Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models

TL;DR - Ripple-Pivot Search is a training-free parallel decoding method for diffusion LLMs that proactively commits selected mid-entropy tokens to reduce uncertainty elsewhere. It accelerates inference substantially while preserving generation quality.

• Uses lookahead evaluation to select pivot positions and token assignments with the greatest downstream benefit.
• Achieves 4–10× wall-clock speedups across three dLLMs and four reasoning/code benchmarks.
• Improves accuracy over a prior lookahead baseline by up to 5.49%.
• Combined with KV caching, reaches up to an 18× speedup over standard decoding.