Tsinghua's DiffuTester generates unit tests with diffusion LLMs 2-3x faster

jiqizhixin · x · 2026-09-21

Tsinghua University's AI Agent Lab (Prof. Jia Li's group) presents DiffuTester, accepted at EMNLP 2026, which uses diffusion LLMs to generate unit tests 2-3x faster without losing coverage.

Key idea: multiple unit tests for the same method share heavy structural redundancy. DiffuTester mines these patterns via Abstract Syntax Trees (AST) and preserves pattern-matching tokens during diffusion decoding, so each denoising step emits more tokens and total inference steps shrink. Unlike KV-cache reuse (per-step cost) or sampling tweaks, it targets redundancy specific to test generation.

Original post →

More from coding & agent

coding & agent channel →