Research map / Large language models / Language, safety and interpretability
Diffusion Language Model Parallel Decoding via Product-of-Experts Bridge
Juntong Shi, Brian L. Trippe, Jure Leskovec, Stefano Ermon, Minkai Xu
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 importance sampling over K=4 candidates (ablation) | PoE-Bridge · GSM8K | accuracy: 95.20 -> 95.30 | Improved |
| add importance sampling over K=4 candidates (ablation) | PoE-Bridge · MATH | accuracy: 73.86 -> 74.42 | Improved |
| add importance sampling over K=4 candidates (ablation) | PoE-Bridge · HumanEval | accuracy: 80.69 -> 79.47 | Hurt |
| add importance sampling over K=4 candidates (ablation) | PoE-Bridge · MBPP | accuracy: 72.20 -> 73.20 | Improved |
| add importance sampling over K=4 candidates (ablation) | PoE-Bridge · MATH | throughput (tokens/s): 99.49 -> 94.94 | Hurt |
| replace direct DLM-to-AR speculative decoding (w=0.0) with PoE bridge w=0.3 (ablation) | PoE-Bridge · HumanEval (K=1) | throughput (tokens/s): 56.55 -> 84.82 | Improved |
| replace direct DLM-to-AR speculative decoding (w=0.0) with PoE bridge w=0.3 (ablation) | PoE-Bridge · HumanEval (K=1) | % all-reject steps: 17.15 -> 4.53 | Improved |
| raise PoE weight w from 0.3 to 0.9 (ablation) | PoE-Bridge · HumanEval (K=1) | accuracy: 80.69 -> 30.49 | Hurt |
| raise PoE weight w from 0.3 to 0.9 (ablation) | PoE-Bridge · HumanEval (K=1) | throughput (tokens/s): 84.82 -> 168.35 | Improved |
| use Elastic Rejection Window V=32 instead of disabled (V=inf), K=4 (ablation) | PoE-Bridge · MATH | throughput (tokens/s): 39.46 -> 94.94 | Improved |
| use Elastic Rejection Window V=32 instead of V=8, K=1 (ablation) | PoE-Bridge · MATH | throughput (tokens/s): 66.22 -> 99.49 | Improved |
| use mixed-temperature candidates instead of uniform temperature with increasing K (ablation) | PoE-Bridge · MATH and MBPP | accuracy: at least one-third of the remaining gap closed | Improved |
Closest papers by meaning
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.90)
- Confidence-Based Decoding is Provably Efficient for Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.90)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · 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.
