The auto-updated technology brief
Reinforcement learning (RL) is key to aligning language models, from RLHF in AI assistants to newer RLVR workflows for reasoning and agent tasks, enabling more accurate domain-specific AI agents.
NVIDIA introduces the BioNeMo Agent Toolkit, enabling AI scientists to assist in life science discovery by reading papers, generating hypotheses, writing code, and calling APIs in an iterative scientific process.
NVIDIA introduces Nemotron 3 Ultra using NVFP4 4-bit floating point quantization with Model Optimizer to efficiently move large model weights for longer context windows.
NVIDIA Omniverse NuRec is a neural reconstruction pipeline that builds high-fidelity 3D representations of real-world environments from multisensor data, optimized using NVIDIA Nsight developer tools.
Training large language models (LLMs) at massive scale faces infrastructure challenges due to long runtimes and thousands of GPUs. Nonuniform tensor parallelism helps mitigate slowdowns caused by device unavailability and resource fluctuations.
NVIDIA researchers have developed an Ising decoding method that reduces logical error rates in quantum color codes by over 300 times, advancing fault-tolerant quantum computing.
NVIDIA's NeMo framework addresses data limitations in financial NLP by generating synthetic data to improve LLM fine-tuning for trading research, risk modeling, and surveillance.
A team at Beijing Institute of Technology developed night vision goggles that translate infrared wavelengths into distinct visible colors, providing a more natural visual experience.
NVIDIA's BioNeMo Agent Toolkit enhances the speed and scalability of biomolecular structure prediction and co-folding workflows using OpenFold3, benefiting drug discovery and protein design.
Robotics foundation models have advanced to follow natural language instructions for manipulating objects, but rigorous evaluation remains a key challenge. This article introduces the main problems and a method to address them.
A study from Nanjing University shows recycling could become the main source of key materials for EV manufacturing in China by 2050, despite challenges from evolving battery chemistries.
This article discusses the balance of accuracy, throughput, and interactivity in AI performance, focusing on how model design choices impact throughput and interactivity in large language models.