Research map / Large language models / Language, safety and interpretability
d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation
Yu-Yang Qian, Junda Su, Lanxiang Hu, Peiyuan Zhang, Zhijie Deng, Peng Zhao, Hao Zhang
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 pseudo-trajectory distillation (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | TPF: 6.41 -> 7.55 | Improved |
| add pseudo-trajectory distillation (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 441.4 -> 517.7 | Improved |
| add curriculum noise (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | TPF and accuracy: TPF 7.55 -> 8.46; Acc 72.1 -> 69.8 | Mixed |
| add curriculum window (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 551.3 -> 637.7 | Improved |
| add multi-block decoding and early stopping (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 492.9 -> 637.7 | Improved |
| add early stopping on EOS (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 635.0 -> 637.7 | Improved |
| use curriculum noise 0.0->0.8 instead of fixed t=0.5 (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 521.7 -> 637.7 | Improved |
| use curriculum noise 0.0->0.5 instead of 0.2->0.5 (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 557.7 -> 616.9 | Improved |
| use curriculum window 16->32 instead of fixed 32 (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 536.0 -> 637.7 | Improved |
| start window curriculum at 16 instead of 0 or 24 (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 603.1 / 584.9 -> 637.7 | Improved |
| swap exponential AUP weighting for power or linear (ablation) | AUP metric ranking · GSM8K-CoT (0-shot), LLaDA-based | method ranking: d3LLM top in all nine configurations | No clear effect |
| raise AUP penalty alpha from 1 to 10 (ablation) | d3LLM-LLaDA · GSM8K-CoT (0-shot), LLaDA-based | AUP score: 659.4 -> 568.4 | Hurt |
Closest papers by meaning
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.92)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.92)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (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.
