- Abstract: Large-scale benchmark datasets have been among the major driving forces in AI, supporting training of models and measuring their progress. The key assumption is that these benchmarks are realistic approximations of the target tasks in the real world. However, while machine performance on these benchmarks advances rapidly --- often surpassing human performance --- it still struggles on the target tasks in the wild. This raises an important question: whether the surreal high performance on existing benchmarks are inflated due to spurious biases in them, and if so, how we can effectively revise these benchmarks to better simulate more realistic problem distributions in the real world. In this paper, we posit that while the real world problems consist of a great deal of long-tail problems, existing benchmarks are overly populated with a great deal of similar (thus non-tail) problems, which in turn, leads to a major overestimation of true AI performance. To address this challenge, we present a novel framework of Adversarial Filters to investigate model-based reduction of dataset biases. We discuss that the optimum bias reduction via AFOptimum is intractable, thus propose AFLite, an iterative greedy algorithm that adversarially filters out data points to identify a reduced dataset with more realistic problem distributions and considerably less spurious biases. AFLite is lightweight and can in principle be applied to any task and dataset. We apply it to popular benchmarks that are practically solved --- ImageNet and Natural Language Inference (SNLI, MNLI, QNLI) --- and present filtered counterparts as new challenge datasets where the model performance drops considerably (e.g., from 84% to 24% for ImageNet and from 92% to 62% for SNLI), while human performance remains high. An extensive suite of analysis demonstrates that AFLite effectively reduces measurable dataset biases in both the synthetic and real datasets. Finally, we introduce new measures of dataset biases based on K-nearest-neighbors to help guide future research on dataset developments and bias reduction.