Research map / Image generation / Fast diffusion sampling
IDLM: Inverse-distilled Diffusion Language Models
David Li, Nikita Gushchin, Dmitry Abulkhanov, Eric Moulines, Ivan Oseledets, Maxim Panov, Alexander Korotin
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 |
|---|---|---|---|
| distill MDLM with IDLM to 16 steps | MDLM · OpenWebText | sampling steps: 1024 -> 16 steps (64x) | Improved |
| distill Duo-DCD with IDLM (greedy sampler) | Duo-DCD · OpenWebText | sampling steps: 1024 -> 4 steps (256x) | Improved |
| distill Duo with IDLM (greedy sampler) | Duo · OpenWebText | sampling steps: 1024 -> 16 steps (64x) | Improved |
| distill SEDD with IDLM | SEDD · OpenWebText | sampling steps: 1024 -> 256 steps (4x) | Improved |
| distill MDLM with IDLM, 128 steps vs 1024-step teacher | MDLM · GSM8K (TinyGSM) | accuracy (%): 18.0 -> 19.86 | Improved |
| distill Duo with IDLM, 64 steps vs 1024-step teacher | Duo · GSM8K (TinyGSM) | accuracy (%): 17.2 -> 19.03 | Improved |
| reduce IDLM-MDLM sampling steps from 128 to 32 (ablation) | IDLM-MDLM · GSM8K (TinyGSM) | accuracy (%): 19.86 -> 12.81 | Hurt |
| reduce IDLM-Duo sampling steps from 128 to 32 (ablation) | IDLM-Duo · GSM8K (TinyGSM) | accuracy (%): 21.38 -> 15.39 | Hurt |
| add Gaussian input noise to IDLM-MDLM (ablation) | IDLM-MDLM · OpenWebText | GenPPL/Entropy | No clear effect |
| use full IDLM-Duo update instead of stop-gradient variant (ablation) | IDLM-Duo · OpenWebText | GenPPL/Entropy stability | Improved |
| use IDLM instead of SDTT on 0.9B MDLM | 0.9B MDLM · OpenWebText | GenPPL/Entropy at low steps | Improved |
Closest papers by meaning
- Energy-Based Diffusion Language Models for Text Generation (ICLR 2025 · similarity 0.91)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.90)
- Theoretical Benefit and Limitation of Diffusion Language Model (NeurIPS 2025 · similarity 0.90)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.90)
- The Diffusion Duality (ICML 2025 · similarity 0.90)
- SemanticDLM+: Improving Diffusion LLMs through Bias-variance Trade-off in Transition Kernel Design (NeurIPS 2026 · similarity 0.89)
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.
