Embark Studios Details Machine Learning Lessons From Arc Raiders and The Finals
The talk shows that for Embark, the hardest part of deploying machine learning in games is not the algorithm but the organizational structure around it.
Reporting from 1 sources: 4Gamer.net.
At NDC26, Embark Studios machine learning lead Martin detailed how the studio uses ML for recommendations, quest generation, and enemy behavior. A key takeaway was that a simple EASE recommendation model increased purchasers by 400% in The Finals, but organizational hurdles nearly killed the project until an engineer physically moved desks to the commercial team.
Embark Studios machine learning lead Martin gave a session at NEXON's NDC26 conference covering three practical ML applications across The Finals and the upcoming Arc Raiders. The first case study focused on The Finals store, where a simple EASE recommendation model-just four lines of Python-was tested in the For You section. The result was a 400% increase in purchasers and a significant rise in in-game currency spending.
Despite the algorithm's success, the project nearly failed due to organizational friction. The ML team was isolated from the commercial team, leading to misaligned baselines and a false initial analysis showing no significant improvement. The fix was mundane: the ML engineer moved his desk to sit with the commercial team. Communication overhead vanished, tests passed, and the model shipped.
Martin also discussed using LLMs analytically-not generatively-to validate quest graph structures, since LLMs can read every permutation of quest order without fatigue. The third example, Arc Raiders robots, was mentioned but not detailed in the report.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.