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

Zhuolin Yang; Zihan Liu; Yang Chen; Wenliang Dai; Boxin Wang; Sheng-Chieh Lin; Chankyu Lee; Yangyi Chen

cs.CLMar 19, 2026

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

Authors: Zhuolin Yang, Zihan Liu, Yang Chen, Wenliang Dai, Boxin Wang, Sheng-Chieh Lin, Chankyu Lee, Yangyi Chen,

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.

Figures & tables

Figure 1

Competition	P1	P2	P3	P4	P5	P6	Overall	Medal
IMO 2025	7	7 †	7	7	7	0	35/42	Gold
IOI 2025	39	88.53	100	100	28.75	83	439.28/600	Gold

Table 2: Performance of Nemotron-Cascade-2-30B-A3B model on IMO 2025, IOI 2025, and ICPC World Finals 2025 competitions. Nemotron-Cascade-2 model achieved solid gold medal on all these top-tier competitions. Our IMO 2025 solutions are evaluated by human expert (IMO 2015 Gold medalist) while IOI 2025 and ICPCWF 2025 solutions are verified through OnlineJudge with official testcases.

Figure 1: (Left) The chat template uses adjacent <think></think> tokens to indicate non-thinking mode, and a single <think> followed by \n to indicate thinking mode. (Right) For tool calling, the available tools are listed in the system prompt. The model is instructed to call tools within the <tool_call> and </tool_call> tags.

Figure 2: Nemotron-Cascade 2 applies Cascade RL with the sequential, domain-wise ordering after SFT, leading to substantial improvements across the corresponding domains.

Panel of Figure 3 — Figure 3: Training dynamics and downstream evaluation.

		ArenaHard v2
Method	Steps	Hard Prompt	Creative Writing
Initial	0	71.5	40.6
RLHF	100	81.7	68.6
RLHF	160	80.7	71.2
MOPD	52	85.5	71.0

Table 3: Comparison of MOPD and RLHF at matched evaluation checkpoints on ArenaHard V2.0.

Scaffold	Agentless Mini		OpenHands
Scaffold	avg@4	pass@4	avg@4	pass@4
Init.	41.9%	55.2%	49.8%	64.2%
after Agentless RL	44.3%	57.4%	50.8%	65.0%

Table 4: Effectiveness of Agentless RL on SWE-bench Verified.

Model	IMO-ProofBench
	Basic (30)	Advanced (30)	Overall (60)
Aletheia ( Feng et al., 2026 )	-	91.9	-
Gemini 3 Deep Think ( Gemini Team, 2026 )	-	76.7	-
Gemini Deep Think (IMO Gold) ( Gemini Team, 2025b )	89.0	65.7	76.7
DeepSeek-Math-V2-671B-A37B ( Shao et al., 2025 )	99.0	61.9	80.2
DeepSeek-Math-V2-671B-A37B ( our reproduced score ) †	99.5	57.7	78.6

Table 5: IMO-ProofBench ( Luong et al., 2025 ) reports scores split into the Basic (30 problems) and Advanced (30 problems) subtasks, as well as Overall (60 problems). Expert-evaluated results are taken from the IMO-ProofBench leaderboard (accessed on 2026/3/9).

Panel of Figure 4 — Figure 4: IMO-ProofBench (Advanced) score graded by LLM ProofAutoGrader (DeepSeek-V3.2-Speciale).

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameters
Global batch size	$64$
Packed sequence length	$256$ K
Max learning rate	$5\times 10^{-5}$
Min learning rate	$5\times 10^{-6}$
Learning rate warmup steps	200
Scheduler	cosine

Appendix

Table 7: Training hyperparameters for Nemotron-Cascade-2-30B-A3B in SFT.

Hyper-parameters	IF-RL	Multi-domain RL	MOPD
Max response length	49K	49K	98K
Batch size	128	128	128
# Rollout size	16	16	4
Learning rate	$3\times 10^{-6}$	$3\times 10^{-6}$	$3\times 10^{-6}$
Steps	180	70	52
	AdamW	Adam	AdamW

Appendix

Table 8: Training hyperparameters of Nemotron-Cascade-2-30B-A3B in Cascade RL (IF-RL, Multi-domain RL, MOPD).

Hyper-parameters	RLHF	Long-context RL	Code RL
Max response length	16K	49K	118K
Batch size	128	128	128
# Rollout size	16	16	16
Learning rate	$3\times 10^{-6}$	$3\times 10^{-6}$	$3\times 10^{-6}$
Steps	25	30	22
	AdamW	Adam	AdamW

Appendix

Table 9: Training hyperparameters of Nemotron-Cascade-2-30B-A3B in Cascade RL (RLHF, Long-context RL, Code RL).

