lifatsastain/structured-latent-text-refinement

Structured Latent Text Refinement

Draft-conditioned latent refinement for non-autoregressive text generation in frozen BERT latent space.

Paper

When Latent Geometry Is Not Enough: Draft-Conditioned Latent Refinement for Non-Autoregressive Text Generation
De Shuai Zhang, Beijing Wuzi University, May 2026
[arXiv link — coming soon]

Checkpoints

• stage1_rocstories_768_best.pt — Stage 1 BERT latent autoencoder with parallel decoder, trained on ROCStories
• draft_prior_rocstories_768_best.pt — DraftPrior network, 768-dimensional, trained with 0% token dropout

Code

https://github.com/saslifat-gif/structured-latent-text-refinement

Usage

See repository README for installation and inference instructions.