Florida State University

2 related articles

Subscribe to updates

arXiv Oct 12, 2026
Same molecule, shuffled spectra: a benchmark asks whether models actually use the evidence
MolP2SBench recasts molecular property-to-structure inference as masked-scaffold multiple choice and adds shuffled and counterfactual property substitutions, separating a correct answer from genuine use of the supplied evidence. Across seven models, accuracy on a property channel does not track how strongly a model responds when that property is replaced, and model performance does not always match chemists' intuition about which properties are easier.
MolP2SBench recasts molecular property-to-structure inference as masked-scaffold multiple choice and adds shuffled and counterfactual property substitutions, separating a correct answer from genuine use of the supplied evidence. Across seven models, accuracy on a property channel does not track how strongly a model responds when that property is replaced, and model performance does not always match chemists' intuition about which properties are easier.
University of Georgia University of California, Berkeley University of Tennessee at Knoxville University of Illinois Urbana-Champaign Florida State University
Yuhang Xie（First author）
MolP2SBench Property-Grounded Recovery (PGR)Molecular Structure Prediction
SIAM Journal on Applied Mathematics Sep 24, 2026
Integrating image inpainting into an ensemble score filter to track surface quasi-geostrophic dynamics under partial observations
The work develops an ensemble score filter (EnSF) that integrates image inpainting to address data assimilation with partial observations: at each filtering step a training-free diffusion model estimates the observed states by incorporating likelihood information into the score function, and image inpainting methods then predict the unobserved state variables, with performance demonstrated by tracking Surface Quasi-Geostrophic (SQG) model dynamics across a variety of scenarios as a proof of concept.
The work develops an ensemble score filter (EnSF) that integrates image inpainting to address data assimilation with partial observations: at each filtering step a training-free diffusion model estimates the observed states by incorporating likelihood information into the score function, and image inpainting methods then predict the unobserved state variables, with performance demonstrated by tracking Surface Quasi-Geostrophic (SQG) model dynamics across a variety of scenarios as a proof of concept.
Colorado State University Florida State University
Siming Liang（First author）
Ensemble Score Filter Data Assimilation Image Inpainting Diffusion Model Surface Quasi-Geostrophic Dynamics

You’re all caught up

Same molecule, shuffled spectra: a benchmark asks whether models actually use the evidence

Integrating image inpainting into an ensemble score filter to track surface quasi-geostrophic dynamics under partial observations