Efficient Diffusion LLMs via Temporal-Spatial Parallel Decoding and Confidence Extrapolation

2026-05-28

AI Summary

This work addresses the inefficiency of diffusion-based large language model inference, which stems from redundant iterative denoising and repeated remasking of already-converged tokens. The authors formulate the decoding process as a dynamic control problem and propose a trajectory-aware decoding framework. By analyzing token-level denoising trajectories—such as confidence, entropy, and momentum—the method identifies convergence states and employs a training-free confidence extrapolation mechanism for proactive decision-making. Integrating spatiotemporal parallel decoding with KV cache-compatible optimizations, this approach significantly reduces unnecessary iterations without compromising generation quality, thereby substantially improving inference efficiency.

Abstract

Diffusion-based large language models (dLLMs) support parallel text generation via iterative denoising, yet inference remains latency-heavy because many steps are spent on redundant refinement and repeated remasking of tokens whose final values are already determined. Prior acceleration methods mainly depend on step-local confidence heuristics or fixed schedules, which are sensitive to prompt and task variation and ignore strong positional effects within a sequence. We cast diffusion decoding as a dynamic control problem and show that token-wise denoising trajectories provide the key signal for reliable control. We propose a trace-aware decoding framework with two components. First, Temporal-Spatial Parallel Decoding (TSPD) uses a lightweight temporalspatial controller that consumes per-token trajectory features, including confidence, entropy, and momentum, together with token position, to decide when a token has converged and can be safely fixed. Second, we introduce Confidence Extrapolation (CE), a training-free state-space module that forecasts future logit trends with uncertainty to support proactive decisions, including safe look-ahead and targeted stabilization when trajectories are oscillatory or underconfident. Together, TSPD and CE reduce unnecessary denoising iterations while preserving output quality, and they compose cleanly with system optimizations such as KV caching.

Problem

• diffusion LLMs
• parallel decoding
• inference latency
• token convergence
• confidence estimation

Innovation

• Temporal-Spatial Parallel Decoding
• Confidence Extrapolation
• diffusion LLMs
• trajectory-aware decoding
• dynamic control