NVIDIA launches Nemotron for Science: RL post-training adds 17% on drug-metabolism prediction
NVIDIAAI · x · 2026-11-17
At SC26, NVIDIA's Ian Buck announced Nemotron for Science — a collection of open models, datasets, tools, and post-training recipes for building AI agents that plan tasks, run simulations, check results, and self-correct.
Key points:
- Nemotron 3.5 Super VL: an open, customizable multimodal model with frontier-level accuracy on scientific benchmarks and the fastest time-to-solution for scientific coding.
- RL Litmus-Bench V2.0: a dataset/benchmark for training and evaluating LLMs and agentic models on molecular properties.
- Post-training recipe: build a custom RL environment from experimental data with NeMo Gym, then apply RL via NeMo RL. In a drug-metabolism example, SFT improved accuracy 12% and RL added another 17% at a 16,384-token output budget.
- Optional distillation: warm-start the student model with reasoning traces from a stronger teacher like Claude Sonnet 5.5 to speed up RL convergence.
A full walkthrough covers building and evaluating an RL environment for your own scientific task.
More from Models
- NVIDIA launches Nemotron for Science: fully open models, datasets and agent recipes — NVIDIAAI · 2026-11-17
- Diffusion images lack controllability: researchers argue for painter-style, tool-call-based generation — josephdviviano · 2026-11-17
- User Has Spent $2,513 on ChatGPT Since 2023 but Eyes a Switch to Claude — MoreLunch1547 · 2026-11-17
- Claude Fable 5.2 writes 43% shorter replies and cuts em dash usage by 99% — ArtificialAnlys · 2026-11-17
- InferenceBench update: Gemini 4 Carbon takes #3, breaking Anthropic's leaderboard sweep — maksym_andr · 2026-11-17
- ChatGPT hallucinates an entire nonexistent client email while sorting the user's inbox — jdjohnson · 2026-11-17