Task design optimization using adaptive task difficulty adjustment has been used to improve behavioral performance [1]; however, its theory-free approach lacks feasibility for personalized cognitive enhancement. We aim to provide a computational framework for Cognitive Bias (CB) training with the goal of reducing cognitive bias (Fig.1A), particularly in memory and inference, that designs individually tailored task based on a metacognitive monitoring and control theory [2]. Our cognitive training strategy is inspired by the successes of neural network training through loss minimization. This framework comprises four layers: stimuli, behavior, latent cognition, and a bias in model parameter layer (Fig.1B). The first layer is built on a task parameter set from a recall-based two-step decision task (Fig. 2A). The other layers correspond to action-and-confidence measures, computational variables (e.g., prediction error) of a computational model of decision-making (Fig. 2B), and the distance between the model parameters of a target and a human trainee (model parameter distance; MPD) (Fig. 2C), which we defined as the loss for CB training. We generated the task stimulus sequences for the training of each subject in silico to minimize this MPD. We obtained the distribution of model parameters from the first experiment (N=97) and defined a target model based on high-performing participants (Fig.2C). This was followed by a second experiment (N=36) that examined the effect of CB training. Beyond the effects of repetition learning shown in the control group, the CB-trained group exhibited significantly improved decision-making performance, including enhanced metacognitive efficiency. The effects of CB training were attributed to the reoriented MPD towards the target model (Fig.3). Our work provides a theoretical foundation for theory-based automatic task generation that potentially reshapes human cognitive processes for each individual, bridging computational cognitive psychology and machine learning.