Sensor-based computing minimizes latency and energy consumption by processing data at the capture point, thereby eliminating extensive data transfer and enabling real-time decision-making. Here, we present a breakthrough in in-sensor computing via circularly polarized light detectors that integrate cholesteric liquid crystal reflectors with two-dimensional van der Waals p-n heterostructures. Our device exhibits a high dissymmetry factor (1.90), allowing effective separation of mixed circularly polarized images, along with a rapid photoresponse (4 μs) and wide linear dynamic range (up to 114.1 dB), suitable for analog multiply-and-accumulate operations in convolution-based in-sensor computing. Harnessing these detectors, we propose mixed-multimodal in-sensor computing using the chiral state of circularly polarized light to dynamically control responsivity, which enables the blending of two arbitrary image processing modes within a single, non-reconfigurable circuit. By effectively integrating polarization-sensitive detectors into the in-sensor computing framework, the proposed architecture preserves kernel optimization capabilities while simplifying circuit complexity.
© 2025. The Author(s).