# ChipMATE: Multi-Agent Training via Reinforcement Learning for Enhanced RTL Generation

**Authors:** Zhongkai Yu, Yichen Lin, Chenyang Zhou, Yuwei Zhang, Kun Zhou, Junxia Cui, Haotian Ye, Zhengding Hu, Zaifeng Pan, Ruiyi Wang, Yujie Zhao, Hejia Zhang, Jingbo Shang, Jishen Zhao, Yufei Ding  
**arXiv:** [2605.12857](https://arxiv.org/abs/2605.12857)  
**Date:** 2026-05-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2605.12857v1/

## Key points

- ChipMATE is a multi-agent framework for RTL code generation that addresses limitations of existing API-based systems.
- It introduces a novel approach inspired by industrial practices, where correctness emerges from cross-comparison between independently written RTL modules and reference models.
- The framework pairs a Verilog agent with a Python reference-model agent for mutual verification of outputs without relying on a golden oracle.
- ChipMATE incorporates backtrack-based inference workflow to prevent error propagation and implements a two-stage training pipeline for agents.
- It utilizes a hybrid data-generation framework to produce high-quality reference model training samples and achieves impressive results in pass rates, outperforming existing models like DeepSeek V4.
- The authors have made the code and model weights publicly available on GitHub, demonstrating their commitment to transparency and reproducibility in research.

## Summary

, , , , 


ChipMATE is a groundbreaking multi-agent framework for RTL code generation that addresses the limitations of existing API-based systems. These systems often assume the availability of a golden testbench at generation time, rely on closed-source APIs that do not align with chip vendors' security requirements, and cannot be trained on vendors' proprietary RTL codebases. To bridge these gaps, ChipMATE introduces a novel approach inspired by industrial practices where correctness emerges from cross-comparison between independently written RTL modules and reference models. The framework pairs a Verilog agent with a Python reference-model agent, enabling mutual verification of outputs without relying on a golden oracle. ChipMATE incorporates a backtrack-based inference workflow to prevent error propagation across turns and implements a two-stage training pipeline. In the first stage, each agent is individually trained to maximize its code-generation capability, followed by joint training to enhance collaboration between the agents effectively. To support the training process, ChipMATE utilizes a hybrid data-generation framework that produces 64.4K high-quality reference model training samples. The framework achieves impressive results, with pass rates of 75.0% and 80.1% on VerilogEval V2 using 4B and 9B base models respectively. Notably, ChipMATE outperforms all existing self-trained models and even surpasses DeepSeek V4 with 1600B parameters. The authors have made both the code and model weights publicly available on GitHub (https://github.com/zhongkaiyu/ChipMATE), showcasing their commitment to transparency and reproducibility in research. With its innovative approach to multi-agent training via reinforcement learning, ChipMATE represents a significant advancement in RTL code generation that holds promise for enhancing industrial practices in this domain.