These papers really changed what I thought was possible in Computer Science today. I hope they are interesting to you too.
The fact that the computer vision problem of recognising images is now solved (i.e. computers are about as good a humans at this task) was a huge surprise to me, but this paper: Generative Adversarial Text to Image Synthesis demonstrates a program that can create a picture of a bird entirely from a text description of it, such as 'this magnificent fellow is almost all black with a red crest, and white cheek patch'
I quite like the concept of structuring programming problems so there is a separatation between the domain model and the 'search' or optimisation method used to process it. This is a level of abstraction beyond normal functional (or OO) design, where we use the full power of a computer to manipulate the 'program'. Microsoft's paper Automated Synthesis of Symbolic Instruction Encodings from I/O Samples (MSR-TR-2011-123) shows an example of how to combine search and SMT solvers to automatically reverse engineer the behaviour of 500 x86 instructions.