The paper formalizes a notion of evaluability that drops the classical held-out test set assumption, asking whether black-box models can still be evaluated when test samples may have appeared in training. The authors fully characterize the distribution classes admitting accurate evaluation and prove tight upper and lower bounds on sample complexity, showing black-box evaluation is feasible if and only if the distribution is close to a specific condition.
Foundation Model
基础模型
Large models pre-trained on massive data, adaptable to various downstream tasks
Curated Intelligence
Baidu released ERNIE-5.5-Flash-Base on Hugging Face, a mixture-of-experts foundation model with 280B total parameters and 14B active per token, pre-trained on 33.5 trillion tokens with a 128K training context that can extend to one million tokens during fine-tuning and inference. It ships deliberately without fine-tuning or post-training, positioning the raw weights for developer customization, with a post-trained version to follow.
Agibot unveiled full-stack embodied AI advancements covering hardware, motion, manipulation, and interaction intelligence at its AI WAVE event, proposing a 'Five Leaps' theory for mature physical AI. The company argues that general-purpose LLMs cannot replace specialized embodied models, emphasizing a software-hardware integrated approach validated through real-world deployments (e.g., a 24-hour convenience store live stream) and introducing 'Worken' to quantify effective intelligent labor hours.
Xiaomi Group received two awards at the 2026 World Internet Conference. Its embodied foundation model demonstrated an 8-hour continuous operation capability in automotive factories and completed long-horizon multi-room task verification in a residential setting. This marks a significant step in closing the loop between technology and real-world deployment for both industrial and home scenarios.
Google Research leverages foundational models to enhance tuberculosis (TB) radiology screening, addressing the critical shortage of specialists in high-burden regions. The initiative evolves from specific AI tools to a more generalizable model architecture, aiming to improve early detection capabilities and healthcare equity in remote and resource-limited settings.
AVEVA's Chief Technologist highlights that as foundation models and agentic AI enter industrial environments, ensuring safety and reliability in physical system interactions becomes a central challenge. The article emphasizes integrating multi-source data like telemetry and logs to enhance real-time decision-making, aiming to address the unpredictability of new AI systems and drive responsible industrial automation.
Biohub, backed by Mark Zuckerberg and Priscilla Chan, is coordinating a $1.8 billion initiative to build AI models that predict cell behavior for drug development. Key partners including Meta, Google DeepMind, Isomorphic Labs, and the US Department of Energy are contributing funds and infrastructure, with commercial funders receiving one year of exclusive data access before public release. The first standardized dataset is expected within a year.
Google introduces planetary geospatial foundation models via Earth AI to address data lags and silos in traditional epidemiological surveillance. By leveraging population dynamics data, these models offer a general-purpose solution for disease outbreak prediction and resource allocation without custom engineering pipelines, with limited free access for academic researchers.
Researchers introduce GlucoFM, a foundation model for Continuous Glucose Monitoring (CGM). It employs a novel dual-stream architecture to explicitly separate slow baseline trends from transient event dynamics, addressing the limitations of existing single-stream models in interpreting complex physiological signals with sparse labels.
Google introduces SensorFM to address the challenges of generalizing wearable health data due to high inter-individual variability and scarce labels. By leveraging self-supervised learning to build universal health representations, the model supports automatic specialization and serves as the core interface for Personal Health Agents, bridging raw physiological signals with personalized preventive care.