Physics law sets diffusion model's noise schedule for better particle spread prediction
bravo_abad · x · 2026-10-12
Research by Reichherzer and colleagues bakes physics into generative diffusion: to predict how a particle beam spreads through plasma, they derive the model's noise schedule from a known physical law of spread over time, so the added noise's variance follows the prescribed evolution. With the neural network and training budget unchanged, this physics-derived schedule tracked the expected spreading better than standard schedules.
More from Research
- Real-time chunking world model demo runs autonomously at 200ms per generation — ___Mufasaa · 2026-10-12
- CL4SM workshop at NeurIPS 2026 pre-event calls for physically constrained ML papers — cici118 · 2026-10-12
- Adobe's Credit Stabilization through Time makes short-sequence training generalize to long contexts — burkov · 2026-10-12
- Compressing a ~500MB RoBERTa AI-text detector to ~97MB that runs fully in-browser — Alternative_Detail31 · 2026-10-12
- PwC upgrades paper PDF reader with faster loading and chapter jump navigation — NielsRogge · 2026-10-12
- Work2Skill Turns Human GUI Demos into Reusable Agent Skills, Beating No-Skill Baselines — nex-agi · 2026-10-12