Tsinghua's DiffuTester accelerates dLLM unit test generation up to 3x via AST pattern mining
机器之心 · wechat · 2026-09-11
Tsinghua's AIAgent group proposes DiffuTester, a training-free framework that speeds up unit test generation with diffusion LLMs while preserving quality.
- Key insight: multiple tests for the same method share many structural code patterns.
- Method: after confidence-based decoding at each denoising step, DiffuTester builds ASTs line-by-line for in-flight tests, mines shared structures, and additionally retains the corresponding tokens — decoding more tokens per step and cutting total inference steps.
- Details: only tokens above a confidence threshold are kept, and AST analysis runs every few steps to limit overhead.
- Results: stable 2–3x speedups on DiffuCoder and Dream across Python, C++, and Java on TestEval, with peak coverage maintained; it also stacks with KV-cache-based acceleration.
Accepted to EMNLP 2026; code open-sourced at THU-Agent/DiffuTester.
More from coding & agent
- PyTorch Lightning checkpointing runs up to 95% faster on Google Cloud — LightningAI · 2026-09-11
- Rive launches CLI and RML text format so AI agents can build interactive designs — farooqib · 2026-09-11
- GitHub's Spec Kit hits 126K stars by forcing AI agents to spec before they code — blaizedsouza · 2026-09-11
- Agent sandboxing in three layers: primitives, runtime, platform — plus an AGENTS.md handbook playbook — blaizedsouza · 2026-09-11
- DIY AI accounts payable agent saved $183, more than its token costs — gregmushen · 2026-09-11
- Word-level diff versioning for AI agent edits, with full traces and citations — SnooPeripherals5313 · 2026-09-11