Research map / Image generation / Fast diffusion sampling

SimSD: Simple Speculative Decoding in Diffusion Language Models

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Masked diffusion language model decoding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SimSD, a training-free temporal causal attention mask with reference tokens and RoPE copy that lets diffusion LLMs verify drafted tokens token-by-token in one forward pass, yielding up to 7.46x throughput on SDAR models.

Problem

Diffusion LLMs use mask tokens and bidirectional attention, so the effective context changes across denoising steps and standard token-level speculative verification cannot be applied directly.

Approach

Duplicate the response into paired data (reference, drafted tokens) and mask sections, apply a token-level temporal attention mask keyed to a draft-induced step map, and copy RoPE indices from data to mask slots so one target forward pass yields valid per-position logits for rejection-sampling verification.

Key results

• SimSD reaches 71.6 tokens/s average at block length 4, a 7.46x speedup over vanilla TP-2 decoding.
• Average accuracy rises from 0.696 (vanilla) to 0.713 (SimSD) at block length 4, and from 0.681 to 0.698 at block length 8.
• Removing aligned RoPE drops GSM8K accuracy from 84.0% to 0.0% and token-level acceptance from 0.902 to 0.718.

Surprising findings

• Removing RoPE alignment collapses GSM8K accuracy to 0.0% despite a moderate drop in acceptance rate.
• Larger draft length gamma improves accuracy and acceptance rate but does not monotonically improve throughput.

Limitations

• The single-forward verification is an approximation of sequential denoising; KL divergence grows across steps (up to 0.50 in cross-model setting).
• Evaluated only on the SDAR model family with two GPUs at batch size 1.
• Throughput does not scale monotonically with draft length gamma, requiring tuning.

Experimental setup

Datasets: GSM8K MBPP TriviaQA MMLU

Benchmarks: GSM8K MBPP TriviaQA MMLU

Metrics: decoded tokens per second speedup accuracy pass@1 token-level acceptance rate KL divergence

Baselines: Vanilla Vanilla + CUDA Graph S2D2

Builds on: speculative decoding block diffusion S2D2

Closest papers by meaning

• Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.93)
• DFlash: Block Diffusion for Flash Speculative Decoding (ICML 2026 · similarity 0.93)
• A Theoretical Perspective for Speculative Decoding Algorithm (NeurIPS 2024 · similarity 0.91)
• DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
• Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.91)
• PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models (NeurIPS 2026 · similarity 0.91)