NVIDIA recipe: RL post-training lifts Nemotron 3.5 Super VL accuracy a further 17% on drug-metabolism task
NVIDIA has published a recipe for post-training Nemotron 3.5 Super VL on scientific tasks using NeMo Gym and NeMo RL. In a drug-metabolism example, supervised fine-tuning improved accuracy by 12% and RL post-training lifted it a further 17% at a 16,384-token output budget.
NVIDIA has published a step-by-step recipe for post-training Nemotron 3.5 Super VL on scientific use cases. In its drug-metabolism example, supervised fine-tuning improved accuracy by 12%, and RL post-training lifted it a further 17% at a 16,384-token output budget. The guide appeared on the NVIDIA Developer blog on Nov 16, 2026, written by Gautham Dharuman, Xiaoyun Wang, Logan Ward, Geetika Gupta and Kibibi Moseley.
What NVIDIA is offering
The recipe sits inside NVIDIA Nemotron for Science, which the company describes as a collection of models, datasets, tools and recipes for agentic scientific workloads. Two components are named:
- Nemotron 3.5 Super VL: an open, customizable multimodal model for specialized scientific workflows.
- NVIDIA Nemotron Litmus-Bench V2.0: an open source chemistry-reasoning dataset and benchmark for training and evaluating LLMs and agentic AI models on molecular properties.
NVIDIA claims Nemotron 3.5 Super VL delivers frontier-level accuracy on scientific benchmarks out of the box, along with the fastest time to solution for scientific coding. Its chart compares four models on 11 science benchmarks. According to the chart description, Nemotron 3.5 Super VL leads on seven, including ChemBench (75.9) and ether0 (39).
The post-training recipe
The workflow starts from a CSV of experimental records, such as sequences, SMILES strings, reaction conditions and metadata paired with measured outcomes. These records are reformatted into question-answer or user/assistant prompt pairs. The experimental ground truth then acts as the verifier that scores model responses.
This forms a reinforcement learning with verifiable rewards (RLVR) environment. NVIDIA says it needs no human preference labelling, because the reward comes directly from the data. Two NVIDIA tools split the work:
- NeMo Gym supplies the environment infrastructure.
- NeMo RL supplies the post-training framework.
The verifier's fidelity can be tuned to the task. A simple sequence-similarity function works as a lightweight verifier. Cheminformatics libraries, molecular simulators or even a live autonomous lab can be wired in as tool calls in the scoring layer.
Worked example: drug metabolism
Given a parent compound and its transforming enzyme, the model predicts the resulting metabolite. The examples use records from MetxBioDB, a database of experimentally confirmed biotransformation reactions. The verifier compares predicted and reference molecular fingerprints using Tanimoto similarity, which for the bit fingerprints used here scores from 0 to 1, with higher meaning more similar.
The reported results: supervised fine-tuning improved accuracy by 12%, and RL post-training added a further 17% at a 16,384-token output budget.
Optional distillation step
Before RL, a stronger teacher model such as Claude Sonnet 5.5 can generate high-quality reasoning traces. These warm-start the student model, which NVIDIA says accelerates convergence. The guide presents direct RL first, then distillation as an optional addition.
What this means
The recipe shows that domain adaptation of Nemotron 3.5 Super VL can run from a plain table of lab measurements, with the measurements themselves acting as the reward signal. Teams with experimental datasets can therefore skip preference labelling. Because the verifier is swappable, the same pipeline can start with a cheap similarity metric and move toward simulators or lab instruments as the task demands. The 12% and 17% gains come from a single drug-metabolism example, so results on other scientific tasks will depend on the data and verifier used. The benchmark leadership and speed claims are NVIDIA's own.
Related Articles
NVIDIA launches Nemotron for Science with Nemotron 3.5 Super VL and a chemistry dataset of nearly 50,000 questions
NVIDIA introduced Nemotron for Science on November 16, 2026. It pairs the open Nemotron 3.5 Super VL model with scientific evaluations, post-training recipes and LitmusBench v0.2, a chemistry dataset of nearly 50,000 questions.
NVIDIA releases two synthetic Japanese datasets, Nemotron-Personas-Finance-Japan and Nemotron-PII-Japan
At NVIDIA Agentic AI Day Japan 2026 in Tokyo on November 10, NVIDIA presented two synthetic datasets for Japanese AI development. Nemotron-Personas-Finance-Japan targets financial AI and user simulation, and Nemotron-PII-Japan covers PII detection and masking across 8 domains. Both are published on Hugging Face.
NVIDIA Alpamayo 2 Edge: 2B-Parameter Driving Model Hits 160-190 ms Latency on DRIVE AGX Thor
NVIDIA has introduced Alpamayo 2 Edge, a 2B-parameter adaptive reasoning vision-language-action model built for real-time operation on DRIVE AGX Thor. NVIDIA says deployment recipes with TensorRT and DFlash2 reach 160-190 ms pixel-to-trajectory latency, and that closed-loop RL improved driving performance by 48.9%.
NVIDIA Vera Rubin NVL72 posts over 10x GB300 NVL72 throughput per kW on MLPerf Endpoints v1.0
Provisional MLPerf Endpoints v1.0 results show NVIDIA Vera Rubin NVL72 delivering over 10x the throughput per kilowatt of GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. The Vera Rubin results are pending MLCommons verification; the GB300 results are verified.
Comments
Loading...