Research map / Large language models / Agents, code and math
QiMeng-ChipV-RTL: Exploiting Information Locality for IP-level Verilog Generation
Hanqi Lyu, Di Huang, Yaoyu Zhu, Kangcheng Liu, Bohan Dou, Chongxiao Li, Pengwei Jin, Shuyao Cheng, Rui Zhang, Zidong Du, Qi Guo, Xing Hu and 1 more
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 |
|---|---|---|---|
| index document into fragments with dual-level descriptions instead of feeding full specification (ablation) | ChipV-RTL (Claude) · RealBench (all) | functional pass rate: 35.0% -> 45.0% | Improved |
| add AST-guided locality-aware debugging (on top of no-index setup) (ablation) | ChipV-RTL (Claude) · RealBench (all) | functional pass rate: 20.0% -> 35.0% | Improved |
| add planner agent (on top of no index, no debug) (ablation) | ChipV-RTL (Claude) · RealBench (all) | functional pass rate: 20.0% -> 20.0% functional; syntax 43.3% -> 51.6% | No clear effect |
| add planner agent (on top of no index, no debug) (ablation) | ChipV-RTL (Claude) · RealBench (all) | syntactic pass rate: 43.3% -> 51.6% | Improved |
| remove indexing on AES IP (ablation) | ChipV-RTL (Claude) · RealBench AES | functional pass rate: 50.0% -> 50.0% | No clear effect |
| use ChipV-RTL multi-agent framework with Claude vs MAGE with Claude | Claude-3.7-Sonnet · RealBench (all) | functional pass rate: 21.6% -> 45.0% | Improved |
| use ChipV-RTL with Claude vs direct Claude prompting | Claude-3.7-Sonnet · CVDP cid003 | functional Pass@1: 48.7% -> 61.5% | Improved |
| swap backbone from Claude to DeepSeek-V3 in ChipV-RTL | ChipV-RTL · RealBench (all) | functional pass rate: 45.0% -> 35.0% | Hurt |
| use information locality to restrict context per agent | ChipV-RTL agents · RealBench | context reduction: 70.8% (RTL), 54.8% (Merger), 61.6% (Debugger) | Improved |
| ChipV-RTL generated RTL vs golden RTL on PPA | ChipV-RTL · RealBench synthesisable modules | area/delay/power: 2.37% Area, 4.34% Delay, 4.63% Power improvement (text); table average 0.25% / 3.00% / -1.15% | Mixed |
| use hardware spec instead of software spec (locality measure) | Qwen3-8B relevance scoring · normalized entropy | normalized entropy: 0.8220 -> 0.6718 | Improved |
Closest papers by meaning
- VeriThoughts: Enabling Automated Verilog Code Generation using Reasoning and Formal Verification (NeurIPS 2025 · similarity 0.88)
- SymRTLO: Enhancing RTL Code Optimization with LLMs and Neuron-Inspired Symbolic Reasoning (NeurIPS 2025 · similarity 0.86)
- DeepRTL: Bridging Verilog Understanding and Generation with a Unified Representation Model (ICLR 2025 · similarity 0.86)
- BetterV: Controlled Verilog Generation with Discriminative Guidance (ICML 2024 · similarity 0.86)
- CraftRTL: High-quality Synthetic Data Generation for Verilog Code Models with Correct-by-Construction Non-Textual Representations and Targeted Code Repair (ICLR 2025 · similarity 0.86)
- VerilogCoder: Autonomous Verilog Coding Agents with Graph-based Planning and Abstract Syntax Tree (AST)-based Waveform Tracing Tool (AAAI 2025 · similarity 0.86)
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.
