Reproduce “Mesh-Native Tokenisation: Retrofitting Foundation Models to Unstructured Data” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Georg Maierhofer, Miles Cranmer, Payel Mukhopadhyay

MNC adapts grid-pretrained emulators to irregular meshes by replacing convolutional input and output layers with mesh-aware versions that preserve physical neighborhoods and reuse weights, while retaining the processor. On shear flow, fine-tuned Overtone+MNC reaches next-step VRMSE 0.43 ± 0.01 × 10⁻², versus 0.66 ± 0.02 × 10⁻² from scratch. The evidence supports sample efficiency, not accuracy gains in every setting or reduced wall-clock execution time.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.