NVIDIA's MOTIVE: Precisely Filtering Video Training Clips
NVIDIAAI · x · 2026-07-08
NVIDIA Research's paper MOTIVE received an Outstanding Paper Honorable Mention at ICML 2026. Targeting video model training, the method re-weights training signals—favoring motion regions while downplaying static backgrounds. By scoring the impact of motion per clip, it filters out small batches of high-impact clips for fine-tuning to improve temporal dynamics. MOTIVE showed improvements in VBench dynamic metrics and achieved a 74.1% human preference win rate.
More from Research
- Deep Learning Course Adds Mechanistic Interpretability Lecture This Winter — CSProfKGD · 2026-10-26
- INCLUDE-2.0 benchmark evaluates regional knowledge across 91 languages and 86 countries — ABosselut · 2026-10-26
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- CoVer at EMNLP 2026: Training Code LLMs to Generate and Self-Verify Simultaneously — algo_diver · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Turing-complete malware can run in PyTorch's backward pass via a zero-valued loss — retr0jirachi · 2026-10-26