VEHBench - Design Search: leaderboard
Metric: P2 final feasible power ratio (ratio, 0 and up, 1 = best known feasible design): mean over tasks of final feasibility times output power relative to the best known feasible power, infeasible final designs scoring zero, after oracle-feedback-guided bounded edits; LLM-assisted vibration energy harvester design against a physics oracle; one complete run per model. Source: arxiv.org. Saturation forecast: Around July 2027. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.39 |
| 2 | DeepSeek R1 | 0.24 |
| 3 | Claude Sonnet 4.6 | 0.24 |
| 4 | Qwen 3.6 Plus | 0.21 |
| 5 | Qwen 3 Max | 0.2 |
| 6 | Hy3-preview (Reasoning) | 0.16 |
| 7 | DeepSeek V3 | 0.16 |
| 8 | O4 Mini | 0.16 |
| 9 | GPT-5.4 | 0.13 |
| 10 | DeepSeek V4 Pro (Non-reasoning) | 0.13 |
| 11 | Llama 3.3 70B | 0.12 |
| 12 | MiMo-V2.5-Pro (Non-reasoning) | 0.11 |
Interactive version: theaggregate.ai/benchmark?slug=vehbench-design-search · How It Works · Data refreshed daily, snapshot 2026-09-29.