Swapping Only the Sampler Lifts Diffusion Model's Sudoku Accuracy From 0.8% to 65.7%
francoisfleuret · x · 2026-10-26
An experiment thread shared by MariiaDML (retweeted by Francois Fleuret) suggests continuous diffusion models are far better at discrete reasoning than standard sampling reveals:
- Subject: SRM's MNIST-Sudoku Hard checkpoint (arXiv 2502.21075)
- Only change: the sampler, with no retraining
- Result: valid boards solved jumped from 0.8% to 65.7%
- The authors argue standard sampling is the bottleneck for discrete reasoning, and break down why it fails in the thread
More from Research
- Top AI forecaster FutureSearch whiffed on Brazil election, showing prediction's intrinsic limits — sebkrier · 2026-10-26
- Ai2's LitPivot helps researchers stress-test early ideas against the literature, study of 17 shows gains — lmoroney · 2026-10-26
- Ropedia W0 trains robot world model on 10,000 hours of human video — liuziwei7 · 2026-10-26
- UDT vs FDT: does policy optimization beat logical counterfactuals on continuous Newcomb problems — jessi_cata · 2026-10-26
- Fine-tuning Qwen3-1.7B on a 16GB GPU to write coherently without the letter 'e' — CatAstro_Piyush · 2026-10-26
- Meta's overlooked JEM SSL framework beats DINOv3 with 12x less data — iScienceLuvr · 2026-10-26