Visual in-sensor computing (ISC) provides a promising route for reducing redundant data transfer and enabling low-power visual intelligence at the sensor front end. Two-dimensional (2D) material-enabled optoelectronic neuromorphic devices are well suited for this paradigm because their light-matter interaction, tunable interfaces, defect states, ionic dynamics, and anisotropic responses can couple optical sensing with conductance modulation and memory. However, the field is still dominated by diverse material systems, isolated device demonstrations, and application-specific reports. A clear framework is needed to connect material properties, photoinduced mechanisms, device architectures, array integration, and visual ISC functions. In this review, recent progress in 2D optoelectronic neuromorphic devices is reorganized from this cross-scale perspective. The discussion highlights how 2D and quasi-2D material platforms enable light-induced synaptic plasticity, how different device architectures translate these mechanisms into programmable conductance states, and how array-level systems support visual preprocessing, dynamic perception, multidimensional encoding, and pattern recognition. Particular attention is given to the transition from single-device proof-of-concept studies to array-level and task-level visual ISC. Key challenges are further analyzed, including material nonuniformity, device drift, array variability, readout noise, limited benchmarking, and insufficient real-scene validation. This review aims to clarify the development path of 2D optoelectronic neuromorphic hardware toward scalable, reliable, and energy-efficient visual ISC systems.

Information is acquired by humans in dynamically changing environments, with vision acting as the pathway for 80% of information intake. As the core of the visual system, the adaptive mechanism enables humans to flexibly respond to complex and variable lighting conditions. In recent years, biological adaptive mechanisms such as light/dark and chromatic adaptation have been the focus of neuromorphic devices inspired by the visual pathway, intending to overcome static perception limitations and build artificial vision systems capable of dynamic environmental adaptation. This paper systematically reviews the material systems and biomimetic mechanisms of visual adaptive neuromorphic devices. The optoelectronic and memory performances of devices are discussed to provide references for achieving efficient adaptive functions. In addition, application scenarios of artificial vision based on visual adaptive neuromorphic devices are summarized, and their prospects and challenges are deeply explored to offer insights for developing high-performance biomimetic vision-adaptive devices.

The development of high-order neuromorphic computing requires device that integrates in-sensor visual sensing-memory-processing. However, integrated of volatile and non-volatile behavior, as well as reconfigurable architecture for antagonistic photoresponse—excitation and inhibition in single device under single-wavelength stimulus remains critical bottleneck in designing all-in-one neuromorphic visual system. Herein, an amorphous-GaO_x/Hf_0.5Zr_0.5O₂ (a-GaO_x/HZO) heterojunction device demonstrates dual-mode functionality switchover between sensing module (SM) and non-volatile module (NVM) by merely adjusting single-wavelength light intensity. The merit parameters of the SM are governed by switchable ferroelectric polarization, forming sufficient foundations for optoelectronic logic gates. The reconfigurable photoresponse—light intensity-dependent excitation and inhibition (i.e., Weber's Law) of the NVM endows the framework with visual self-adaptation, namely photopic and scotopic adaptation. Representative self-adaptation photosensitivity and adaptive index are strongly correlated with switchable ferroelectric polarization, thereby boosting responsiveness and self-adaptability. Leveraging the dual-mode switchover mechanism, the monolithic a-GaO_x/HZO heterojunction units integrated SM with NVM serve as sensing and computing building blocks for designing in-sensor processing: the SM with distinguishable photoresponse for pre-filtering of interference information and reconfigurable conductance of the NVM for performing anti-interference transmission of digits. This work provides a programmable framework for designing multimodal integrated neuromorphic vision chips and establishing brain-like sensory system for anti-interference communication.