AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation
Authors: Prashanth Vijayaraghavan, Luyao Shi, Ehsan Degan, Vandana Mukherjee, Xin Zhang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results show that AUTOCIRCUIT-RL generates 12% more valid circuits and improves efficiency by 14% compared to the best baselines, while reducing duplicate generation rates by 38%. It achieves over 60% success in synthesizing valid circuits with limited training data, demonstrating strong generalization. These findings highlight the framework s effectiveness in scaling to complex circuits while maintaining efficiency and constraint adherence, marking a significant advancement in AI-driven circuit design. |
| Researcher Affiliation | Industry | 1IBM Almaden Research Center, San Jose, CA 95120, USA 2IBM Thomas J. Watson Research Center, Yorktown Heights, NY 10598. Correspondence to: Prashanth Vijayaraghavan <EMAIL>, Luyao Shi <EMAIL>, Xin zhang <EMAIL>. |
| Pseudocode | No | The paper describes methods in prose, such as in Section 3 'Proposed Approach' and its subsections, without presenting any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not explicitly state that the source code for AUTOCIRCUIT-RL is open-source or provide a link to a code repository for the methodology described. It only references external model repositories for baselines in Appendix A. |
| Open Datasets | No | We generated a dataset of switching power converter topologies with 4 10 components (Figure 2(a)). Using Random Search (RS) (Fan et al., 2021), we generated numerous unique netlists. |
| Dataset Splits | Yes | For training, we randomly sample approximately 100,000 unique netlists (for 4and 5-component circuits)... We apply a weighted sampling strategy in each batch, prioritizing Group 4 (weight 0.4) and giving the lowest priority to Group 1 (weight 0.1). |
| Hardware Specification | Yes | We provide the implementation details of our experiments conducted with the official Py Torch v2.2.0 release binary package, compiled with CUDA 11.8, utilizing NVIDIA V100 GPUs with 32 GB of memory. |
| Software Dependencies | Yes | We provide the implementation details of our experiments conducted with the official Py Torch v2.2.0 release binary package, compiled with CUDA 11.8, utilizing NVIDIA V100 GPUs with 32 GB of memory. |
| Experiment Setup | Yes | Training is conducted over 4-6 epochs using shuffled data from the training split... We use the Adam W optimizer (Loshchilov & Hutter, 2017), setting beta parameters to 0.9 and 0.95, and an epsilon value of 1.0e-8. The learning rate is set to 0.95e-5, and the seed is fixed at 42 to ensure reproducibility. During training, we assess performance by evaluating a subset of 100 sample generations, using consistent evaluation settings. If the performance in the current epoch exceeds that of the previous one, we save the checkpoint. The RL convergence curves in Figure 3 are shown over 25,000 training steps with a batch size of 16. |