VEHBench - Corrupted-State Recovery: leaderboard

Metric: P3 success (%; share of corrupted or misleading design trajectories that end feasible within the recovery budget; LLM-assisted vibration energy harvester design against a physics oracle; one complete run per model). Source: arxiv.org. Saturation forecast: Around January 2028. 12 models tracked.

Top models

#ModelScore
1Hy3-preview (Reasoning)47.4
2MiMo-V2.5-Pro (Non-reasoning)45.5
3DeepSeek R142.9
4GPT-5.442.3
5Gemini 3.1 Pro (Preview)37.2
6DeepSeek V335.3
7Qwen 3 Max30.1
8DeepSeek V4 Pro (Non-reasoning)28.8
9Qwen 3.6 Plus27.6
10O4 Mini26.3
11Claude Sonnet 4.616
12Llama 3.3 70B3.2

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