Trace and Chase: Empowering LLMs to Identify the Root Cause of Test Failures in RISC-V SoCs via Textualized Waveform
🚧 Work in Progress — WaveAce is under active development. The full codebase will be open-sourced after the formal paper is published. If you need early access for research or collaboration, please contact the author (see Contact).
WaveAce is an LLM-based Root Cause Analysis (RCA) framework for test failures in RISC-V System-on-Chip (SoC) designs. It bridges the modality gap between hardware waveforms and LLMs' textual reasoning, enabling LLMs to view waveforms and perform substantiated root cause analysis — not just fault localization.
- >35% of verification time is spent debugging test failures.
- Fault localization gives locations but lacks an explanatory causal chain.
- Symbolic methods need cycle-accurate oracles — impractical at SoC scale.
- LLM-based methods lack dynamic behavior awareness and hallucinate at SoC scale.
WaveAce comprises two core components:
| Component | Role |
|---|---|
| Tracer | A waveform textualization toolset. It interprets signals by type (enums, bus transactions, CPU instructions) and assembles them hierarchically using the software call tree as a backbone, producing a readable, semantic textual trace. |
| Chaser | An iterative RCA pipeline that alternates between waveform behavior analysis and code-based deviation attribution, cross-validating each modality until the root cause converges. |
- UVLLM benchmark (126 defects / 28 DUTs): WaveAce outperforms or matches the baseline in 79% of cases on pass@5.
- OpenTitan SoC (4 implanted + 1 real bug): WaveAce achieves 73% pass@5 across 5 realistic SoC bugs.
| Item | Status |
|---|---|
| Paper | Under formal submission |
| Code | Private — open-sourcing after paper publication |
| Poster | ✅ Available — trace-and-chase-dac2026-wip-poster.pdf |
- 📄 Poster — trace-and-chase-dac2026-wip-poster.pdf
WaveAce will be fully open-sourced after the formal paper is accepted and published. Until then, this repository serves as a project landing page with the poster attached.
If you are interested in early access to the code, please reach out to the author directly.
Yicheng Zhong
- 📧 Email: zhongyicheng241@mails.ucas.ac.cn
- 🐙 GitHub: @zhong-yic
Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences
Yicheng Zhong, Jingzheng Wu, Xiang Ling, Hao Lyu, Zhiqing Rui, Tianyue Luo, Chen Zhao
To be announced upon open-source release.