Nemotron-Cascade 2: Post-Training LLMs with Cascade RL (research.nvidia.com)

Nemotron-Cascade 2 has been introduced as an advanced open-source large language model (LLM) with 30 billion parameters, optimized using a post-training approach that integrates Cascade Reinforcement Learning (RL). Building on the foundation of the Nemotron-Nano-V3 model, it achieves remarkable performance in mathematical reasoning and coding, attaining gold-medal level in prestigious competitions such as the International Mathematical Olympiad and the International Olympiad in Informatics. This accomplishment is noteworthy as it demonstrates high intelligence density, delivering strong capabilities with significantly fewer parameters compared to larger models. The enhancements in Nemotron-Cascade 2 over its predecessor include an expanded Cascade RL framework that covers diverse reasoning and agentic domains, along with the innovative use of multi-domain on-policy distillation. This method allows the model to leverage insights from top-performing teacher models throughout the training, effectively mitigating performance regressions and enhancing overall capabilities across various tasks. The release of this model, along with its checkpoints and training datasets, represents a significant advancement for the AI/ML community, highlighting the increasing efficiency and effectiveness of LLMs in addressing complex reasoning challenges.