SWEGAMEBENCH

Video games captivate billions of players worldwide. Yet, developing video games remains challenging due to specialized engines, complex infrastructure, and highly interactive execution environments. Despite advances in Large Language Model (LLM)-based automated program repair (APR), existing benchmarks and evaluations focus primarily on traditional software, leaving their effectiveness in video game repair largely unexplored. To address this gap, we introduce SWEGAMEBENCH, the first benchmark for evaluating automated repair in Unity-based game development projects. SWEGAMEBENCH includes real-world issue reports, fixing commits, executable Unity test cases, and an automated evaluation pipeline. Using this benchmark, we evaluate state-of-the-art LLM-based program repair agents on repair tasks on Unity-based projects. Live LeaderboardView the SWEGAMEBENCH leaderboard