NVIDIA framework lets frozen LLMs/VLMs keep learning from deployment, +34.2% on medical tasks
nvidia · hf · 2026-10-10
NVIDIA presents a model-agnostic framework that lets frozen LLMs and VLMs keep learning from deployment experience without touching weights — targeting the knowledge-staleness problem in medicine.
- Three forms of external expertise: a Skill guiding reasoning and tool use, a Knowledge Memory storing evidence-backed facts, and a Multimodal Knowledge Base retaining visual examples with retrieval guidance
- Validation strategy: an update is kept only if it helps on new cases without degrading earlier performance, avoiding overfitting to a fixed validation set
- Across 6 benchmarks (clinical diagnosis, workflows, medical reasoning, visual reasoning) and 4 open- and closed-weight base models, the framework improves medical tasks by up to 34.2%
It generalizes to unseen cases, transfers to other models without further optimization, and works in non-medical domains.
More from Models
- Three criticisms of how a leaderboard treats Sol: version, timing metric, and API tuning — mgostIH · 2026-10-10
- Giffmana notices new models Argon and Astra share the same first letter and length — giffmana · 2026-10-10
- OpenAI says Codex usage data will never power its predictions, even after testing — btibor91 · 2026-10-10
- It's 2026 — Duke Libraries explainer on why LLMs still hallucinate, from guess-favoring benchmarks — ArtificialOther · 2026-10-10
- Dev's hands-on Opus 5.5 review: big coding leap, zero personality left — ryunuck · 2026-10-10
- Step 5 Preview, 600B MoE with 1M context, goes free and matches GPT-6 Luna — NousResearch · 2026-10-10