Research map / Image generation / Fast diffusion sampling
Fast Byte Latent Transformer
Julie Kallini, Artidoro Pagnoni, Tomasz Limisiewicz, Gargi Ghosh, Luke Zettlemoyer, Christopher Potts, Xiaochuang Han, Srini Iyer
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 |
|---|---|---|---|
| add verification (BLT-DV) to diffusion drafting (ablation) | BLT-D-16 3B · FLORES-101 Fr-En 4-shot (3B) | BLEU: 34.05 -> 35.23 | Improved |
| add verification (BLT-DV) to diffusion drafting (ablation) | BLT-D-16 3B · HumanEval 0-shot (3B) | pass@1: 9.76 -> 14.02 | Improved |
| add verification (BLT-DV) to diffusion drafting (ablation) | BLT-D-16 3B · FLORES-101 Fr-En 4-shot (3B) | memory bandwidth (GB): 233.87 -> 568.61 | Hurt |
| add self-speculation (BLT-S, k=16) to BLT (ablation) | BLT 3B · FLORES-101 Fr-En 4-shot (3B) | memory bandwidth (GB): 1920.99 -> 727.17 | Improved |
| add self-speculation (BLT-S, k=16) to BLT (ablation) | BLT 3B · FLORES-101 Fr-En 4-shot (3B) | BLEU: 40.72 -> 40.72 | No clear effect |
| increase speculation window k from 4 to 16 (ablation) | BLT-S 3B · FLORES-101 Fr-En 4-shot (3B) | acceptance rate (%): 94.93 -> 69.93 | Hurt |
| increase diffusion block size from 4 to 16 (ablation) | BLT-D 3B (alpha=0.7) · HumanEval 0-shot (3B) | pass@1: 18.90 -> 9.76 | Hurt |
| increase diffusion block size from 4 to 16 (ablation) | BLT-D 3B (alpha=0.7) · FLORES-101 Fr-En 4-shot (3B) | memory bandwidth (GB): 797.58 -> 233.87 | Improved |
| raise confidence threshold alpha from 0.5 to 0.7 (ablation) | BLT-D-16 3B · FLORES-101 Fr-En 4-shot (3B) | BLEU: 31.64 -> 34.05 | Improved |
| use EB sampling (gamma=0.8) instead of confidence alpha=0.5 (ablation) | BLT-D-4 1B · HumanEval 0-shot (1B) | pass@1: 8.54 -> 10.37 | Improved |
| add block diffusion objective to BLT training (BLT-D-16) (ablation) | BLT 3B · MMLU (likelihood, 3B) | accuracy: 41.15 -> 37.08 | Hurt |
| add block diffusion objective to BLT training (BLT-D-4) (ablation) | BLT 3B · PIQA (likelihood, 3B) | accuracy: 79.38 -> 79.60 | No clear effect |
Closest papers by meaning
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.89)
- DFlash: Block Diffusion for Flash Speculative Decoding (ICML 2026 · similarity 0.89)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.88)
- Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion (NeurIPS 2026 · similarity 0.88)
- LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding (ICML 2026 · similarity 0.88)
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.
