An old trick for letting AI agents learn from their failures just got an upgrade to understand plain-language instructions.
The paper, first posted to arXiv in July 2023, describes a problem with Hindsight Experience Replay, a popular method that lets reinforcement learning agents learn from failed attempts by pretending they were aiming for whatever goal they actually reached. That trick only works when goals are simple states a computer can check automatically, and it falls apart when instructions are written in natural language, since there is no built-in way to tell if a sentence like "pick up the red ball" was satisfied. The researchers' fix, called ETHER, trains two AI systems - a speaker and a listener - to invent their own simple language describing what happened in the environment, then loosely matches that invented language to human instructions using patterns in how the two co-occur. Tested on a single pickup task in the BabyAI simulated environment, ETHER's speaker and listener stood in for the missing functions and made learning more sample-efficient, even with imperfect alignment between the invented and human language.
This matters less for the specific result than for what it signals. Hindsight Experience Replay has been a reliable workhorse since 2017 for teaching agents from sparse rewards, but it has never played well with the natural-language goals that instruction-following AI increasingly relies on. ETHER is an early, narrow attempt to close that gap, borrowing emergent communication - a subfield better known for agents inventing their own protocols - to stand in for something humans usually hand-engineer.
The update itself is worth noting: this is a 2023 study getting a fresh arXiv listing in October 2026, not new research, and the only evidence so far comes from one toy pickup task in a block world - promising, but a long way from anything that follows instructions in the real world.