arXiv cs.LG

Position: Explainability Research Must Prioritize Foundations over Ad-hoc Methods

Michal Moshkovitz, Suraj Srinivas, Lesia Semenova…

This XAI position paper points out that while fragmented explainable AI methods are emerging constantly, very few can be implemented to guide actual decision-making, with the core issue being the lack of a basic research framework adapted to the full human-in-the-loop process. By analyzing related papers from three top AI conferences and surveying practitioners, the authors confirm that XAI generally has common problems such as vague problem definitions, inaccurate evaluation, and missing feedback loops. They call on the academic community to prioritize solving basic structural challenges, and also provide practical implementation guidelines.

CARPRT: Class-Aware Zero-Shot Prompt Reweighting for Black-Box Vision-Language Models

Ruijiang Dong, Zesheng Ye, Jianzhong Qi…

For the zero-shot classification scenario of black-box pre-trained vision-language models, existing prompt ensembles use class-shared weights, which do not adapt to the prompt preferences of different classes. This paper proposes CARPRT, a training-free class-aware zero-shot prompt reweighting method that calculates exclusive adaptation weights for different prompts individually for each class. Experiments show that it outperforms class-agnostic reweighting schemes on standard classification benchmarks, verifying the effectiveness of modeling prompt-class dependencies.

Explainable Geospatial AI for Satellite Ground Station Siting Using LiDAR-Derived Terrain Intelligence

Shohini Sarkar, Smithi Mahendran, Rishi Chudasama…

To address the problem that existing ITU standards for estimating ground clutter height (RCH) do not consider differences between similar ground features, leading to large errors in LEO satellite ground station siting, this paper proposes an explainable geospatial AI framework: based on LiDAR-annotated RCH data, it fuses multi-source public geographic remote sensing data to train a LightGBM model, paired with SHAP explainable analysis. The model reduces error by more than 60% compared to the ITU baseline, with an MAE of 1.79 meters and R² of 0.765. It identifies core influencing factors such as canopy coverage, balancing both accuracy and global deployability.

OpenAI

A scorecard for the AI age

OpenAI

OpenAI CFO Sarah Friar has launched a practical quantitative scorecard adapted to the AI era for AI implementation scenarios, used to measure the return on investment of AI spending. The scorecard sets four core evaluation dimensions: effective workload, cost per successful task, system reliability, and return on computing power investment. It solves the pain point of difficulty in quantifying input and output in AI deployment, and provides clear, implementable judgment standards for enterprises to evaluate the actual value of AI applications.

Why teens deserve access to safe AI

OpenAI

This article responds to teenagers’ demands for access to safe AI services, and introduces a series of safety measures launched by OpenAI to adapt ChatGPT for teenage use: setting age-appropriate content protection rules, launching exclusive learning assistance tools, opening up parental control functions, while collaborating with domain experts to iterate on protection mechanisms. On the premise of building a solid safety barrier, it ensures that teenagers can fully access the learning support brought by AI.

Anthropic News

Inviting hard questions

Anthropic

This is a public interaction project initiated by the AI research team: the team has opened a submission channel to the whole society, specifically collecting the most controversial and challenging difficult questions about the AI field from the public. At the same time, the team has made a public commitment that when conducting research and giving responses to the collected questions in the future, it will fully disclose the entire process of research and demonstration, ensure work transparency, and proactively accept public supervision.

Redeploying Claude Fable 5

Anthropic

Anthropic officially announced that with the lifting of relevant export control measures, it will restart the deployment of the Claude Fable 5 large model starting July 1. The version launched this time has completed security upgrades, is equipped with an updated cybersecurity protection system, and has also added a jailbreak prevention framework adapted to industry scenarios. It can better respond to security risks in various usage scenarios under the premise of compliance, and ensure the stability and safety of model implementation.

Google DeepMind

Our approach to bioresilience

Google DeepMind

Google DeepMind and Isomorphic Labs recently released their joint technical path and exclusive AI model in the field of bioresilience. Bioresilience research aims to improve human society’s ability to warn, respond to, and recover from biological risks such as new infectious diseases and sudden biosafety incidents. The AI solution announced this time will provide important technical support for the intelligent upgrading of the global biological risk prevention and control system.

