LLM Training
New research introduces neuron-level activation functions that leverage 2:4 structured sparsity to dramatically accelerate LLM pre-training while maintaining model quality.
Synthetic Data
NVIDIA partners with Singapore to create culturally-aware synthetic training data using Nemotron personas, advancing sovereign AI development through co-designed data generation methodologies.
Reinforcement Learning
LinkedIn shares practical insights from training agentic reinforcement learning systems for GPT-OSS, covering infrastructure challenges, reward modeling, and lessons learned from real-world deployment.
LLM Training
New research introduces SIGMA, a scalable spectral method using eigenvalue analysis to detect model collapse during LLM training before performance degrades catastrophically.
LLM Training
Modern LLMs undergo three critical training stages: pretraining on massive text corpora, supervised fine-tuning for instruction following, and RLHF for alignment. Here's how the pipeline works.
LLM Training
New EDGC method uses entropy to dynamically compress gradients during LLM training, reducing communication overhead while preserving model accuracy across distributed systems.
LLM Training
New research compares three reinforcement learning approaches for enhancing LLM reasoning capabilities, offering insights into parametric tuning strategies for PPO, GRPO, and DAPO algorithms.
LLM Training
Microsoft's DeepSpeed optimization library transforms large language model training through ZeRO memory optimization, 3D parallelism, and infrastructure innovations that make training trillion-parameter models feasible on consumer hardware.