NVIDIA Introduces AceReason-Nemotron: Enhancing Math and Code Reasoning through Reinforcement Learning

 NVIDIA has unveiled AceReason-Nemotron, a 14-billion-parameter open-source model designed to enhance mathematical and coding reasoning through large-scale reinforcement learning (RL). This model demonstrates that RL can significantly improve reasoning capabilities in small to mid-sized models, surpassing traditional distillation-based approaches.

Key Features and Innovations

• Sequential RL Training Strategy: The model undergoes a two-phase RL training process—initially on math-only prompts, followed by code-only prompts. This approach not only boosts performance in respective domains but also ensures minimal degradation across tasks. 

• Enhanced Benchmark Performance: AceReason-Nemotron-14B achieves notable improvements on various benchmarks:

  • AIME 2025: 67.4% (+17.4%)

  • LiveCodeBench v5: 61.1% (+8%)

  • LiveCodeBench v6: 54.9% (+7%) 

• Robust Data Curation Pipeline: NVIDIA developed a comprehensive data curation system to collect challenging prompts with verifiable answers, facilitating effective verification-based RL across both math and code domains. 

• Curriculum Learning and Stability: The training incorporates curriculum learning with progressively increasing response lengths and utilizes on-policy parameter updates to stabilize the RL process. 

Implications for AI Development

AceReason-Nemotron's success illustrates the potential of reinforcement learning in enhancing the reasoning abilities of AI models, particularly in mathematical and coding tasks. By releasing this model under the NVIDIA Open Model License, NVIDIA encourages further research and development in the AI community.

Nemotron-Tool-N1: Revolutionizing LLM Tool Use with Reinforcement Learning

 In the rapidly evolving field of artificial intelligence, enabling large language models (LLMs) to effectively utilize external tools has become a focal point. Traditional methods often rely on supervised fine-tuning, which can be resource-intensive and may not generalize well across diverse tasks. Addressing these challenges, researchers have introduced Nemotron-Tool-N1, a novel approach that employs reinforcement learning to train LLMs for tool use with minimal supervision.

Moving Beyond Supervised Fine-Tuning

Conventional approaches to teaching LLMs tool usage typically involve supervised fine-tuning (SFT), where models learn from annotated reasoning traces or outputs from more powerful models. While effective to an extent, these methods often result in models that mimic reasoning patterns without truly understanding them, limiting their adaptability.

Nemotron-Tool-N1 diverges from this path by utilizing a reinforcement learning framework inspired by DeepSeek-R1. Instead of relying on detailed annotations, the model receives binary rewards based on the structural validity and functional correctness of its tool invocations. This approach encourages the model to develop its own reasoning strategies, leading to better generalization across tasks.

Impressive Performance Benchmarks

Built upon the Qwen-2.5-7B and Qwen-2.5-14B architectures, Nemotron-Tool-N1 has demonstrated remarkable performance. In evaluations using the BFCL and API-Bank benchmarks, the model not only achieved state-of-the-art results but also outperformed GPT-4o, showcasing its superior capability in tool utilization tasks.

Implications for the Future of AI

The success of Nemotron-Tool-N1 underscores the potential of reinforcement learning in training LLMs for complex tasks with minimal supervision. By moving away from traditional fine-tuning methods, this approach offers a more scalable and adaptable solution for integrating tool use into AI systems.

As the demand for more versatile and efficient AI models grows, innovations like Nemotron-Tool-N1 pave the way for future advancements in the field.