---
title: "Peking University's Saber Runs Diffusion Code Models 3.5x Faster Without Losing Quality"
subtitle: "A training-free sampling algorithm for diffusion language models delivers 251% faster code generation while lifting Pass@1 accuracy by 1.9%."
url: "https://alphasignal.ai/news/peking-university-s-saber-runs-diffusion-code-models-3-5x-faster-without-losing"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-15T00:00:00.000Z"
updated: "2026-08-23T21:23:06.659Z"
type: "Paper"
topics:
  - "Development"
  - "Llms"
subtopics:
  - "Code Generation"
  - "Structured Output"
upvotes: 4500
source_url: "https://arxiv.org/abs/2510.18165"
comments: 0
truncated: true
---

# Peking University's Saber Runs Diffusion Code Models 3.5x Faster Without Losing Quality

> A training-free sampling algorithm for diffusion language models delivers 251% faster code generation while lifting Pass@1 accuracy by 1.9%.

Diffusion language models were supposed to fix the slow, one-token-at-a-time bottleneck of autoregressive generation. They can fill in many masked positions per step and use bidirectional context, which suits structured outputs like code. In practice, cranking up the parallelism to actually harvest that speedup makes quality fall off a cliff. A new paper called [Saber](https://arxiv.org/abs/2510.18165), from Peking University and Alibaba's Tongyi Lab, proposes a training-free sampling recipe that breaks the trade-off for code generation, lifting Pass@1 by 1.9% on average while running about 3.5x faster.

## Why parallel decoding keeps collapsing

A diffusion language model (DLM) starts from a sequence of `[MASK]` tokens and iteratively unmasks positions across many denoising steps. Unmask one high-confidence token per step and you get quality similar to autoregressive models, but you also burn as many forward passes as the sequence is long. Unmask many tokens per step and you skip forward passes but risk committing to bad tokens.

The paper pins the collapse on two structural properties of DLM sampling:

* **Non-uniform difficulty.** Early in decoding, the context is mostly mask tokens, so the model is uncertain. Later, once scaffolding like function signatures, brackets, and variable names is in place, predictions become much easier. A static parallelism setting is either too timid at the end or too reckless at the start.
* **Irreversible errors.** Unlike autoregressive models, a DLM decides both which position to fill and what to put there. Once a token is unmasked in mainstream samplers, it is locked in forever. If an early guess conflicts with the context that emerges later, that error corrupts every subsequent prediction.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Saber: An Efficient Sampling with Adaptive Acceleration and Backtracking Enhanced Remasking for Diffusion Language Model](https://arxiv.org/abs/2510.18165)
- [Saber GitHub repository](https://github.com/zhaoyMa/Saber)
- [Large Language Diffusion Models (LLaDA)](https://arxiv.org/abs/2502.09992)
- [DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation](https://arxiv.org/abs/2506.20639)
- [Fast-dLLM: Training-Free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding](https://arxiv.org/abs/2505.22618)