Research map / Image generation / Fast diffusion sampling
Planner Aware Path Learning in Diffusion Language Models Training
Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Michael M. Bronstein, Anru Zhang, Joey Bose, Alexander Tong
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 |
|---|---|---|---|
| train with PAPL instead of standard masked diffusion loss | DLM-150M protein model · protein generation | foldability (%): 42.43 -> 59.40 | Improved |
| train with PAPL instead of standard loss | DLM-150M protein model · protein generation | diversity (%): 92.45 -> 91.73 | No clear effect |
| train with PAPL instead of standard MDLM loss | MDLM on OpenWebText · OWT unconditional, T=32 | generative perplexity: 100.45 -> 40.19 | Improved |
| train with PAPL instead of standard MDLM loss | MDLM on OpenWebText · OWT unconditional, T=128 | MAUVE: 0.015 -> 0.067 | Improved |
| train with PAPL instead of standard loss | DLM 0.5B code model · HumanEval | pass@10: 31.1 -> 38.4 | Improved |
| train with PAPL instead of standard loss | DLM 0.5B code model · MBPP | pass@1: 17.6 -> 16.7 | Hurt |
| train with PAPL instead of standard loss | DLM 0.5B code model · MBPP+ | pass@10: 32.6 -> 53.6 | Improved |
| train with PAPL instead of standard loss | DLM 0.5B code model · HumanEval-Infill | pass@1: 30.0 -> 32.5 | Improved |
| lower softmax temperature tau below default 1 (ablation) | PAPL · protein foldability | foldability | Improved |
| increase path-learning weight alpha up to 5 (ablation) | PAPL · protein foldability | foldability | Improved |
| increase alpha beyond 5 (ablation) | PAPL · protein foldability | foldability | Hurt |
| interpolate planner-weighted loss with vanilla DLM loss (ablation) | PAPL · training stability | training variance | Improved |
Closest papers by meaning
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (NeurIPS 2026 · similarity 0.91)
- Think while You Generate: Discrete Diffusion with Planned Denoising (ICLR 2025 · similarity 0.91)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.90)
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.89)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 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.
