Research map / Image generation / Fast diffusion sampling
Set Diffusion: Interpolating Token Orderings between Autoregression and Diffusion for Fast and Flexible Decoding
Marianne Arriola, Volodymyr Kuleshov
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 efficient token-level training (all L conditionals per ordering) (ablation) | SW-SetDLM S<=L · LM1B | test perplexity: 54.49 -> 42.02 | Improved |
| add efficient token-level training (ablation) | SW-SetDLM S<=L · LM1B | gradient variance: 7.50 x 10^-2 -> 2.73 x 10^-2 | Improved |
| add causal mask (KV-cache-compatible architecture) (ablation) | SW-SetDLM S<=L · LM1B | test perplexity: 46.00 -> 42.02 | Improved |
| add causal mask (ablation) | SW-SetDLM S<=L · LM1B | gradient variance: 15.01 x 10^-2 -> 2.73 x 10^-2 | Improved |
| add efficient training and causal mask with KV caching (ablation) | SW-SetDLM S<=L · GSM8K (order-agnostic schedule) | 0-shot pass@1: 6.37 -> 26.61 | Improved |
| add efficient training and causal mask with KV caching (ablation) | SW-SetDLM S<=L · GSM8K (order-agnostic schedule) | decoding throughput: 24.48 -> 40.83 | Improved |
| replace block diffusion S=16 with set diffusion S<=32 | BD3LM · GSM8K | 0-shot pass@1: 50.49 -> 61.94 | Improved |
| replace block diffusion S=8 with set diffusion S<=16 | BD3LM · GSM8K | decoding throughput: 54.67 -> 56.50 | Improved |
| replace block diffusion S=4 with set diffusion S<=8 | BD3LM · GSM8K | 0-shot pass@1: 63.53 -> 66.41 | Improved |
| replace block diffusion S=4 with set diffusion S<=8 | BD3LM · GSM8K | perplexity: 1.41 -> 1.42 | No clear effect |
| replace block diffusion with set diffusion S<=32 | BD3LM S=16 · ROCStories infilling 1/5 | ROUGE-1: 9.2 -> 11.6 | Improved |
| replace block diffusion with set diffusion S<=32 | BD3LM S=16 · ROCStories infilling 1/5 | decoding throughput: 105.8 -> 132.0 | Improved |
Closest papers by meaning
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.92)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.92)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.92)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.92)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion (NeurIPS 2026 · similarity 0.91)
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.
