Diffusion guidance halves temperature-extreme biases across 17 climate models
StephanMandt · x · 2026-11-17
A new preprint from Stephan Mandt's team trains a conditional diffusion emulator on 17 CMIP6 models plus ERA5 reanalysis, then uses diffusion guidance to steer each emulated climate-model distribution toward observations, with guidance strength tuned via perfect-model experiments on withheld simulated futures.
Key results:
- Across 17 synthetic targets, guidance more than halves median biases in the 1st/99th percentiles of daily temperature
- Median annual global-mean temperature RMSE drops 32–57% across three emission scenarios
- Observational influence decreases with projection horizon and forcing
- Efficient sampling yields large ensembles without extra climate simulations, and guidance beats established bias corrections on probabilistic scores
More from Research
- Physicist pans AI "superconductivity breakthrough" as repackaged 1975 result — ZeeshanZiaML · 2026-11-17
- Studying retrosynthetic disconnection from Vogel's textbook: "really freaking cool" — CatAstro_Piyush · 2026-11-17
- AI drives 30% surge in arXiv theoretical physics submissions; models solve ~60% of private-problem benchmark — binarybits · 2026-11-17
- FrontierCS 2 benchmark challenges AI agents to beat published CS research blind — arena · 2026-11-17
- Statistician: Three Systemic Incidents in One Month Is Enough to Reject Low Base Rate (p=0.002) — wfithian · 2026-11-17
- Isomorphic Labs raises $5B, its second mega-round in six months — The Decoder · 2026-11-17