GlycoGen-MDLM (50M, seed 2)
Converged PyTorch checkpoint (best_model.pt, random initialization seed 2) for "GlycoGen: Glycan Structure Prediction with Crystallizing Flows" (NeurIPS 2026).
- Architecture / Objective: 128-step Masked Diffusion Language Model + Crystallizing Flows
- Test-set top-1 structural accuracy (
n = 76,774):92.47% ± 0.11% (K=32, 3-seed mean) - Code & inference scripts: github.com/rlacombe/GlycoGen
- Project page: rlacombe.github.io/GlycoGen
- Benchmark splits:
rlacombe/candycrunch-v2-benchmark
Citation
@inproceedings{lacombe2026glycogen,
title = {GlycoGen: Glycan Structure Prediction with Crystallizing Flows},
author = {Lacombe, Romain},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026}
}