Research map / Large language models / Language, safety and interpretability
Gumbel Distillation for Parallel Text Generation
Chi Zhang, Xixi Hu, Bo Liu, qiang liu
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 Gumbel Distillation | MDLM · OpenWebText | generative perplexity: 38.34 -> 34.33 | Improved |
| add Gumbel Distillation | MDLM · OpenWebText | MAUVE: 0.217 -> 0.282 | Improved |
| add Gumbel Distillation | MDLM · LM1B | generative perplexity: 78.74 -> 67.64 | Improved |
| add Gumbel Distillation | BD3-LM (L'=4) · LM1B | generative perplexity: 56.98 -> 46.06 | Improved |
| add Gumbel Distillation | BD3-LM (L'=4) · OpenWebText | MAUVE: 0.251 -> 0.304 | Improved |
| add Gumbel Distillation | MDLM · zero-shot QA average | accuracy (%): 34.3% to 36.1% | Improved |
| add Gumbel Distillation | BD3-LM · zero-shot QA average | accuracy (%): from 38.6% to 39.4% | Improved |
| add Gumbel Distillation | BD3-LM · Gemini-judged creativity (OWT) | LLM score: 2.75 -> 2.68 | Hurt |
| add Gumbel conditioning to Medusa heads | Medusa on Vicuna-7B · ShareGPT | accepted length: 1.745 -> 1.891 | Improved |
| add Gumbel conditioning to Medusa heads | Medusa on GPT-2-Small · OpenWebText | accepted length: 1.609 -> 1.691 | Improved |
| use parallel instead of sequential Gumbel extraction (ablation) | MDLM + Gumbel Distillation · LM1B | generative perplexity: 86.38 -> 67.64 | Improved |
| use Gumbel instead of Gaussian noise (ablation) | MDLM + Gumbel Distillation · LM1B | generative perplexity: 81.43 -> 67.64 | Improved |
Closest papers by meaning
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.87)
- Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding (ICML 2025 · similarity 0.87)
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.87)
- Learning to Keep a Promise: Scaling Language Model Decoding Parallelism with Learned Asynchronous Decoding (ICML 2025 · similarity 0.87)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.86)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.86)
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.
