VEHBench - Policy-Conditioned Selection: leaderboard

Metric: P4 Kendall tau-b (-1 to 1): rank agreement between the model ranking of a feasible candidate pool and the oracle ranking under an explicit deployment policy; LLM-assisted vibration energy harvester design against a physics oracle; one complete run per model. Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.

Top models

#ModelScore
1GPT-5.40.89
2Qwen 3.6 Plus0.88
3DeepSeek V30.86
4MiMo-V2.5-Pro (Non-reasoning)0.84
5Claude Sonnet 4.60.84
6Hy3-preview (Reasoning)0.84
7Qwen 3 Max0.84
8DeepSeek R10.83
9Gemini 3.1 Pro (Preview)0.82
10DeepSeek V4 Pro (Non-reasoning)0.79
11O4 Mini0.78
12Llama 3.3 70B0.71

Interactive version: theaggregate.ai/benchmark?slug=vehbench-policy-conditioned-selection · How It Works · Data refreshed daily, snapshot 2026-09-29.