Search or Accelerate: Confidence-Switched Position Beam Search for Diffusion Language Models
This addresses decoding inefficiencies in DLMs for tasks like mathematical reasoning and code generation, offering a practical balance between quality and efficiency, though it is incremental as it builds on existing DLM frameworks.
The paper tackles the problem of suboptimal decoding in Diffusion Language Models (DLMs) due to greedy unmasking, which can lead to poor performance on reasoning-heavy prompts. It introduces SOAR, a training-free decoding algorithm that adapts to model uncertainty, improving generation quality on benchmarks like GSM8K, MBPP, and HumanEval while maintaining competitive inference speed.
Diffusion Language Models (DLMs) generate text by iteratively denoising a masked sequence, repeatedly deciding which positions to commit at each step. Standard decoding follows a greedy rule: unmask the most confident positions, yet this local choice can lock the model into a suboptimal unmasking order, especially on reasoning-heavy prompts. We present SOAR, a training-free decoding algorithm that adapts its behavior to the model's uncertainty. When confidence is low, SOAR briefly widens the search over alternative unmasking decisions to avoid premature commitments; when confidence is high, it collapses the search and decodes many positions in parallel to reduce the number of denoising iterations. Across mathematical reasoning and code generation benchmarks (GSM8K, MBPP, HumanEval) on Dream-7B and LLaDA-8B, SOAR improves generation quality while maintaining competitive inference speed, offering a practical way to balance quality and efficiency in DLM decoding.