CHORUS: Post-Training Framework for High-Coverage Testbench Stimulus Generation
A new framework named CHORUS has been developed by researchers to improve the capabilities of large language models (LLMs) in creating comprehensive testbench stimuli for hardware verification. This framework employs staged supervised fine-tuning (SFT) to generate a variety of behaviorally distinct checkpoints, which are subsequently enhanced using dense-reward reinforcement learning (RL) to create proficient experts that, while maintaining similar overall performance, exhibit unique strengths for specific tasks. These complementary abilities can be utilized through either model merging without additional training or further post-training, enabling CHORUS to surpass the performance of any single expert. This research tackles the critical issue of hardware verification, which represents a significant portion of contemporary chip design efforts. The study can be found on arXiv with the identifier 2608.10090.
Key facts
- CHORUS is a post-training framework for LLMs.
- It focuses on high-coverage testbench stimulus generation.
- Hardware verification is a key application of code generation.
- Staged SFT produces behaviorally diverse checkpoints.
- Dense-reward RL turns checkpoints into strong experts.
- Experts have comparable aggregate performance but distinct strengths.
- Complementary strengths can be exploited via model merging or further post-training.
- The paper is available on arXiv (2608.10090).
Entities
Institutions
- arXiv