Research map / Image generation / Fast diffusion sampling
Time Is a Feature: Exploiting Temporal Dynamics in Diffusion Language Models
Wen Wang, Bozhen Fang, Chenchen Jing, Yongliang Shen, Yangyi Shen, Qiuyu Wang, Hao Ouyang, Hao Chen, Chunhua Shen
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 exponential-weighted temporal voting | LLaDA-8B-Instruct · GSM8K (gen length 256) | accuracy: 76.3 -> 78.7 | Improved |
| add exponential-weighted temporal voting | LLaDA-8B-Instruct · Countdown (gen length 128) | accuracy: 20.3 -> 25.0 (+4.7) | Improved |
| use fixed weighting instead of exponential in temporal voting (ablation) | LLaDA-8B-Instruct · GSM8K (gen length 256) | accuracy: 76.3 -> 73.4 | Hurt |
| use linear weighting for temporal voting (ablation) | LLaDA-8B-Instruct · MATH500 (gen length 256) | accuracy: 33.4 -> 34.4 | Improved |
| vary exponential alpha from 1 to 11 in temporal voting (ablation) | LLaDA-8B-Instruct · four math benchmarks | average accuracy gain: peak average gain of 1.5% at alpha=5 | Improved |
| add negative TSE reward in RFT (no labels) | LLaDA-8B-Instruct SFT · Countdown (gen length 256) | accuracy: 19.9 -> 53.5 | Improved |
| replace accuracy reward (d1) with negative TSE reward | LLaDA-8B-Instruct SFT · Countdown | average accuracy gain: 15.1% (d1) -> 24.7% (TSE) | Improved |
| combine TSE reward with accuracy reward | LLaDA-8B-Instruct SFT · SVAMP (gen length 512) | accuracy: 88.7 -> 92.3 | Improved |
| combine TSE reward with accuracy reward | LLaDA-8B-Instruct SFT · Countdown (gen length 512) | accuracy: 21.5 -> 54.7 | Improved |
| add temporal voting after RFT with negative TSE reward | LLaDA-8B-Instruct RFT · GSM8K (gen length 128) | accuracy: 72.2 -> 72.6 | Improved |
| add temporal voting to the Sudoku-evaluated model | LLaDA-8B-Instruct SFT · Sudoku (gen length 512) | accuracy: 5.5 -> 2.8 | Hurt |
| use negative TSE reward alone in RFT | LLaDA-8B-Instruct SFT · Sudoku (gen length 512) | accuracy: 5.5 -> 3.3 | Hurt |
Closest papers by meaning
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.90)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.90)
- TEST-TIME SCALING IN DIFFUSION LLMS VIA HIDDEN SEMI-AUTOREGRESSIVE EXPERTS (ICLR 2026 · similarity 0.90)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.90)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · 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.
