Research map / Large language models / Language, safety and interpretability
Early Decisions Matter: Proximity Bias and Initial Trajectory Shaping in Non-Autoregressive Diffusion Language Models
Jiyeon Kim, Sungik Choi, Yongrae Jo, Moontae Lee, Minjoon Seo
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 learned planner for first-step position selection (ablation) | Top-1 Prob · GSM8K (T=32) | accuracy: 46.6 -> 55.0 | Improved |
| add learned planner for first-step position selection (ablation) | Top-1 Prob · Sudoku (T=32) | accuracy: 71.2 -> 65.2 | Hurt |
| add EOS temperature annealing (ablation) | Top-1 Prob · GSM8K (T=32) | accuracy: 46.6 -> 50.9 | Improved |
| add EOS temperature annealing (ablation) | Top-1 Prob · Countdown (T=32) | accuracy: 42.2 -> 46.1 | Improved |
| add planner and EOS annealing together | Top-1 Prob · Average of 4 tasks (T=32) | accuracy: 44.8 -> 47.6 | Improved |
| add planner and EOS annealing together | Prob Margin · Average of 4 tasks (T=32) | accuracy: 45.0 -> 49.1 | Improved |
| replace random initial position with learned planner (ablation) | Init. Position · GSM8K (T=32) | accuracy: 52.6 -> 55.0 | Improved |
| randomize initial position uniformly (ablation) | Top-1 Prob · Sudoku (T=32) | accuracy: 71.2 -> 48.3 | Hurt |
| randomize initial position uniformly (ablation) | Top-1 Prob · GSM8K (T=32) | accuracy: 46.6 -> 52.6 | Improved |
| increase diffusion steps in NAR confidence decoding (ablation) | Top-1 Prob · Average of 4 tasks | accuracy: 44.8 -> 39.7 -> 37.7 | Hurt |
| add planner and EOS annealing at higher steps | Top-1 Prob · GSM8K (T=128) | accuracy: 23.3 -> 61.0 | Improved |
| add planner and EOS annealing at higher steps | Top-1 Prob · Average of 4 tasks (T=128) | accuracy: 37.7 -> 52.8 | Improved |
Closest papers by meaning
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.90)
- Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.90)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Confidence-Based Decoding is Provably Efficient for Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Latent Refinement Decoding: Enhancing Diffusion Language Models by Refining Belief States (NeurIPS 2026 · similarity 0.90)
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · 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.
