Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models
Nemotron-Cascade: Scaling Reasoning Models with Cascaded RL
Our innovative Cascade RL approach orchestrates sequential, domain-wise reinforcement learning, drastically reducing engineering complexity and delivering unparalleled performance across diverse benchmarks. Achieve silver-medal status in IOI 2025 and outperform leading models in complex code and math challenges.
Unlocking Next-Gen AI Performance
Nemotron-Cascade models set new benchmarks across critical enterprise AI domains, demonstrating superior alignment and reasoning capabilities.
0
IOI 2025 Total Score (Silver Medal)
0
LiveCodeBench v6 Pass@1 (14B)
0
IFEval Strict Prompt (8B Unified)
0
SWE-bench Verified (14B)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Supervised Fine-Tuning (SFT)
Cascade Reinforcement Learning (RL)
Competitive Coding Deep Dive
RLHF & Model Alignment
Software Engineering (SWE) RL
Supervised Fine-Tuning (SFT)
The foundational stage, equipping models with core skills across general domains, math, coding, science, and tool use. Critical for establishing robust conversational abilities and reasoning capabilities before RL refinement.
Up to 32K
Max Token Length in Stage 2 SFT
Cascade Reinforcement Learning (RL)
Our innovative sequential, domain-wise RL approach. This strategy significantly enhances reasoning performance and alignment while mitigating catastrophic forgetting. It adapts to unique domain requirements, from human feedback to verifiable rewards.
Enterprise Process Flow
Multi-Stage SFT
→
RLHF
→
Instruction-Following RL
→
Math RL
→
Code RL
→
SWE RL
→
Nemotron-Cascade Model
| Benchmark | DeepSeek-R1-0528 (SFT Teacher) | Nemotron-Cascade-8B (RL) | Nemotron-Cascade-14B-Thinking (RL) |
|---|---|---|---|
| MMLU-Pro (EM) | 85.0 | 75.7 ↑1.3 | 77.0 ↑1.0 |
| IFEval (strict prompt) | 84.1 | 90.2 ↑19.4 | 81.9 ↑12.1 |
| AIME 2025 (no tools) | 87.5 | 80.1 ↑7.3 | 83.3 ↑2.2 |
| LiveCodeBench v6 | 73.3 | 71.1 ↑14.4 | 74.6 ↑11.5 |
| SWE Verified (Agentless) | 57.6 | 37.2 ↑11.1 | 43.1 ↑8.6 |
Competitive Coding Deep Dive
An in-depth look at how Nemotron-Cascade achieves state-of-the-art results in competitive programming benchmarks like LiveCodeBench and the International Olympiad in Informatics (IOI).
90.37 points
IOI 2025 Problem 2 Triples Score (14B-Thinking)
IOI 2025: Silver Medal Performance
Our Nemotron-Cascade-14B-Thinking model achieved a silver medal in the challenging International Olympiad in Informatics (IOI) 2025. This was made possible by a feedback-driven, test-time scaling pipeline with a 128K token thinking budget, demonstrating robust performance on real, high-stakes competitive programming problems.
- ✓ Multi-round generate-select-submit process (up to 50 rounds per problem).
- ✓ Feedback-driven prompt updates based on official judge verdicts and submission history.
- ✓ Cross-subtask insights leveraging solved subtasks to inform unsolved ones.
- ✓ Achieved 90.37 points on IOI 2025 Problem 2 Triples, surpassing OpenAI's internal IOI-gold model.
RLHF & Model Alignment
Explore the critical role of Reinforcement Learning from Human Feedback (RLHF) in improving model alignment, reducing verbosity, and enhancing overall response quality across diverse tasks.
Uses 72B
Optimal Reward Model Size for Stable RLHF
| LLM Backbone | Overall RewardBench | ArenaHard (SC) | AIME25 |
|---|---|---|---|
| Qwen2.5-Instruct-7B | 91.98 | 85.09 | 52.55 (7B RM) |
| Qwen2.5-Instruct-72B | 95.15 | 89.69 | 56.23 (72B RM) |
| WorldPM-72B | 94.08 | 86.40 | 54.67 (72B RM) |
Software Engineering (SWE) RL
Delve into the specialized Reinforcement Learning for Software Engineering, focusing on code localization, repair, and patch validation to resolve real-world GitHub issues.
Up to 24K
Optimized Input Context Length for SWE RL
Advanced ROI Calculator: Quantify Your AI Advantage
Estimate the potential annual cost savings and efficiency gains by integrating Nemotron-Cascade into your enterprise operations.
Estimated Annual Savings
$0
Annual Hours Reclaimed
0
Your Path to AI Excellence: The Nemotron-Cascade Implementation Roadmap
Our proven, cascaded reinforcement learning pipeline systematically builds and refines AI capabilities, ensuring robust, scalable, and domain-specific performance.
RLHF: Alignment & Foundational Reasoning
Enhance overall response quality, reduce verbosity, and align models with human preferences, providing a strong basis for advanced reasoning.
Instruction-Following RL: Precision & Control
Develop strict instruction adherence, enabling models to precisely follow complex multi-step user commands and constraints.
Math RL: Specialized Problem Solving
Boost mathematical reasoning and problem-solving through verifiable rewards and length extension training for complex problems.
Code RL: Competitive Programming Mastery
Achieve state-of-the-art performance in code generation and competitive programming challenges with rule-based reward functions.
SWE RL: Software Engineering Expertise
Refine code repair accuracy for real-world GitHub issues using an execution-free reward model and multi-stage context extension.
Ready to Transform Your Enterprise with AI?
Book a strategic consultation to discover how Nemotron-Cascade can drive unparalleled performance and efficiency for your organization.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat