EA's research team built a system to automate one of game development's most tedious jobs: finding ways to break the goalie AI.
The system, called Reward-Adaptive Iterative Discovery (RAID), uses reinforcement learning to train a population of virtual goal-scoring agents against a development build of NHL 26. The agents iterate until they find behavioral exploits — patterns a skilled human player could abuse to score at will. In its first deployment, RAID surfaced six distinct scoring strategies in a single experiment, strategies that human playtesters had previously needed hours of manual sessions to uncover. The key design choice was forcing diversity: standard RL agents tend to latch onto one working exploit and stop, so the researchers added a mechanism to push the population toward multiple qualitatively different solutions.
Game testing is expensive enough that studios routinely ship with known bugs simply because they ran out of time or headcount to find them all. A tool that front-loads exploit discovery — and re-runs automatically after every AI tweak — compresses that feedback loop significantly. For a title like NHL 26, where goalie behavior is central to the competitive experience, catching exploits in development rather than after launch matters.
RAID is not a finished product and the paper is careful to frame this as a case study on one game mode. The harder question is whether the approach generalizes — finding exploits in a bounded hockey sim is a cleaner problem than stress-testing the open-ended behavior of an RPG or a battle royale. EA isn't the first studio to experiment with RL for automated testing, but publishing the method suggests they'd rather establish the research claim than keep it proprietary.