vehbench-d82a2d0a·1 events·first seen Aliases: VEHBench
Researchers introduce VEHBench, a 763-task engineering benchmark for evaluating LLMs across distinct stages of vibration energy harvester (VEH) design, scored by an analytical physical oracle. The benchmark covers four design roles: specification triage, verifier-guided search, corrupted-state recovery, and policy-conditioned selection. Key finding is that LLM capability is strongly stage-dependent — no single model dominates the full workflow — suggesting that model routing and specialization matter for engineering applications.