Research map / Architectures and efficiency / Attention and transformers
DiffuMamba: High-Throughput Diffusion LMs with Mamba Backbone
Vaibhav Singh, Oleksiy Ostapenko, Pierre-Andre Noel, Eugene Belilovsky, Torsten Scholak
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 one attention mixer every N=5 Mamba mixers (hybrid) | DiffuMamba · DCLM validation (Quokka budget) | validation perplexity: 21.41 -> 20.17 | Improved |
| replace attention mixers with bidirectional Mamba | DiffuTran · DCLM validation (Quokka budget) | validation perplexity: 22.72 -> 21.41 | Improved |
| use hybrid Mamba-attention instead of pure attention | DiffuTran 1.3B · Zero-shot PPL avg | perplexity: 55.43 -> 51.18 | Improved |
| use Mamba backbone instead of attention at small scale | DiffuTran 240M · Zero-shot PPL avg | perplexity: 81.79 -> 78.11 | Mixed |
| use Mamba instead of attention at 0.5B | DiffuTran 0.5B · Zero-shot PPL avg | perplexity: 61.29 -> 66.01 | Hurt |
| Mamba-based backbone for downstream zero-shot tasks | DiffuTran 1.3B · Downstream average accuracy | accuracy: 33.81 -> 38.12 | Improved |
| hybrid versus pure Mamba on downstream tasks | DiffuMamba 1.3B · Downstream average accuracy | accuracy: 38.12 -> 38.24 | No clear effect |
| evaluate at double training length (L=2048 vs 1024) | DiffuTran 1.3B · Length generalization | avg zero-shot PPL: 32.20 -> 56.57 | Hurt |
| evaluate at double training length (L=2048 vs 1024) | DiffuMamba-H 1.3B · Length generalization | avg zero-shot PPL: 28.88 -> 28.53 | No clear effect |
| evaluate at double training length (L=2048 vs 1024) | DiffuMamba-H 1.3B · Length generalization | MAUVE: 0.744 -> 0.796 | Improved |
| add hybrid interleaving (N=5) to pure attention (ablation) | DiffuTran 0.5B · Hybrid interleaving ablation | MAUVE: 0.672 -> 0.719 | Improved |
| add hybrid interleaving (N=5) to pure Mamba (ablation) | DiffuMamba 0.5B · Hybrid interleaving ablation | MAUVE: 0.674 -> 0.719 | Improved |
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.90)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.90)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.90)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.90)
- Theoretical Benefit and Limitation of Diffusion Language Model (NeurIPS 2025 · 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.
