Compelling evidence suggests that the interaction between electromagnetic metasurfaces and deep learning gives rise to the proliferation of intelligent metasurfaces in the past decade. In general, deep learning offers a transformative force to reform the design and working style of metasurfaces. A majority of the inverse-design literature announce that, given a user-defined input, the pre-trained deep learning models can quickly output the metasurface candidates with high fidelity. However, they largely ignore an important fact, that is, the practical input is always semi-known. In this work, we introduce a generation-elimination network that is robust to semi-known input and information pollution. The network is composed of a generative network to generate a number of possible answers and then a discriminative network to eliminate suboptimal answers. We benchmark the feasibility via two scenes, the on-demand metasurface design of the reflection spectra and the far-field pattern. In the microwave experiment, we fabricated and measured the reconfigurable metasurfaces to automatically meet the semi-known beam steering requirement that widely exist in wireless communication. Our work for the first time answers the question of how to cope with semi-known input, which is ubiquitous in a panoply of real-world applications, such as imaging, sensing, and communication across noisy environment.