Research map / Graph learning / Graph self-supervised learning
Beyond Tokens: Enhancing RTL Quality Estimation via Structural Graph Learning
Yi Liu, Hongji Zhang, Yiwen Wang, Dimitris Tsaras, Lei Chen, Mingxuan Yuan, Qiang Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove structure-aware masked node modeling (ablation) | StructRTL · area prediction, w/o KD | R2: 0.7249 -> 0.7463 | Improved |
| add edge prediction pretraining (ablation) | StructRTL · area prediction, w/o KD | R2: 0.7018 -> 0.7463 | Improved |
| add knowledge distillation from post-mapping netlist predictor | StructRTL · delay prediction | R2: 0.7630 -> 0.8872 | Improved |
| add knowledge distillation from post-mapping netlist predictor | StructRTL · area prediction | R2: 0.7463 -> 0.8676 | Improved |
| use CDFG graph self-supervised model instead of Verilog-LLM embeddings | CodeV-QW-7B · area prediction, w/o KD | R2: 0.6353 -> 0.7463 | Improved |
| mask post-GNN embeddings instead of raw input features (ablation) | StructRTL · area and delay prediction | downstream R2 | Improved |
| add global positional embeddings (ablation) | StructRTL · area and delay prediction | downstream R2 | Improved |
| combine GNN with Transformer instead of either alone (ablation) | StructRTL · area and delay prediction | downstream R2 | Improved |
| add stratified masking (ablation) | StructRTL · masked node modeling | overall validation accuracy: 82.44% -> 82.27% overall; rare operators improve | Mixed |
| vary edge prediction sampling ratio from 10% to 40% (ablation) | StructRTL · area and delay R2 | R2: 0.7448 to 0.7463 (area) | No clear effect |
| feed textualized CDFG (.dot) to LLM instead of RTL code | VeriDistill CodeV models · area prediction, w/o KD | R2: 0.6353 -> 0.5639 (CodeV-QW-7B) | Hurt |
| fully fine-tune LLM on RTL code for quality estimation | CodeV-QW-7B · area validation, w/o KD | R2: 0.6353 -> 0.2685 | Hurt |
Closest papers by meaning
- Bridging Layout and RTL: Knowledge Distillation based Timing Prediction (ICML 2025 · similarity 0.87)
- SymRTLO: Enhancing RTL Code Optimization with LLMs and Neuron-Inspired Symbolic Reasoning (NeurIPS 2025 · similarity 0.83)
- VeriLoC: Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code (NeurIPS 2025 · similarity 0.82)
- Functional Matching of Logic Subgraphs: Beyond Structural Isomorphism (NeurIPS 2025 · similarity 0.82)
- CircuitFusion: Multimodal Circuit Representation Learning for Agile Chip Design (ICLR 2025 · similarity 0.82)
- Circuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
