Fast-dLLM++: An Accelerated Decoding Method for Diffusion Language Models Based on Fréchet Distance Analysis

A study accepted by the ICML 2026 SPIGM Workshop proposes Fast-dLLM++, which significantly accelerates the inference of diffusion language models (dLLMs) through Fréchet Contour Decoding technology, providing new ideas for efficient generative AI. This research was published by kasakh on GitHub (Project link: https://github.com/kasakh/fast-dllm-plus-page, release date: 2026-06-08).

Diffusion models have made breakthroughs in the field of image generation, but their iterative denoising process leads to inference latency; when extended to language models (dLLMs), the efficiency bottleneck becomes more prominent. Traditional autoregressive models have limitations in sequential generation, while diffusion models support parallelism but require a large number of iterations. Against this background, how to accelerate dLLMs inference without sacrificing quality has become a focus, and Fast-dLLM++ emerged as a solution.

Fast-dLLM++ proposes the 'Fréchet Contour Decoding' technology: 1. Fréchet distance is a metric for measuring curve similarity. Treating diffusion denoising as latent space trajectories, necessary steps are identified by analyzing the Fréchet distance of these trajectories; 2. The contour decoding mechanism first models the diffusion process to extract the ideal path contour, and dynamically determines the computation intensity based on the deviation between the current state and the target contour during inference, focusing resources on areas that require fine-grained processing.

The implementation optimizations of Fast-dLLM++ include: 1. Trajectory analysis and prediction: Establish a statistical model of diffusion trajectories through reference sampling offline, providing a roadmap for online inference; 2. Adaptive step scheduling: Based on Fréchet analysis, use small steps in high curvature regions (high information density) and increase step size or skip redundant computations in low curvature regions; 3. Quality monitoring module: Continuously evaluate generation quality to ensure that the accelerated output is consistent with full sampling.

Fast-dLLM++ shows significant performance improvements in standard benchmark tests, typically achieving 2-10x inference speedup while maintaining generation quality; this method was accepted by the ICML 2026 SPIGM Workshop, and its academic value has been recognized.

The application value of Fast-dLLM++ includes: 1. Reduce deployment costs: Reduce computation steps, improve the experience of real-time applications (dialogue, code completion, translation), and lower service costs; 2. Expand scenarios: Enter latency-sensitive fields such as edge devices and mobile applications; 3. Promote technology integration: Cross-disciplinary combination of geometric analysis, statistical learning, and efficient computing, providing methodological inspiration for subsequent research.

Fast-dLLM++ needs to balance speed, quality, and generality; future explorations can include: 1. Automatically adjust acceleration strategies for specific scenarios; 2. Combine with compression technologies such as quantization, pruning, and knowledge distillation to further improve efficiency.

Fast-dLLM++ provides a new path for dLLMs inference acceleration through Fréchet distance analysis and contour decoding, with both theoretical innovation and practical value. Such efficient inference technologies will promote the democratization of large models. Interested users can visit the project page to obtain the paper and code.