Reviews: Improving Neural Program Synthesis with Inferred Execution Traces

Summary: This paper provides a novel program synthesis approach by inferring the intermediate states: execution trace. Unlike conventional methods that treat program synthesis as an end-to-end task based on input/output (i/o) examples, in this manuscript, the authors attempt to infer the execution trace from i/o examples, and further synthesize the program given the i/o and execution trace. The motivation is clear and straightforward since generating programs from execution traces should be easier. Experiments on Karel program synthesis task show the effectiveness of the proposed approach. The research question is significant and very hard.