News

NVIDIA Nemotron 3.5 Super VL Enhances Scientific Accuracy with Post-Training

NVIDIA's Nemotron 3.5 Super VL leverages RL post-training to boost accuracy in scientific workflows, including drug-metabolism predictions.

NVIDIA Nemotron 3.5 Super VL Enhances Scientific Accuracy with Post-Training

NVIDIA has introduced a comprehensive solution for advancing scientific research through its Nemotron for Science platform. This initiative integrates the Nemotron 3.5 Super VL model, a multimodal AI system, with domain-specific datasets and tools to enhance accuracy and efficiency in complex scientific workflows. The platform addresses challenges such as error propagation, latency, and computational costs in agentic AI-driven research.

Key Features of Nemotron for Science

Nemotron for Science includes the Nemotron 3.5 Super VL model, which is open and customizable, and the Nemotron Litmus-Bench V2.0, an open-source dataset and benchmark for evaluating molecular properties. The platform delivers frontier-level accuracy on scientific benchmarks and accelerates time-to-solution for scientific coding tasks.

One of the standout features is the ability to post-train the Nemotron 3.5 Super VL model for specific scientific domains. NVIDIA provides a practical recipe for this process, which involves building a custom reinforcement learning (RL) environment using NeMo Gym and applying RL with NVIDIA NeMo RL. This approach enables researchers to fine-tune the model for tasks such as drug-metabolism prediction, where accuracy improvements of 12% through supervised fine-tuning (SFT) and an additional 17% through RL post-training were demonstrated.

Post-Training Workflow and Tools

The post-training process begins with experimental data formatted as input-output pairs, such as molecular structures represented by SMILES strings. These pairs are used to create a reinforcement learning with verifiable rewards (RLVR) environment. The model predicts outcomes, which are scored using tools like Tanimoto similarity to compare molecular fingerprints. This dense reward signal eliminates the need for human annotation.

Training is conducted using group relative policy optimization (GRPO) with low-rank adapters (LoRA). The environment infrastructure, reward server, and rollout workers are fully self-contained, allowing the pipeline to generalize across scientific domains. NVIDIA's approach ensures that even highly capable instruction-tuned models can achieve meaningful improvements when given targeted RL feedback.

Optional Distillation for Enhanced Performance

For scenarios where the base model requires additional capability before RL training, NVIDIA recommends a distillation step. This involves using a stronger teacher model to generate high-quality reasoning traces, which are then used to fine-tune the student model. This "warm-start" technique accelerates convergence and raises the performance floor, making subsequent RL training more efficient and effective.

In the drug-metabolism example, the teacher model's reasoning traces guided the student model to better predict metabolite structures, leading to significant accuracy gains. The process highlights the flexibility and adaptability of the Nemotron platform for specialized scientific tasks.

Implications for Scientific Research

By combining advanced AI models, domain-specific datasets, and robust post-training workflows, NVIDIA's Nemotron for Science platform enables researchers to tackle complex scientific challenges with greater accuracy and efficiency. The platform's ability to integrate seamlessly into agentic AI workflows positions it as a critical tool for accelerating discoveries in fields such as chemistry, biology, and beyond.

Sign in to save this article Free account: build your research collection and return to read it anytime.