Reproduce “QiMeng-SALV: Signal-Aware Learning for Verilog Code Generation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yang Zhang, Rui Zhang, Jiaming Guo, Lei Huang, Di Huang, Yunpu Zhao, Shuyao Cheng, Pengwei Jin and 5 more

QiMeng-SALV introduces a signal-aware reinforcement learning approach for Verilog code generation, enhancing the functional correctness of large language models. The method improves functional reward signal density during DPO training by leveraging individual correct output signal implementations within partially erroneous modules, achieving state-of-the-art pass@1 scores on multiple benchmarks like RTLLM v1.1 and VerilogEval.

Key insights

  • The parallel and structural characteristics of Verilog code enable the functional independence of individual output signals, allowing for partial correctness within a module.
  • Fine-grained, signal-level functional feedback provides a denser and more informative reward signal for reinforcement learning compared to sparse, module-level correctness.
  • Exploiting partial correctness within modules, where some signals are correct even if the overall module is flawed, greatly expands the usable training data for RL optimization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.