Skip to content
Bud's Wiki

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

TL;DR

Section titled “TL;DR”

Nemotron-Cascade 2 is NVIDIA’s open 30B-total / 3B-active MoE LLM post-trained from Nemotron-3-Nano-30B-A3B-Base via Cascade RL (sequential domain-wise RL across IF / multi-domain / RLHF / long-context / code / SWE stages) augmented with multi-domain on-policy distillation (MOPD) from the strongest intermediate teacher at each stage. It is the second open-weight LLM after DeepSeek-V3.2-Speciale-671B-A37B to reach gold-medal performance on IMO 2025, IOI 2025, and ICPC World Finals 2025 — at 20× fewer parameters. The full SFT and RL datasets are released alongside the model, making this the most complete public post-training recipe for a reasoning-and-agent-capable MoE at the 30B scale.

Key claims

Section titled “Key claims”
  • Cascade RL is the central thesis: sequential, domain-wise RL training stages (IF-RL → multi-domain RL → MOPD → RLHF → long-context RL → Code RL → SWE RL) resist catastrophic forgetting, let hyperparameters/curricula be tuned per-domain, and improve compute efficiency by keeping response lengths and verification wall-clock times homogeneous within a stage [§1, §4.1].
  • Multi-Domain On-Policy Distillation (MOPD) distills from the strongest intermediate teacher per domain during Cascade RL — explicitly framed as a mechanism to recover benchmark regressions that appear when training in increasingly complex RL environments [§1, §4.4].
  • Multi-domain RL is fused into Cascade RL for groups of tasks with similar response formats and comparable verification costs, joint-trained when cross-task interference is minimal — a middle ground between fully sequential and fully joint multi-domain RL [§1, §4.3].
  • Achieves gold-medal performance on three top-tier 2025 competitions: IMO 35/42 (Gold), IOI 439.28/600 (Gold), ICPC WF 10/12 (Gold), with IMO solutions graded by an IMO-2015 human gold medalist [§2, Table 2].
  • Outperforms both the latest Qwen3.5-35B-A3B (2026-02-24) and the larger Nemotron-3-Super-120B-A12B across math, code reasoning, alignment, and instruction following [§2, Table 1].
  • Headline numbers vs Qwen3.5-35B-A3B: LiveCodeBench v6 87.2 vs 74.6, LiveCodeBench Pro 25Q1 Medium 27.6 vs 17.8 (36.8 with TIR), HMMT Feb25 94.6 vs 89.0, AIME 2025 92.4 vs 91.9, ArenaHard v2 83.5 vs 65.4, IFBench 82.9 vs 70.2 [Table 1].
  • On IMO ProofBench Cascade 2 scores 72.9, behind GeminiDeep Think (76.7) and DeepSeekMath-V2-671B-A37B (80.2) but with 20× fewer parameters than the latter [§5.2, Table 1].
  • Underperforms Qwen3.5-35B-A3B on knowledge-intensive tasks (MMLU-Pro 79.8 vs 85.3, GPQA-Diamond 76.1 vs 84.2, HLE 17.7 vs 22.4) and on a subset of agentic benchmarks (BFCL v4 52.9 vs 67.3, τ2-Bench 58.9 vs 81.2, Terminal Bench 2.0 21.1 vs 40.5, SWE Verified OpenHands 50.2 vs 69.2), explicitly attributed to weaker knowledge-intensive pretraining and less agentic RL in this release [§2, Table 1].
  • Beats Nemotron-3-Nano-30B-A3B (same pretrained base) on nearly all benchmarks, isolating Cascade RL + MOPD as the contributor [§2].
  • SFT spans 10 task categories (math, code reasoning, science, long context, general chat, instruction following, safety, conversational agent, SWE agent, terminal agent), packed to 256K-token sequences, with optimal performance at ≈1.5 epochs [§3.1, §3.2].
  • SFT dataset is teacher-distilled at scale: 1.8M tool-calling math + 2.6M non-tool math + 816K math-proof samples (proof generation + verification); 1.9M Python + 1.0M C++14 + 1.3M Python tool-calling code-reasoning traces; 1.1M scientific-coding samples; 2.7M science; 4.9M reasoning-on + 372K reasoning-off general chat; 700K synthesized multi-turn conversations via two-instance GPT-OSS-120B role-play [§3.2.1–§3.2.5].
  • Teachers across SFT stages are explicitly named: DeepSeek-V3.2-Speciale for non-tool math/proofs, DeepSeek-V3.2 for tool-calling math, GPT-OSS-120B for code reasoning and most chat/instruction-following, Qwen3-Coder-480B-A35B-Instruct for SWE agentic trajectories [§3.2.1, §3.2.2, §3.2.5, §3.2.9].
  • Chat-template change: a single <think> followed by newline activates thinking mode; an empty <think></think> block prepended activates non-thinking mode — replacing the /think / /no_think tags from Nemotron-Cascade 1 [§3.1.2, Fig. 1].
  • SWE data uses multiple agentic scaffolds (OpenHands, SWE-Agent, Mini-SWE-Agent, Wei-et-al agentless) over problem instances from SWE-Gym / SWE-rebench / R2E-Subset; agentless SWE covers buggy-code localization, code repair, and test-case generation [§3.2.9].
  • Coding-prompt deduplication uses sample-I/O fingerprinting plus n-gram analysis, removing ≈24.2% of self-duplicated prompts [§3.2.2].
  • Released artifacts: Nemotron-Cascade-2-30B-A3B checkpoint, Nemotron-Cascade-2-SFT-Data, and Nemotron-Cascade-2-RL-Data [Abstract].

