The paper introduces a set of synthetic tasks designed to evaluate reading comprehension and question answering (QA) abilities, aiming to measure progress towards building intelligent dialogue agents. The tasks are intended to be prerequisites for any system capable of conversing with humans, testing various aspects of understanding, such as chaining facts, simple induction, deduction, and more. The authors argue that many existing learning systems cannot solve these tasks, and they classify these tasks into skill sets to help researchers identify and rectify system weaknesses. They also extend the Memory Networks model and show that it can solve some but not all of the tasks. The tasks are generated using a simulation of a physical world, similar to a classic text adventure game, where actors move around and interact with objects. The simulation generates grounded text and question-answer pairs, allowing for clear evaluation of performance. The paper includes a detailed description of the tasks, a discussion of related work, and experimental results comparing different methods on the tasks. The authors conclude that while some existing machine learning methods, particularly Memory Networks, perform well on some tasks, they still fail on several others, highlighting the need for further research to minimize the required supervision and training examples.The paper introduces a set of synthetic tasks designed to evaluate reading comprehension and question answering (QA) abilities, aiming to measure progress towards building intelligent dialogue agents. The tasks are intended to be prerequisites for any system capable of conversing with humans, testing various aspects of understanding, such as chaining facts, simple induction, deduction, and more. The authors argue that many existing learning systems cannot solve these tasks, and they classify these tasks into skill sets to help researchers identify and rectify system weaknesses. They also extend the Memory Networks model and show that it can solve some but not all of the tasks. The tasks are generated using a simulation of a physical world, similar to a classic text adventure game, where actors move around and interact with objects. The simulation generates grounded text and question-answer pairs, allowing for clear evaluation of performance. The paper includes a detailed description of the tasks, a discussion of related work, and experimental results comparing different methods on the tasks. The authors conclude that while some existing machine learning methods, particularly Memory Networks, perform well on some tasks, they still fail on several others, highlighting the need for further research to minimize the required supervision and training examples.