Computer and Information Research Scientists

ABot-Manipulation, a new general VLA foundation model for robot manipulation, has been open-sourced by AMAP CV Lab. It utilizes an innovative action manifold learning paradigm, integrates over 6 million open-source trajectories, supports plug-and-play 3D perception, and shows excellent performance in industry benchmarks.

Machine Heart has upgraded its "SOTA! Model" resource platform, offering access to over 15,000 AI model resources that cover more than 1,000 AI tasks. Users can efficiently search for models based on hardware, computing platform, framework, and model scale.

An interview with AWS Shanghai AI Research Director Zhang Zheng discusses the strategic direction of AI, including insights into the AI4S competition and the release of new programming models like Alibaba's Qwen3.6-Plus. The discussion focuses on finding the right path for AI prosperity.

Microsoft announced the availability of three new world-class MAI models within its Foundry platform, enhancing AI capabilities for enterprise use, workflow automation, and productivity tools, directly impacting AI research and development.

Microsoft has launched "Critique," a new multi-model deep research system within M365 Copilot, aiming to provide advanced research capabilities for professionals in the field.

The article profiles Ruslan Salakhutdinov, a Machine Learning Professor at CMU and Head of AI Research at Apple, detailing his academic background in Machine Learning (Computer Science) from the University of Toronto and his postdoctoral work at MIT AI Lab.

Hugging Face's blog post highlights the development of a fast multilingual Optical Character Recognition (OCR) model, which was built using synthetic data. This advancement provides a new capability in processing diverse textual data and demonstrates an innovative approach to model training.

Xingyun, a startup founded by a former Huawei "genius youth," secured over 400 million RMB in funding to develop a new generation of AI inference chips. The company focuses on creating GPGPU products with ultra-large video memory and CUDA compatibility, using non-3D DRAM architectures to reduce memory costs for large AI models. This technical innovation addresses the memory bottleneck by utilizing lower-cost memory like LPDDR/NAND instead of HBM, emphasizing system-level design and software-hardware coordination to achieve cost-effective and efficient AI inference, particularly for edge devices and consumer electronics.

Salesforce launched "Headless 360," a major architectural shift exposing its entire platform via APIs, MCP tools, and CLI for AI agents. This initiative introduces over 100 new developer tools, including native React support, and Agent Script—a new open-source domain-specific language for defining agent behavior. It also includes new lifecycle management tools for testing, evaluating, and orchestrating AI agents, transforming how developers build on Salesforce by allowing AI agents to operate the system programmatically without a traditional UI.

Hugging Face introduced Ecom-RLVE, a framework for creating adaptive and verifiable environments for e-commerce conversational agents, enhancing model evaluation and development workflows for AI researchers.