Research
VEHBench: Stage-Local Diagnostics Expose Where LLMs Fail in Physical Design Loops
Depeng Su, Yuyu Luo and Guobiao Hu introduce VEHBench (arXiv 2607.18181, cs.CL/cs.SE), a benchmark for LLM-assisted vibration energy harvester design for battery-free IoT. The design is the interesting part: rather than scoring only the end artifact, it diagnoses per-stage failures under coupled physical constraints. That stage-local evaluation pattern generalizes to any agent benchmark where a single end-to-end score hides which step broke.
Source
↳ Follow the thread