Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models
Published in Preprint; under submission to NeurIPS 2026, 2026
This manuscript is a preprint currently under submission to NeurIPS 2026. We study dynamic boundary evaluation for language models: a framework for probing what models can and cannot do beyond fixed benchmarks and static worst-case attacks. The work formulates evaluation as an adaptive process that searches along model capability boundaries, guided by skills and calibrated by task difficulty.
The corresponding Chinese thesis, 大语言模型动态能力边界评测：技能引导搜索与难度校准, was selected as an Outstanding Undergraduate Thesis.
Recommended citation: Wang, H., Yu, D., & Zhang, H. (2026). "Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models." Preprint; under submission.