GDC Festival of Gaming is part of the Informa Festivals Division of Informa PLC

This site is operated by a business or businesses owned by Informa PLC and all copyright resides with them. Informa PLC's registered office is 5 Howick Place, London SW1P 1WG. Registered in England and Wales. Number 8860726.

March 9-13, 2026
Moscone CenterSan Francisco, CA

Agenda

Generative Recommender Post-Trained by Reinforcement Learning in Video Games

Dong Fang  (Lead Data Scientist, LIGHTSPEED STUDIOS)
Location: Room 3004, West Hall
Date: Monday, March 9
Time: 11:50 am - 12:20 pm
Pass Type: Festival Pass, Game Changer Pass - Get your pass now!
Audience Level: Advanced
Track: Game & Production Technology
Format: Lecture
Vault Recording: Video
Audience Level: Advanced

The exponential growth of online gaming generates vast amounts data but also presents unique challenges for recommendation systems—including fluctuating player interests, complex social graphs, and diverse content types. This lecture introduces a novel generative recommendation framework built on large language models and enhanced by reinforcement learning (RL). Designed specifically for gaming ecosystems, it unifies multiple recommendation tasks - such as content, product, social suggestions - into a single architecture, significantly reducing engineering overhead while enabling rapid feature deployment. The RL post-training enables near real-time adaptation to evolving player behavior, optimizing long-term engagement and experience diversity. The solution not only supports dynamic configuration optimization within gameplay but also enhances content delivery across game communities. It demonstrates substantial improvements in retention and engagement compared to current online solutions. Finally, the talk will address scalability, latency, and fairness challenges encountered in real-world deployment.

Takeaway

Attendees will learn how to design a unified, generative recommender system using large language models and reinforcement learning (RL), specifically tailored for gaming environments. They will explore concrete strategies for handling diverse recommendation scenarios - such as game configuration, content feed recommendation, and social suggestions - within this unified architecture. The system dynamically adapts via RL to optimize key metrics like retention and engagement, while also addressing production challenges such as latency and bias at scale.

Intended Audience

This session targets platform architects, ML engineers, data scientists, and product leads building recommendation systems for online games. Foundational knowledge of machine learning concepts is recommended. Game designers seeking deeper insights into advanced personalization mechanics (friend systems, UGC discovery, dynamic advertising) will also benefit significantly.