Discrete Beckmann Transport Models: One-Step Discrete Generators Without a Teacher
rishabh16_ · x · 2026-09-16
A new Harvard collaboration introduces Discrete Beckmann Transport Models, a family of discrete generative models that provably carries any latent point to a fixed point on the simplex vertices in a single step, without a teacher model.
- Achieves the diversity-coherence Pareto frontier on unconditional language modeling.
- SOTA few-step reasoning: 84.6% on Sudoku-Hard with 4 NFEs, 16.8% on TinyGSM/GSM8K with 32 NFEs.
More from Research
- Fudan team: LLaMA3-70B and Qwen25-72B can self-replicate without human help — hey_abusiddik · 2026-09-16
- Ke Holdings Open-Sources PanoWorld: Consistent Whole-House 360° Panorama From Floorplans — rsasaki0109 · 2026-09-16
- Thompson's Compiler Attack Explains Why AI Evals Fail When the Model Helps Build Them — vishalmisra · 2026-09-16
- Docusign and NVIDIA Built a 900M-Parameter Extractor That Reads Contract Tables 20x Faster — AI Engineer · 2026-09-16
- Princeton's Elad Hazan reflects on two decades of online convex optimization — HazanPrinceton · 2026-09-16
- RL on private lab data to predict experimental outcomes, immune to web contamination — hsu_byron · 2026-09-16