Abstract: Auto-exposure （AE） is a pivotal component in imaging systems， playing a decisive role in achieving balanced image brightness and enhancing the accuracy of high-level vision tasks. However， existing techniques face considerable challenges： traditional rule-based algorithms are constrained by the “semantic gap” and struggle with semantic ambiguities in complex lighting conditions， while end-to-end deep learning approaches frequently operate as physically unconstrained “black boxes” leading to significant temporal instability. To address these issues， this paper introduces a physics-aware white-box auto-exposure framework， named PhysAEC. Departing from traditional parameter regression and image enhancement methods， we redefined the core AE challenge as a “multi-target luma prediction” task to establish optimal exposure anchors for the ISP control loop and ensure semantic adaptability and physical interpretability. PhysAEC adopts a three-stream decoupled architecture to facilitate the integration of heterogeneous information： an RGB semantic stream extracts high-level scene priors to eliminate semantic ambiguities （e.g.， in backlight scenarios）， while the raw-domain spatial grid and global histogram streams provide precise local intensity distributions and dynamic-range boundary constraints， respectively. Furthermore， to mitigate temporal oscillation during continuous inference， we introduced a tolerance-aware loss （TAL） that incorporates the hysteresis characteristics of photometric control. By optimizing physical regularization at the target level， TAL effectively suppressed parameter jitter resulting from minor fluctuations. Experiments conducted on our Balanced-AE-Dataset， comprising 10000 high-quality samples， revealed that PhysAEC achieves a prediction accuracy of 94.05% under standard conditions， with the mean absolute error decreasing from 21.12 to 2.53. In complex high-dynamic-range scenarios， the method yielded a PSNR of 38.98 dB and an SSIM of 0.994. These results underscored the proposed method’s successful integration of semantic understanding and robust physical control， establishing a new paradigm for low-level ISP control tasks.