Banger paper from NVIDIA. Training general-purpose reasoning models with RL is complicated. Different domains have wildly different response lengths and verification times. Math uses fast symbolic verification. Code requires slow execution-based verification. Alignment needs reward model scores. Blending all these heterogeneous prompts together makes the infrastructure complex, slows training, and makes hyperparameter tuning difficult. This new research introduces Cascade RL, a framework that trains models sequentially across domains rather than mixing everything together. First RLHF for alignment, then instruction-following RL, then math RL, then code RL, then software engineering RL. This sequential approach is resistant to catastrophic forgetting. In RL, the model generates its own experience, so old behaviors remain if they stay reward-relevant. Unlike supervised learning, where previous data disappears, RL optimizes cumulative reward rather than fitting exact targets. RLHF, as a pre-step, actually boosts reasoning ability far beyond mere preference optimization by reducing verbosity and repetition. Subsequent domain-specific RL stages rarely degrade earlier performance and may even improve it. Here are the results: Their 14B model outperforms its own SFT teacher, DeepSeek-R1-0528 (671B), on LiveCodeBench v5/v6/Pro. Nemotron-Cascade-8B achieves 71.1% on LiveCodeBench v6, comparable to DeepSeek-R1-0528 at 73.3% despite being 84x smaller. The 14B model achieved silver medal performance at IOI 2025. They also demonstrate that unified reasoning models can operate effectively in both thinking and non-thinking modes, closing the gap with dedicated thinking models while keeping everything in a single model. Paper: Learn to build effective AI Agents in our academy: