RTL Generation on RTLLM common subset, N=44 and N=49 v2.0
0.48CoverageTTT-RTL
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| TTT-RTLBackbone LLM=Qwen3-8B, Training Strategy=domain SFT plus per-design test-time training, Evaluation Flow=Yosys / OpenSTA / Nangate 45 nm2026.06 | 0.48 | 0.349 | 0.527 | 0.38 | 33.5 | 0.341 | |
| EvolVEBackbone LLM=GPT-4o-mini, Training Strategy=frozen agentic search, Evaluation Flow=Yosys / OpenSTA / Nangate 45 nm2026.06 | 0.46 | 0.872 | 0.909 | 0.17 | 0.5 | 0.892 | |
| REvolutionBackbone LLM=GPT-4o-mini, Training Strategy=frozen agentic search, Evaluation Flow=Yosys / OpenSTA / Nangate 45 nm2026.06 | 0.45 | 0.739 | 0.813 | 0.29 | 8 | 0.762 | |
| VeriAgentBackbone LLM=GPT-4o-mini, Training Strategy=frozen agentic search, Evaluation Flow=Yosys / OpenSTA / Nangate 45 nm2026.06 | 0.44 | 0.813 | 0.853 | 0.25 | 2 | 0.83 |