DiffuTester: Accelerating Diffusion LLM Unit Test Generation via Structural Pattern Mining

Research Background

Diffusion Large Language Models (dLLMs) generate multiple tokens in parallel through iterative denoising, offering stronger parallel generation potential than traditional autoregressive models. This makes them attractive for efficient code generation, especially unit test generation (UTG) where many test cases are needed to cover statements and branches.

However, dLLMs face a clear speed–quality trade‑off : keeping only a few high‑confidence tokens per denoising step preserves quality but underutilizes parallelism; decoding more tokens per step speeds up generation but degrades quality. This problem is particularly acute in UTG, where developers must produce numerous test cases for many methods.

A key observation is that unit tests for the same method share repetitive structural patterns (e.g., common setup, assertion skeletons). DiffuTester exploits this redundancy to let the model decode more tokens per step without sacrificing test quality.

Problem Statement

The core question: How to accelerate dLLM unit test generation without lowering generation quality? Existing training‑free acceleration methods fall into two categories: (1) cache‑based methods that reuse intermediate representations or KV Cache across denoising steps to reduce per‑step compute, and (2) sampling‑based techniques that decode more tokens per step or adjust sampling dynamically to cut total steps. Both categories typically degrade quality. DiffuTester belongs to the sampling‑based family but avoids quality loss and remains compatible with cache‑based methods.

Method: DiffuTester

DiffuTester is a training‑free acceleration framework for dLLM unit test generation. Its central idea: dynamically mine shared structures among multiple test cases in a batch and use those structures to retain additional tokens during decoding, thereby reducing the total number of denoising steps.

Decoding Pipeline

At each decoding step:

Confidence‑based decoding – the standard dLLM step keeps the tokens the model is most confident about.

Structure‑pattern decoding – DiffuTester builds Abstract Syntax Trees (ASTs) for the partially generated test cases in the current batch, merges the ASTs to identify shared structural nodes, and retains the tokens corresponding to those common structures. This allows the model to finalize more tokens in a single step.

Because early‑stage diffusion output may contain syntax errors, DiffuTester does not parse complete test cases. Instead it uses line‑by‑line parsing to limit the impact of local syntax errors on structure mining.

Auxiliary Designs

Confidence threshold – only tokens whose confidence exceeds a preset threshold are kept, preventing low‑confidence tokens from harming code quality.

Periodic AST analysis – structure mining is not run at every denoising step; it is performed every k steps to reduce overhead.

In summary, DiffuTester augments the original confidence decoding with structural information mined from multiple test cases, enabling more tokens to be decoded per step and achieving more efficient dLLM unit test generation.

Experimental Results and Analysis

Experiments were conducted on two representative dLLMs – DiffuCoder and Dream – using the TestEval benchmark extended to three programming languages: Python, C++, and Java .

Main Results: Significant Speedup with Preserved Quality

DiffuTester consistently improves generation efficiency across models and languages while maintaining peak coverage. The framework achieves up to 2–3× speedup without reducing test coverage.

Comparison with Sampling‑Based Acceleration Methods

Against other sampling‑based accelerators, DiffuTester delivers a better balance of speed and quality, outperforming them on both metrics.

Compatibility with Cache‑Based Methods

Cache‑based methods (e.g., KV Cache reuse) reduce per‑step compute, while DiffuTester reduces the number of steps. The two approaches are orthogonal and can be combined for additive gains.

Artifacts

Paper: https://arxiv.org/pdf/2509.24975 Code:

https://github.com/THU-Agent/DiffuTester