LEAP technique cuts dLLM decoding steps by 30% without training
New training-free method detects early-converging tokens to unlock faster parallel decoding.
A team of researchers from Shanghai Jiao Tong University has introduced LEAP (Lookahead Early-Convergence Token Detection for Accelerated Parallel Decoding), a novel method that dramatically speeds up Diffusion Language Models (dLLMs). Current dLLMs rely on high-confidence thresholds to ensure conditional independence for parallel decoding, but this conservative approach limits parallelism. Through systematic token-level statistical analysis, the team discovered that many tokens converge to correct predictions early in the denoising process without meeting standard confidence criteria. LEAP addresses this by employing a training-free, plug-and-play mechanism that uses future context filtering and multi-sequence superposition to detect these early-converging tokens, enabling reliable early decoding.
Benchmark results are compelling: LEAP reduces the average number of denoising steps by approximately 30% compared to confidence-based decoding. On the GSM8K math reasoning dataset, combining LEAP with dParallel accelerates decoding to 7.2 tokens per step while maintaining model precision. The method breaks the reliance on high-confidence priors, offering a new paradigm for parallel decoding in dLLMs. This work has been published on arXiv (2605.10980) and represents a significant step toward making diffusion-based language models practical for real-time applications where latency matters.
- LEAP is training-free and plug-and-play, requiring no model modification or additional compute overhead.
- Reduces average denoising steps by ~30% compared to standard confidence-based decoding.
- On GSM8K, combined with dParallel, achieves 7.2 tokens per step while preserving model precision.
Why It Matters
Makes diffusion LLMs practical for real-time applications by slashing inference latency without sacrificing accuracy.