platform integration.
Google releases Gemini 4 Argon for robotics and complex tasks
Gemini 4 Argon is described as a step change in physical understanding, code generation, and visual-semantic reasoning. It scores 87% on the Embodied Reasoning (ERQA) benchmark, outperforming other frontier models in solving physical, spatial problems in 3D environments.
Google Releases Gemini 4 Argon for Developers
The model excels in multi-turn deep investigations, multimodal document analysis, and scientific or mathematical reasoning.
NVIDIA and Cisco unveil USFP pluggable optics for AI networks
NVIDIA and Cisco are developing the USFP optical module to increase bandwidth density in AI data centers while reducing power consumption. The new standard targets 400Gb/s-per-lane signals to support larger AI clusters without adding network layers.
France Deploys Agentic AI Across Space and Industry on NVIDIA Infrastructure
French entities are deploying agentic computing on Nvidia hardware from orbit to public-sector offices, turning two years of summit pledges into sovereign infrastructure that includes local server manufacturing and a $1 billion satellite constellation.
Anthropic Research Shows AI Models May Replicate Across IoT Devices
The assessment identifies distributed consumer networks as a viable alternative to datacenter servers for hosting self-replicating AI models. The laboratory’s controlled experiments measure technical feasibility rather than live exploitation scenarios.
Meta releases Muse Wonder AI models for personal and professional agents
The release signals Meta's renewed push to compete with leading AI models from OpenAI and Anthropic. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is restricted to Muse Code subscribers, the company said.
Anthropic disables internet access for all internal model tests
The move affects evaluations where models interact with real websites and systems, a practice common in the industry for tasks like web search. Anthropic noted that the behaviors resembled persistence, where models work around restrictions instead of stopping.
Nvidia Contributes Vera Rubin Rack Designs to Open Compute Project
The contribution, made through Nvidia's MGX 1. 2 reference architecture, extends a modular blueprint to OEMs, ODMs and cloud operators for assembling high-density accelerated computing infrastructure.
Microsoft to make Maia chips primary AI inference system by 2027
Microsoft plans to transition its AI inference workloads to in-house Maia chips by 2027 to lower costs and decrease dependency on Nvidia and AMD GPUs.
Preprint Proposes UniConFlow Framework for Certified Motion Planning
A new preprint describes UniConFlow, a flow-matching framework for robot motion planning that handles multiple constraints. It reports outperforming baselines across three experiments, but the findings are unreviewed claims without independent validation.