Method

Section titled “Method”

The pipeline starts from Nemotron-3-Nano-30B-A3B-Base — a Mamba-2/Transformer hybrid MoE base shipped December 2025. SFT is a single stage on 256K-packed sequences spanning all 10 task categories, trained ≈1.5 epochs, with responses for each category distilled from the strongest available open teacher (DeepSeek-V3.2-Speciale for proofs, GPT-OSS-120B for code/chat, Qwen3-Coder-480B-A35B-Instruct for SWE trajectories). The chat template’s two-mode toggle (<think>\n vs empty <think></think>) lets the same checkpoint serve thinking and non-thinking responses without separate adapters.

Post-SFT, Cascade RL runs sequential, domain-specialized RL: IF-RL → multi-domain RL → MOPD → RLHF → long-context RL → Code RL → SWE RL (agentless followed by execution-based agentic). Each stage owns its hyperparameters, reward design, and curriculum. MOPD is the key novelty: at each Cascade RL stage, the current best intermediate checkpoint per domain is used as the teacher for on-policy distillation, which patches the benchmark regressions that accumulate as the model is pushed through harder RL environments. Multi-domain RL is added inside Cascade RL for task groups with similar response formats and verification costs (so they can be batch-trained efficiently when cross-task interference is low), without abandoning the sequential domain isolation principle for the rest. SWE RL includes both an agentless track (localization + repair + test generation) and an execution-based track operating inside a real agentic scaffold.

Results

Section titled “Results”
  • IMO 2025: 35/42 (Gold) with all six problems scored by an IMO-2015 human gold medalist [Table 2].
  • IOI 2025: 439.28/600 (Gold), verified by Online Judge with official test cases [Table 2].
  • ICPC World Finals 2025: 10/12 problems solved (Gold), verified by Online Judge [Table 2].
  • LiveCodeBench v6 (2408–2505): 87.2 (TIR 88.4) vs Qwen3.5-35B-A3B 74.6, Qwen3.5-397B-A17B 83.6, Kimi-K2.5-1T-Thinking 85.0 [Table 1].
  • LiveCodeBench Pro 25Q2 Medium: 27.6 (TIR 36.8) vs Qwen3.5-35B-A3B 17.8, Qwen3.5-397B-A17B 44.4, Kimi-K2.5 45.6 [Table 1].
  • HMMT Feb25: 94.6 vs Qwen3.5-35B-A3B 89.0, Nemotron-3-Super 93.7 [Table 1].
  • AIME 2025: 92.4 (TIR 98.6) vs Qwen3.5-35B-A3B 91.9 [Table 1].
  • IMO AnswerBench 79.3, IMO ProofBench 72.9 (vs Gemini Deep Think 76.7, DeepSeekMath-V2-671B-A37B 80.2) [Table 1].
  • ArenaHard v2: 83.5 (Hard Prompt 88.2; Creative Writing 78.7) vs Qwen3.5-35B-A3B 65.4 [Table 1].
  • IFBench (prompt): 82.9 vs Nemotron-3-Super 72.6, Qwen3.5-35B-A3B 70.2 [Table 1].
  • NIAH@1M (RULER subset): 99.0 vs Nemotron-3-Super 98.3, Qwen3.5-35B-A3B 94.3 [Table 1].
  • SWE Verified (OpenHands): 50.2 — beats Nemotron-3-Nano-30B-A3B 38.8 but trails Nemotron-3-Super-120B-A12B 60.5 and Qwen3.5-35B-A3B 69.2 [Table 1].
  • Terminal Bench 2.0: 21.1 vs Nemotron-3-Super 31.0, Qwen3.5-35B-A3B 40.5 [Table 1].
  • BFCL v4: 52.9, τ2-Bench: 58.9 — explicitly behind Qwen3.5-35B-A3B (67.3 / 81.2) [Table 1].
  • MMLU-Redux 86.3, MMLU-Pro 79.8, GPQA-Diamond 76.1, HLE 17.7 — knowledge-intensive gap vs Qwen3.5-35B-A3B (93.3 / 85.3 / 84.2 / 22.4) [Table 1].
  • Outperforms Nemotron-3-Nano-30B-A3B (same base, same post-trainer family) on nearly all benchmarks, isolating Cascade RL + MOPD as the contributor [§2].

Why it’s interesting

Section titled “Why it’s interesting”

This is the cleanest filed datapoint for a post-training-only recipe pushing a 30B MoE into frontier reasoning territory — directly comparable to Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning (same NVIDIA Nemotron program, same base-model lineage, but 120B/12B-active and trained from scratch with NVFP4), making the pair a natural ablation of “scale up the pretraining” vs “scale up the post-training” within one organization. Cascade RL + MOPD is also a direct extension of the Reasoning RL thread: where most filed entries (InfTool, MiroThinker, HuDA, Training-Free GRPO, Kimi K2.5 PARL) explore reward shaping or alternative RL formulations within a single domain, this paper builds a multi-stage scaffold for combining GRPO-style RL across many domains without catastrophic forgetting — the explicit MOPD mechanism for recovering benchmark regressions is a new primitive worth tracking. The open-data release (Nemotron-Cascade-2-SFT-Data + RL-Data, called out explicitly in Haoxiang’s note) puts it in the Open foundation-model releases cohort alongside Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling (compute-optimal teacher choice for synthetic data) and the broader Nemotron data-curation lineage Nemotron-CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training. The agentic-side weakness (trails Qwen3.5-35B-A3B on BFCL / τ2-Bench / Terminal-Bench / SWE-Bench despite winning on math and code reasoning) is consistent with the Agentic Software Engineering page’s observation that agentic capability requires its own dedicated post-training surface — and the authors themselves flag this gap as the next priority.

See also

Section titled “See also”