AdaRoboVLG introduces a modular framework that decouples task understanding from grasp synthesis, using composable foundation model priors to generate stable, task-appropriate grasps across different robot hands, validated on multiple benchmarks and real-world cluttered and dynamic scenes.
Imperial College London researchers developed an AI model that analyzes standard 10-second ECGs to detect heart failure and valve disease with 81-90% accuracy, identifying subtle patterns humans miss, and spun off Cardiovolt.ai for clinical deployment.
Google's TimesFM-3 introduces native multivariate time series forecasting with non-autoregressive decoding, enabling joint predictions across multiple targets and covariates in a single forward pass, outperforming competitors on benchmarks while offering zero-shot generalization and probabilistic outputs.
The ECCV‑2026 paper introduces DriveTeach‑VLA, a vision‑language‑action model that improves autonomous driving by distilling traffic‑aware visual cues and projecting BEV trajectories onto image pixels, achieving state‑of‑the‑art PDMS scores of 90.4 on NAVSIM and up to 92.7 with a trajectory selector, while detailing the training pipeline, visual distillation, 2D‑TGP prompting, and extensive ablations.
Shanghai AI Lab’s Shushu team launches ArchSpace, an open‑source platform that releases full model checkpoints, training logs, architecture decisions and validation results, inviting the community to submit and verify foundation‑model innovations such as Inter‑layer Attention and Sparse Attention across 1B‑8B scale models, with transparent logs on W&B and HuggingFace.
The article surveys the latest advances in large language model agents, covering multi‑agent collaboration, long‑horizon planning, self‑evolution, trust and safety, test‑time scaling techniques, new foundation and multimodal models, open‑source and closed‑source breakthroughs, world‑model integration, and emerging vertical applications.
This survey examines the emerging field of audio reasoning, distinguishing it from simple audio perception, and systematically classifies four major research directions—Audio-to-Text, Audio-to-Speech, Audio-Visual, and Agentic Audio—while highlighting challenges in data, evaluation, and real‑time multimodal integration.
The PSI Lab at USC, led by Wang Yue, secured three CVPR 2026 awards—Psi‑0, PhysWorld and Humanoid Everyday—each tackling a distinct stage of humanoid robot learning: large‑scale human video pre‑training, embodiment‑aligned fine‑tuning, and physics‑aware world modeling, together forming a coherent data‑model‑prediction pipeline.
In a Tencent Cloud AI conference interview, Yao Shunyu argues that AI is a long‑term, increasingly diverse field, explains the "second half" concept, stresses the importance of context, culture, and trust, and outlines a three‑layer AI strategy covering foundation, product, and frontier research.
The article reviews CVPR 2026 medical imaging papers, highlighting a shift from pure image recognition toward efficient model adaptation, clinical semantic understanding, and cross‑modal reasoning, with examples ranging from simple AI agents optimizing workflows to multimodal foundation models for CT, ultrasound, spatial transcriptomics, IMU‑video alignment, and dual‑view X‑ray analysis.
This article systematically analyzes the most influential open‑source foundation models—Meta Llama, Alibaba Qwen, Mistral AI, and others—detailing their core architectures, lightweight, instruction‑tuned, multimodal, and industry‑specific derivatives, and outlining current ecosystem characteristics and future development trends.
The article details a multi‑agent AI system that transforms manual short‑video creation for cross‑border e‑commerce PPC campaigns into a deterministic, high‑throughput production line, describing its evolution through three phases, the underlying foundation models, modular agent roles, quality‑control mechanisms, and a self‑improving feedback loop.
In a detailed interview, Ted Xiao, former Google DeepMind researcher, walks through the existence‑proof, foundation‑model, and scaling eras of embodied robot learning, explaining the technical challenges, pivotal decisions, and the evolving role of large language and vision models in robotics.
This survey systematically reviews how foundation models reshape embodied navigation, covering problem definition, taxonomy of tasks and robot forms, system architecture from perception to control, data sources and training strategies, edge deployment techniques, benchmark metrics, and future research directions.
This survey reviews over 200 papers to propose a three‑dimensional classification framework for foundation‑agent memory, analyzes paradigm shifts from model‑centric to utility‑centric AI, and outlines memory substrates, cognitive mechanisms, operation strategies, learning paradigms, evaluation metrics, applications, and future research directions.
Researchers at Oak Ridge National Laboratory introduced D‑CHAG, a distributed cross‑channel hierarchical aggregation method that cuts memory consumption by up to 75% and more than doubles throughput when training massive multi‑channel foundation models on up to 1,024 AMD GPUs, as demonstrated on hyperspectral imaging and weather‑forecasting workloads.
The EDEN series of foundation models, trained on the massive BaseData macro‑genomic dataset covering over one million species and 9.7 trillion nucleotides, achieve state‑of‑the‑art genome and protein prediction while enabling functional recombinase design, antimicrobial peptide generation, and synthetic microbiome construction with minimal task‑specific data.
The author explains why he transitioned from search advertising algorithms to foundation model research, outlines the four typical activities of base‑model teams, and shares detailed technical insights, experimental practices, and scaling strategies that led the EvoCUA GUI Agent to achieve open‑source SOTA on OSWorld.
