Beyond Scattered Acceptance: Fast and Coherent Inference for DLMs via Longest Stable Prefixes

arxiv:2603.05454

Original source · Paper JSON

LLMs: architectures, context, training · 2026-03-05

Cites (1) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• Beyond Scattered Acceptance : Fast and Coherent Inference for DLMs via Longest Stable Prefixes
• Abstract
• 1 Introduction
• 2 Related Work
• 2.1 Diffusion Large Language Model
• 2.2 Acceleration Methods for Diffusion Language Models
• 3 Method
• 3.1 Preliminary
• Forward Corruption Process.
• Reverse Generation Process.
• The Inefficiency of Conventional Schedulers.
• 3.2 The Longest Stable Prefix (LSP) Scheduler
• Single-Pass Prediction and Stability Assessment.
• Distinction from Blockwise Autoregressive Decoding.
• Approximate KV Caching for Prefix States.
• Targeted Block Sizing via Adaptive Thresholding.
• Structural Coherence via Boundary Snapping and Monotone Progress.
• 4 Experiments
• 4.1 Experimental Setup
• Models and Benchmarks.
• Baseline and LSP Configuration.
• 4.2 Main Results and Analysis
• Global Planning in Creative Writing.
• 4.3 Ablation Studies and Analysis
• 4.3.1 The Critical Role of Adaptive Sizing
• Motivation.
• Analysis.
• 4.3.2 Enhancing Coherence with Structural Snapping
• Motivation.
• Analysis.
• Qualitative Evidence.
• 4.3.3 Prefix-First vs. Scattered: The Power of Topology
• Motivation.
• Analysis.
• 4.3.4 Quantifying Repair Costs via Token Flip Rate
• 5 Conclusion
• Limitations