Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation

2026-03-19

AI Summary

This work addresses the challenge of enhancing large language models’ performance on complex reasoning and agent tasks—such as mathematics and programming—under strict parameter constraints. The authors propose a multi-domain online policy distillation mechanism that integrates a substantially expanded Cascade Reinforcement Learning (Cascade RL) framework with a Mixture-of-Experts (MoE) architecture, enabling efficient knowledge transfer from state-of-the-art teacher models across domains while activating only 3 billion parameters. The resulting open-source 30B MoE model (with 3B active parameters) achieves near-state-of-the-art performance in mathematical and programming reasoning, becoming the second open-source model—after DeepSeek-V3.2—to attain gold-medal-level proficiency across all three major international competitions: the International Mathematical Olympiad (IMO), International Olympiad in Informatics (IOI), and International Collegiate Programming Contest (ICPC), with approximately 20× improvement in parameter efficiency.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

We introduce Nemotron-Cascade 2, an open 30B MoE model with 3B activated parameters that delivers best-in-class reasoning and strong agentic capabilities. Despite its compact size, its mathematical and coding reasoning performance approaches that of frontier open models. It is the second open-weight LLM, after DeepSeekV3.2-Speciale-671B-A37B, to achieve Gold Medal-level performance in the 2025 International Mathematical Olympiad (IMO), the International Olympiad in Informatics (IOI), and the ICPC World Finals, demonstrating remarkably high intelligence density with 20x fewer parameters. In contrast to Nemotron-Cascade 1, the key technical advancements are as follows. After SFT on a meticulously curated dataset, we substantially expand Cascade RL to cover a much broader spectrum of reasoning and agentic domains. Furthermore, we introduce multi-domain on-policy distillation from the strongest intermediate teacher models for each domain throughout the Cascade RL process, allowing us to efficiently recover benchmark regressions and sustain strong performance gains along the way. We release the collection of model checkpoint and training data.

Problem

• reasoning
• agentic capabilities
• intelligence density
• mathematical reasoning
• coding reasoning

Innovation

• Cascade RL
• Multi-Domain On-Policy Distillation
• Mixture-of-Experts (MoE)
• Intelligence Density
• Open-Weight LLM