CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation

CHORUS：用于高覆盖率测试平台激励生成的互补专家模型

Abstract: Large language models (LLMs) have advanced code generation, where executable feedback provides a more reliable learning signal than textual imitation alone. Hardware verification is an important application of code generation and accounts for a substantial fraction of modern chip design effort, with high-coverage testbench stimulus generation as a key task.

摘要： 大型语言模型（LLM）推动了代码生成技术的发展，其中可执行反馈比单纯的文本模仿提供了更可靠的学习信号。硬件验证是代码生成的一个重要应用领域，占据了现代芯片设计工作量的很大一部分，而高覆盖率的测试平台（Testbench）激励生成则是其中的关键任务。

We present CHORUS, a post-training framework that pushes performance beyond what a conventional supervised fine-tuning (SFT)-to-reinforcement learning (RL) pipeline achieves. CHORUS builds on two observations. First, staged SFT produces behaviorally diverse checkpoints, and dense-reward RL turns them into strong experts with comparable aggregate performance but distinct task-level strengths.

我们提出了 CHORUS，这是一个训练后框架，其性能超越了传统的“监督微调（SFT）转强化学习（RL）”流程。CHORUS 基于两个观察结果：首先，分阶段的 SFT 会产生行为各异的检查点（checkpoints），而密集奖励的强化学习将它们转化为强大的专家模型，这些模型在总体性能上相当，但在任务层面上各具优势。

Second, these complementary strengths can be exploited through either training-free model merging or further post-training to outperform the best individual expert. By consolidating the resulting specialists into a single 4B model, CHORUS achieves 88.0% Pass@1 on CVDP-ECov, outperforming DeepSeek-R1 (671B) by 13.5 percentage points.

其次，通过无需训练的模型合并或进一步的训练后处理，可以利用这些互补优势，从而超越表现最好的单一专家模型。通过将由此产生的专家模型整合到一个 4B 参数的模型中，CHORUS 在 CVDP-ECov 测试集上达到了 88.0% 的 Pass@1 准确率，比 DeepSeek-R1 (671B) 高出 13.5 个百分点。

---

Paper Details:

• Authors: Hejia Zhang, Sheng Lu, Zhongming Yu, Chia-Tung Ho, Brucek Khailany, Jishen Zhao
• arXiv ID: 2608.10090
• Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)

论文详情：

• 作者： Hejia Zhang, Sheng Lu, Zhongming Yu, Chia-Tung Ho, Brucek Khailany, Jishen Zhao
• arXiv ID: 2608.10090
• 学科分类： 人工智能 (cs.AI)；机器学习 (cs.LG)