Research map / Probabilistic and scientific modeling / Variational and generative sampling
DISCS: A Benchmark for Discrete Sampling
Katayoon Goshvadi, Haoran Sun, Xingchao Liu, Azade Nova, Ruqi Zhang, Will Sussman Grathwohl, Dale Schuurmans, Hanjun Dai
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use weight function t/(t+1) instead of sqrt(t) in locally balanced samplers (ablation) | GWG/DMALA/PAS/DLMC · Text infilling (BERT) | Self-BLEU (lower better): 85.93 -> 81.15 | Improved |
| Use weight function t/(t+1) instead of sqrt(t) (ablation) | GWG · Text infilling (BERT) | Corpus BLEU: 16.75 -> 16.68 | No clear effect |
| Use weight function t/(t+1) instead of sqrt(t) (ablation) | DLMC · Text infilling (BERT) | Self-BLEU (lower better): 85.28 -> 84.55 | Improved |
| Use weight function t/(t+1) instead of sqrt(t) (ablation) | DLMCf · Text infilling (BERT) | Self-BLEU (lower better): 88.39 -> 80.12 | Improved |
| Use locally balanced samplers (GWG) instead of RWM | random walk Metropolis · Text infilling (BERT) | Self-BLEU (lower better): 92.41 -> 85.93 | Improved |
| Use gradient-based PAS instead of RWM | random walk Metropolis · MIS on ER[9000-11000] density 0.15 | set size lower bound: -555.674 -> 377.123 | Improved |
| Use PAS instead of DLMC | DLMC · MIS on ER[9000-11000] density 0.15 | set size lower bound: 355.058 -> 377.123 | Improved |
| Use PAS instead of GWG-nA | GWG-nA · MIS on ER[700-800] density 0.25 | set size lower bound: 28.187 -> 28.500 | Improved |
| Use weight function t/(t+1) when inverse temperature is above critical point 0.4407 (ablation) | locally balanced samplers · Ising model 50x50 and 100x100 | ESS | Improved |
| Use weight function sqrt(t) when inverse temperature is below critical point 0.4407 (ablation) | locally balanced samplers · Ising model 50x50 and 100x100 | ESS | Improved |
| Move inverse temperature closer to critical point 0.4407 (ablation) | samplers · Ising model | sampling difficulty | Hurt |
Closest papers by meaning
- Jump Your Steps: Optimizing Sampling Schedule of Discrete Diffusion Models (ICLR 2025 · similarity 0.86)
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.86)
- Learnable Sampler Distillation for Discrete Diffusion Models (NeurIPS 2025 · similarity 0.85)
- Split Gibbs Discrete Diffusion Posterior Sampling (NeurIPS 2025 · similarity 0.85)
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.85)
- Gradient-based Discrete Sampling with Automatic Cyclical Scheduling (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