Hyperparameters
# prompts per step	$16$
# rollout	$64$
Temperature	$0.8$
Max sequence length	$256k$
Max turn	$200$
Max learning rate	$3\times 10^{-6}$

Appendix

Table 10: Training hyperparameters of Nemotron-Cascade-2-30B-A3B model in execution-based agentic SWE-RL.

Contest Name	Contest Problems									Score	Penalty	Est. Rank	ELO
Hello 2025	A	B	C	D	E1	E2	F	G	H	10779.46	-	13/16703	3449
Hello 2025	500.00	1000.00	1493.75	2235.71	0.0	1900.00	0.0	3650.00	0.0	10779.46	-	13/16703	3449
Codeforces Round 996 (Div. 2)	A	B	C	D	E	F				5793.75	-	2/21232	2198
Codeforces Round 996 (Div. 2)	500.00	993.75	1475.00	0.0	2825.00	0.0				5793.75	-	2/21232	2198
Codeforces Round 997 (Div. 2)	A	B	C	D	E	F1	F2			9378.75	-	1/18823	2198
Codeforces Round 997 (Div. 2)	493.75	1250.00	1475.00	0.0	2225.00	2710.00	1225.00			9378.75	-	1/18823	2198

Appendix

Table 11: Nemotron-Cascade-2-30B-A3B performance details on 40 Div.1 and Div.2 Codeforces Rounds ranging from 2501 to 2507 without python-tool use. We attempt each problem with

N=8

times in total. For regular codeforces rounds, we present the score after considering expected penalties for each problem. For ICPC style rounds, we mark passed/failed problems as + and - correspondingly. We compute the estimated rank to human contestants and the corresponding Elo score as shown in rightmost two columns.

Contest Name	Contest Problems									Score	Penalty	Est. Rank	ELO
Hello 2025	A	B	C	D	E1	E2	F	G	H	11712.50	-	11/16703	3497
Hello 2025	500.00	1000.00	1500.00	2225.00	937.50	1900.00	0.0	3650.00	0.0	11712.50	-	11/16703	3497
Codeforces Round 996 (Div. 2)	A	B	C	D	E	F				5025.00	-	2/21232	2198
Codeforces Round 996 (Div. 2)	500.00	975.00	1475.00	2075.00	0.0	0.0				5025.00	-	2/21232	2198
Codeforces Round 997 (Div. 2)	A	B	C	D	E	F1	F2			11187.50	-	1/18823	2198
Codeforces Round 997 (Div. 2)	493.75	1250.00	1500.00	1900.00	2075.00	2743.75	1225.00			11187.50	-	1/18823	2198

Appendix

Table 12: Nemotron-Cascade-2-30B-A3B performance details on 40 Div.1 and Div.2 Codeforces Rounds ranging from 2501 to 2507 with python-tool use. We attempt each problem with

N=8

RL for Language Model Reasoning

Reinforcement learning for language model reasoning uses reward-driven training to improve multi-step problem solving, tool use, and reliable answers.

See all 1,087 papers

Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
Jun 12, 2026NVIDIA, :, Aaron Blakeman +571Mixture-of-Experts Language Models Long-Context Language Model Inference
Allan
An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics
Sep 9, 2026Ivan Moshkov, Stephen Ge, George Armstrong +3Mathematical Reasoning Benchmarks Test-Time Scaling
NVIDIA
OLLM: Options-based Large Language Models
Apr 21, 2026Shashank Sharma, Janina Hoffmann, Vinay NamboodiriLLM Alignment RL for Language Model Reasoning
Department of Computer Science University of Bath · Department of Psychology University of Bath

Jun 12, 2026J/K move · Enter open · S save

Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

NVIDIA, :, Aaron Blakeman, Aaron Thomas+570

Allan

Mixture-of-Experts Language Models Long-Context Language Model Inference LLM Inference Acceleration Agentic Reasoning Hybrid Mamba-Transformer Architectures

Open paper AskPDF arXiv

We introduce Nemotron 3 Ultra, a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model. We pre-trained Nemotron 3 Ultra on 20 trillion text tokens, then extended the context length to 1M tokens, and post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD). Nemotron 3 Ultra is our most capable model yet, employing multiple key technologies - LatentMoE, Multi Token Prediction (MTP), NVFP4 pre-training, multi-environment RLVR, MOPD, and reasoning budget control. Nemotron 3 Ultra achieves up to ~6x higher inference throughput as compared to state-of-the-art publicly available LLMs while attaining on-par accuracy. The state-of-the-art accuracy, high inference throughput, and 1M token context length make Nemotron 3 Ultra ideal for long-running autonomous agentic tasks. We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace.

Abstract

RL for Language Model Reasoning

Explore similar work

Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics

OLLM: Options-based Large Language Models