AI Research Radar is an autonomous tool for tracking AI news related to research and coding. It collects signals such as new papers, model and developer-tool updates, research-writing tools, AI-for-science work, mathematical reasoning, robotics, hardware, and selected company news. The results are filtered through a set of research-oriented keywords.
20 visible items
Featured result
OpenAI has introduced Prism, an AI-powered LaTeX editor that operates directly within web browsers. This tool aims to streamline the creation and editing of LaTeX documents by leveraging AI capabilities for enhanced user experience. Prism integrates AI assistance to simplify complex typesetting tasks traditionally associated with LaTeX.
Interleaved papers, lab updates, and research-tool signals.
NVIDIA has introduced Vera, a new class of CPU designed specifically for the agentic AI era, delivering up to 1.8x sustained per-core performance over traditional x86 CPUs. Vera features NVIDIA's Olympus core with 50% higher instructions per cycle and supports up to 88 cores with high memory bandwidth, enabling faster AI agent loops critical for reasoning and task execution. AI innovators like Perplexity are adoptin...
OpenAI has released the GPT-5.6 model family, including specialized versions for frontier capability, cost-efficiency, and high-volume workloads, alongside new features like Programmatic Tool Calling, explicit prompt caching, and multi-agent orchestration in beta. The update also introduces GPT-Realtime-2.1 models for improved real-time voice applications, a Safety Usage Dashboard, and enhanced moderation tools, whi...
Researchers have introduced MIRA-Math, a new benchmark designed to test AI's ability to solve mathematical problems when exactly one critical fact is missing. The AI must request this missing information in natural language within a strict limit and then use it to produce an exact answer. This approach isolates the capability of minimal information requesting combined with precise mathematical reasoning.
NVIDIA has enhanced the performance of molecular dynamics simulations by enabling GPU-initiated communication in GROMACS, a widely used MD package. By replacing CPU-orchestrated MPI communication with GPU-native NVSHMEM, the new approach eliminates CPU-GPU handoffs, significantly reducing latency and improving scalability across multiple GPUs. This innovation allows simulations to run faster and more efficiently, re...
AMD Ryzen processors are increasingly being recognized for their growing impact in agentic AI technologies, suggesting potential for further advancements and market growth. This development highlights AMD's strengthening position in AI hardware beyond traditional computing tasks. However, specific details on performance improvements or product launches remain limited.
Tesla Inc. is gaining attention as a top robotics stock due to its advancements in robotaxi technology and the development of the Optimus humanoid robot. These initiatives highlight Tesla's expanding role beyond electric vehicles into robotics and AI-driven automation. The company's progress in these areas positions it at the forefront of robotics innovation.
Microsoft Research has introduced Flint, an open-source visualization language designed to help AI agents generate expressive and polished charts from compact, human-editable specifications. Flint automates complex design decisions by inferring semantic data types and chart parameters, enabling reliable and attractive visualizations across multiple backends like Vega-Lite, Apache ECharts, and Chart.js. This approach...
Xiaomi has begun releasing the HyperOS 3.3 beta update for its Xiaomi 17 and 15T Pro smartphones, laying the groundwork for compatibility with the upcoming Android 17 and HyperOS 4 operating systems. This update signals Xiaomi's commitment to keeping its flagship devices current with the latest software advancements. The beta phase allows users to experience new features and improvements ahead of the official releas...
OpenAI's recent analysis has uncovered significant reliability and accuracy concerns in SWE-Bench Pro, a widely used coding benchmark for evaluating AI models. This finding calls into question the dependability of performance assessments based on this benchmark. The study highlights the need for more robust evaluation tools in AI coding performance measurement.
NVIDIA and Hugging Face highlight the critical role of open and synthetic data in developing robust AI agents capable of real-world interactions beyond static benchmarks. NVIDIA's Nemotron open data products, including over 10 trillion pre-training tokens and millions of post-training samples, enable transparent, inspectable agent behaviors and foster a diverse AI ecosystem. Synthetic data helps preserve proprietary...
OpenAI has integrated its Codex coding agent into the ChatGPT desktop app for macOS and Windows, allowing users to maintain their projects and workflows seamlessly. The update includes general availability of Codex Remote for mobile control, new plugins like DigitalOcean provisioning, expanded regional access in Europe, and the introduction of advanced models such as GPT-5.5 for complex coding tasks. Additional feat...
Anthropic and AE Studio have introduced GRAM, a novel method that compartmentalizes dual-use knowledge—such as virology and cybersecurity—within AI models, allowing selective removal or retention of sensitive capabilities without retraining multiple models. Tested on models up to 5 billion parameters, GRAM effectively isolates and deletes specific knowledge modules without degrading overall performance, offering a m...
Google DeepMind's Accelerator programs support over 2,000 AI-focused startups, developers, and non-profits across 88 countries by providing access to advanced AI tools, expert mentorship, and strategic guidance. These equity-free programs help bridge the gap between research and market-ready AI solutions, with a focus on addressing critical environmental and societal issues. Participants also gain opportunities to s...
Former President Trump has approved Anthropic's expansion to global markets, positioning the AI startup as a significant competitor to Google DeepMind. This move could intensify competition in the AI research and development space. Details on the scope and timeline of Anthropic's global operations remain limited.
NVIDIA Nemotron 3 Ultra, tuned by LangChain's Deep Agents harness, achieves the highest accuracy among open AI models while running at 10 times lower inference cost than leading closed models. This performance boost was achieved without retraining the model, relying instead on engineering the environment around it, enabling faster, more cost-effective AI agent deployment. Enterprises like Abridge, Amdocs, Box, and E...
OpenAI has introduced a Computer Use feature in the ChatGPT desktop app that allows the AI to visually interact with and operate graphical user interfaces on macOS and Windows. This capability enables ChatGPT to perform complex tasks beyond command-line tools, such as navigating desktop apps, adjusting settings, and reproducing GUI-specific bugs. Users must grant explicit permissions and can control which apps ChatG...
A new study introduces institutional red-teaming, a method that isolates the effect of deployment rules on multi-agent AI safety by holding agents and tasks constant while varying rules. Testing across 33,924 games showed that changing consequence rules can shift fatality rates by up to 58 percentage points, with identity-targeting rules consistently causing unsafe outcomes. The research highlights that not just AI...
NVIDIA has developed a method to improve the accuracy of LangChain Deep Agents by creating a harness profile tailored for the Nemotron 3 Ultra model, enabling better performance without the need for costly fine-tuning. This approach automates the tuning of agent harnesses to more closely match model training data, using iterative testing and adjustments to optimize results. The process leverages existing NVIDIA clou...
Researchers have developed MedPMC, a systematic framework designed to scale high-quality, multimodal medical data from PubMed Central to enhance foundation models in medicine. This framework addresses limitations in existing datasets by improving data fidelity, reproducibility, and clinical validation, enabling better integration of diverse clinical information. MedPMC aims to support the creation of more accurate a...