Research map / Image generation / Fast diffusion sampling
From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models
Hengyu Fu, Baihe Huang, Virginia Adams, Charles Wang, Junkeun Yi, Mohammad Mahdi Kamani, Venkat Krishna Srinivasan, Jiantao Jiao
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 cross-block parallel decoding and look-ahead beam-search exploration (ETE) to confidence-based decoding | LLaDA-8B-Instruct · GSM8K, HumanEval, MATH, MMLU-Pro | inference steps at fixed accuracy: over 30% fewer steps at almost all accuracy levels | Improved |
| add ETE with KV caching to confidence-based decoding | LLaDA-8B-Instruct with Fast-DLLM · GSM8K | throughput at comparable accuracy: 50% higher throughput | Improved |
| add ETE with KV caching to confidence-based decoding | LLaDA-8B-Instruct with Fast-DLLM · MMLU-Pro | throughput at comparable accuracy: up to 70% higher throughput | Improved |
| add ETE with KV caching to confidence-based decoding | LLaDA-8B-Instruct with Fast-DLLM · MATH | throughput at comparable accuracy: 30% higher throughput | Improved |
| add ETE with KV caching to confidence-based decoding | LLaDA-8B-Instruct with Fast-DLLM · GSM8K | peak accuracy: 0.779 -> 0.797 | Improved |
| add ETE with KV caching to confidence-based decoding | LLaDA-8B-Instruct with Fast-DLLM · MMLU-Pro | peak accuracy: 0.352 -> 0.386 | Improved |
| lower the confidence factor f in dynamic confidence decoding (stricter threshold) (ablation) | dynamic confidence decoding · GSM8K (200 samples) | decoding rounds | Hurt |
| batch the beam-search exploration candidates with shared KV cache (ablation) | ETE · LLaDA-8B-Instruct forward latency | wall-clock time per batched forward: free lunch up to about 2 on H100, 4 on H200, 7 on B200 | No clear effect |
| exploration (low-confidence token commits) versus exploitation rounds in confidence decoding | dynamic confidence decoding · GSM8K (200 samples) | share of information per share of rounds: exploration 23.5% of rounds, 66.7% of bits at f=0.4 | Improved |
Closest papers by meaning
- Confidence-Based Decoding is Provably Efficient for Diffusion Language Models (NeurIPS 2026 · similarity 0.93)
- PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.92)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.92)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
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.