Empowering India’s next generation of innovators with ATL Saathi

Google DeepMind

Google, together with India’s Atal Innovation Mission (AIM), has jointly launched the AI tool ATL Saathi. Developed based on the Gemini large model, this tool is intended for use by teachers in Indian campus robotics science and innovation laboratories. Its core function is to lower the threshold for science and innovation experiment teaching, assist teachers in better carrying out science and innovation guidance for teenagers, and help cultivate the next generation of local innovative talents in India.

Hugging Face Blog

Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers

Hugging Face

This paper proposes a large-scale fine-tuning solution for vision models that combines NVIDIA NeMo Automodel and the Hugging Face Diffusers library. It unifies the capabilities of the two frameworks, optimizes the distributed multi-GPU/multi-node training process, and is compatible with multiple types of diffusion models for text-to-image, text-to-video, and other tasks. This solution greatly lowers the threshold for fine-tuning large-scale visual generative models, significantly improves training efficiency, and can quickly meet customized generation needs in vertical scenarios.

Newer Models, Same Advantage

Hugging Face

This paper titled Newer Models, Same Advantage focuses on the problem of group performance bias in large models. It constructs a standardized question-and-answer test set covering different demographic characteristics, and compares the task accuracy of multiple generations of old and new models including the GPT and Claude series. The study found that model iteration has not narrowed the performance gap between different groups: the performance advantages of specific groups that existed in older versions still remain stable in new-generation models, and biases have not been effectively alleviated with model updates.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment research is based on the perspective of virtue ethics, refuting the classic orthogonality hypothesis that intelligence and goals are independent: human rational actions are not anchored to preset ultimate goals, but rely on normative calibration of behavior through the network of action practice. To achieve AI that is safe, controllable, and can collaborate with humans, its decision-making logic must match the human practice-based action paradigm. This path is compatible with both ethical alignment and core security requirements at the same time.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of Recursive Self-Improvement (RSI) can be traced back to the ultra-intelligent machine concept proposed by I. J. Good in 1965, referring to a system that can surpass all human intellectual activities and autonomously design better machines to achieve self-iteration. In 2008, Eliezer Yudkowsky clearly defined its core as a feedback loop: AI uses existing intelligence to optimize the cognitive mechanism that produces its own intelligence. Current RSI in the AI field can refer both to models directly rewriting their own weights, and broadly extend to models optimizing their own training pipelines.

量子位

全球首款720°连续后空翻机器狗来了！宇泛智能携“灵猫”双馆联袂首秀WAIC

量子位

At the 2026 World Artificial Intelligence Conference (WAIC), Yufan Intelligence’s full-stack self-developed quadruped robot “Lingmao” made its debut in two pavilions. As the world’s first quadruped robot that can complete 720° continuous backflips, 15 “Lingmao” units performed difficult movements such as individual stunts and group synchronized dancing, demonstrating its technological accumulation in motion control, cluster scheduling, etc., echoing the conference theme and exploring the industrial implementation path of embodied intelligence.

PPIO发布智能模型网关，打造面向Agent时代的智能Token工厂

量子位

During WAIC 2026, PPIO released its Agentic Cloud positioning and intelligent model gateway, building an intelligent Token factory for the agent era. Based on its self-developed core formula for agent productivity, it has built a two-layer product system: the intelligent model gateway reduces costs, increases efficiency, and raises the intelligent density of Tokens, while the Agent Harness layer extends the running time of agents, creating a full-stack cloud service natively adapted for agents, with related business growing rapidly in recent years.

逛完WAIC 2026我悟了：国产AI芯片的真对手，根本不是英伟达的GPU

量子位

After visiting WAIC 2026, it can be seen that the core competitive barrier for domestic AI chips is not NVIDIA GPU hardware, but the development habits, engineering standards, and migration costs accumulated by the CUDA ecosystem. The industry competition logic has shifted from comparing individual chip parameters to building full-stack system capabilities. Manufacturers such as Qingwei Intelligence have explored an ecological path that integrates architecture, chips, software stacks, and applications, focusing on solving customer needs for low-cost migration, stable use, and scalability, which is the core direction of domestic substitution.