June 2026 monthly work summary focusing on MoE routing reproducibility and RL training pipeline enhancements. Delivered Router Replay (R3) to ensure consistent MoE expert assignments across rollout and training, with validation, fallbacks for missing routes, and comprehensive tracing. Extended NeMo Gym to propagate routed expert indices through training outputs to enable downstream reinforcement learning tasks and router replay in async rollouts.

April 2026 monthly summary for NVIDIA-NeMo/Automodel: Delivered end-to-end supervised fine-tuning support for discrete diffusion LLMs (dLLMs), establishing training recipes, data handling procedures, and loss functions to enable rapid domain adaptation and customization. No major bugs fixed this month; focus was on feature delivery and pipeline robustness. The work aligns with Automodel goals to empower customers with flexible, trainable discrete diffusion models.