FieldAI, a 2023 startup backed by Bezos, Gates, Nvidia and Intel, has raised over $405 million to develop a physics‑first “general robot brain” (FFMs) that closes the real‑world data gap, leverages NASA‑honed autonomy research, and targets industrial tasks while riding a surge in global robotics investment.
CoRA freezes pretrained time‑series foundation models, extracts multimodal covariate embeddings, evaluates their causal relevance with a trainable Granger‑Causal Embedding, and injects them via a zero‑initialized condition module, achieving up to 31.1% MSE reduction across single‑ and multi‑modal forecasting tasks.
Google Research, X, and Cloud teams introduced Earth AI, a interoperable GeoAI model family that fuses image, population, and environmental data via a Gemini‑driven reasoning Agent, achieving state‑of‑the‑art performance and a 64% reasoning boost over Gemini 2.5 Pro while enabling non‑experts to run real‑time cross‑domain analyses.
A recent breakthrough by Tsinghua researchers introduces AutoGFM, an adaptive graph neural architecture search framework that dramatically improves the performance and generalization of graph foundation models across diverse tasks and domains, as validated by extensive ICML‑2025 experiments.
This article introduces Apple’s Foundation Models framework, explaining its core features, system requirements, and step‑by‑step Swift code for importing the framework, checking model availability, creating sessions, handling generation options, guided output, tool calling, streaming responses, and maintaining conversational context for privacy‑preserving on‑device AI.
This article introduces Apple’s Foundation Models framework for on‑device large language models, explains its advantages and limitations, details technical features like guided generation, snapshot streaming, and tool calling, and provides a complete Swift‑based travel‑itinerary example with performance‑tuning tips.
Apple’s WWDC 2025 revealed iOS 26 (formerly iOS 19) with a sweeping “Liquid Glass” UI redesign, AI‑powered photo tools, a dynamic lock‑screen clock, revamped camera and gallery, plus a new Foundation Models Framework for on‑device AI, while iPadOS gains multi‑window and file‑management features.
Apple unveiled a suite of developer tools at WWDC—including a privacy‑focused Foundation Models framework, Xcode 26 with built‑in LLM support, enhanced App Intents, Swift 6.2 upgrades, Linux container capabilities, and App Store accessibility labels—signaling major shifts for iOS developers.
Apple’s WWDC revealed delayed Siri updates, a new Xcode 26 with built‑in ChatGPT, and a Foundation Models framework, but investors were disappointed, causing a $750 billion market‑cap drop and sparking criticism that Apple’s AI strategy lags behind competitors.
This article critically examines R1‑Zero‑style training by analyzing foundation models and reinforcement learning, uncovering pre‑training and optimization biases, proposing an unbiased Dr. GRPO method, and demonstrating a minimalist 7B‑model recipe that achieves new state‑of‑the‑art performance on AIME 2024.
Physical Intelligence's new Pi-zero model, built on a vision‑language foundation and fine‑tuned with extensive robot data, outperforms prior baselines across multiple tasks, showcasing the promise of large multimodal foundation models for flexible, robust robot control.
This article introduces graph foundation models, explains their relationship with large language models, reviews recent advances in graph neural networks and representation learning, presents the authors' own research on PT‑HGNN, Specformer and GraphTranslator, and discusses challenges, future research directions, and a Q&A session.
A year after the LLM boom, the talent market has moved from severe supply shortage to a high‑water equilibrium, with foundation‑model teams still heavily funded while application‑focused deployments struggle, highlighting changing hiring dynamics, investment pressures, and the looming risk of demand contraction.
This article provides a comprehensive overview of Graph Foundation Models (GFMs), covering their definition, key characteristics, historical development of graph machine learning, recent research trends such as PT‑HGNN, Specformer, and GraphTranslator, and discusses future challenges and research directions.
In a shareholder letter, Amazon CEO Andy Jassy outlines the company’s generative AI strategy, describing it as the most significant technological transformation since cloud computing, detailing AWS’s infrastructure investments, new services like Bedrock, and the broader impact on developers and customers.
TaskMatrix.AI, a Microsoft‑designed AI ecosystem, links large foundation models with millions of APIs through a multimodal conversational model, an API platform, selector and executor, enabling tasks from image processing to robot control while highlighting its learning mechanisms, advantages, use cases, and remaining challenges.
Tool learning enables foundation models to follow human instructions and use external tools, and this article reviews Tsinghua’s research, introduces the WebCPM interactive web‑search QA framework, BMTools and ToolBench packages, and discusses background, frameworks, applications, and future challenges in AI tool usage.
This article reviews the emerging field of tool learning for large foundation models, outlining its background, categorization, core framework components, training strategies, and applications such as WebCPM, BMTools, and ToolBench, while highlighting recent research results and open‑source resources.
This article explores the concept and evolution of large foundation models, their transformative impact on AI-generated content, the underlying technologies such as transformers, diffusion, and CLIP, and discusses the challenges, emerging abilities, and future prospects of these models across multiple modalities.
In this conference talk, He Zhengyu of Ant Group outlines the challenges of large‑model AI, proposes a green, shared, model‑centric infrastructure built on foundation models, cloud‑native MLOps, and Model‑as‑a‑Service (MaaS) to lower cost and accelerate AI adoption across industries.
