Research map / Image generation / Fast diffusion sampling
Self-Speculative Decoding Accelerates Lossless Inference in Any-Order and Any-Subset Autoregressive Models
Gabe Guo, Stefano Ermon
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 any-subset speculative decoding with self as draft | finetuned AS-ARM · WikiText 95% masked | Model NFE: 486.0 -> 434.1 | Improved |
| use any-subset speculative decoding with self as draft | finetuned AS-ARM · WikiText 95% masked | Gen PPL: 107.9 -> 107.6 | No clear effect |
| use context n-gram draft instead of self draft | ASSD · WikiText 95% masked | Time (s): 16.50 -> 16.80 | Hurt |
| use speculative decoding on off-the-shelf XLNet | XLNet · WikiText 95% masked | Time (s): 18.04 -> 9.36 | Improved |
| speculate only k=2 tokens (ablation) | ASSD · WikiText length 256 | speedup (%): 8.8 -> -2.7 | Hurt |
| increase speculated tokens k from 4 to 20 (ablation) | ASSD · WikiText length 128 | speedup (%): 8.8 -> 9.0 | No clear effect |
| increase sequence length from 128 to 1024 (ablation) | ASSD · WikiText 95% masked | speedup (%): 8.94 -> 11.00 | Improved |
| add KV cache (ablation) | XLNet 340M ASSD · infill 128 of 1024 | time per NFE (s): 0.1407 -> 0.0706 | Improved |
| add epsilon acceptance tolerance 0.1 (ablation) | ASSD · WikiText length 256 | Spec PPL: 122.70 -> 185.26 | Hurt |
| add epsilon acceptance tolerance 0.1 (ablation) | ASSD · WikiText length 256 | speedup (%): 8.28 -> 14.22 | Improved |
| apply speculative decoding to left-to-right prompting | XLNet · WikiText prefix prompting | Time (s): 14.3 -> 13.3 | Improved |
| finetune on wider masking ratios instead of pretrained weights (ablation) | AS-ARM · ROCStories infill 1/5 | ROUGE-1: 14.4 -> 13.1 | Hurt |
Closest papers by meaning
- Speculative Speculative Decoding (ICLR 2026 · similarity 0.90)
- A Theoretical Perspective for Speculative Decoding Algorithm (NeurIPS 2024 · similarity 0.90)
- DFlash: Block Diffusion for Flash Speculative Decoding (ICML 2026 · similarity 0.90)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.90)
- Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment (ICLR 2025 · similarity 0.90)
- Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios (AAAI 2026 · similarity 0.90)
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.
