A summary of recent AI research papers, research-tool releases, and developer-tool updates.
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.
NVIDIA's BioNeMo Recipes demonstrate how Low-Rank Adaptation (LoRA) can fine-tune large biological foundation models like ESM2-3B for protein structure prediction and Evo2-1B for DNA splice-site classification using only about 1% of parameters. This approach allows training billion-parameter models on a single NVIDIA RTX 6000 GPU in under an hour, significantly reducing computational and storage demands. The recipes integrate performance optimizations such as NVIDIA Transformer Engine and sequence packing to maxim...
Keywords: drug discovery, fine-tuning, Foundation Models, NVIDIA, protein · score 38.9
Source Mix
Interleaved papers, lab updates, and research-tool signals.
EconCSLib is a new Lean 4 library that integrates large language models to assist researchers in formalizing economics and computation papers by generating Lean code verified by humans. This tool preserves the original research structure while simplifying the conversion of paper claims into formal proofs, improving both rigor and efficiency. It represents a novel combination of AI and formal methods to support acade...
SpaceX and Google have entered a $30 billion deal to collaborate on AI infrastructure, signaling a major partnership despite their competitive history. This collaboration aims to leverage SpaceX's satellite network and Google's cloud capabilities to enhance AI computing power and data accessibility. The deal marks a significant move in the AI and tech industries, combining strengths from two leading companies.
Keywords: AI infrastructure, infrastructure deal · score 23.9
Xiaomi has released a HyperOS update that includes the June security patch for several of its smartphones. This update aims to enhance device security and system stability across the supported models. Users of these Xiaomi phones can now benefit from improved protection against vulnerabilities.
Tesla is reportedly considering deploying a fleet of up to 5,000 robotaxis in the Las Vegas region. This move would mark a significant expansion of Tesla's autonomous vehicle services, leveraging its self-driving technology in a major urban market. Details on timelines and operational specifics remain limited.
The article outlines 13 significant instances where NVIDIA CEO Jensen Huang's announcements and actions have notably impacted technology markets. These moments underscore his role in shaping trends in AI, hardware, and computing industries. However, specific details on each event are not provided in the source.
World-Action Models (WAMs), which leverage pretrained video or world-model backbones to predict both future states and robot actions, are rapidly gaining traction as a complementary approach to Vision-Language-Action (VLA) models in robotics. Unlike VLAs that rely on vision-language models and face challenges in grounding language to action, WAMs offer a promising alternative by modeling scene dynamics directly from...
Researchers have developed APPO, a reinforcement learning technique that improves large language model agents by accurately assigning credit to intermediate actions instead of relying on broad heuristics. This method optimizes multi-turn tool use and decision branching, potentially boosting AI agents' task performance and procedural reasoning. The approach is detailed in a recent preprint on arXiv.
Keywords: agentic, agents, large language model, tool use · score 25.1
Qualcomm's Dragonfly AI initiative has failed to capture market attention as Nvidia's high-profile Computex presentations dominate headlines, leading to a decline in Qualcomm's stock price premarket. Nvidia's aggressive product launches and AI advancements at Computex have set a new benchmark in the AI hardware space, overshadowing Qualcomm's developments. This shift highlights the competitive dynamics in AI chip in...
NVIDIA has developed advanced fused MLP kernels using its CuTe DSL to significantly boost training throughput for mixture-of-experts (MoE) AI models. These kernels address key bottlenecks by fusing GLU activation functions with GEMM operations, eliminating CPU synchronization overhead, and supporting low-precision quantization, resulting in up to 2x kernel-level speedups and up to 93% end-to-end training improvement...
Researchers have developed EvoArena, a new benchmark suite that evaluates how large language model (LLM) agents adapt and update their knowledge in changing environments across terminal, software, and social domains. Unlike static tests, EvoArena simulates progressive environmental changes to assess continual learning and memory robustness over time. This tool provides valuable insights into LLM performance in real-...
Keywords: agents, benchmark, large language model, LLM · score 21.8
EurekAgent introduces a novel approach to autonomous scientific discovery by optimizing the environments in which AI agents operate, rather than improving the agents themselves. Utilizing large language model-based agents, this method automates the proposal, validation, and iteration of scientific solutions, potentially outperforming human-designed methods. The research highlights environment engineering as the crit...
Keywords: agents, large language model, scientific discovery · score 21.8
Agents-K1 is a new AI pipeline that transforms raw scientific documents into detailed knowledge graphs, capturing entities, claims, evidence, and methodological lineages often missed by existing models. This approach enables AI agents to orchestrate scientific knowledge more deeply than traditional summaries or citation networks, enhancing their scientific reasoning capabilities. Presented as a preprint on arXiv, th...