Open-Ended Video Game Glitch Detection with Agentic Reasoning and Temporal Grounding

A research team led by Muyang Zheng has developed GliDe, a novel agentic framework designed to tackle the complex challenge of open-ended video game glitch detection. Unlike traditional image recognition tasks, this requires understanding game mechanics, physics, and expected state transitions over continuous video to distinguish true bugs from unusual but valid gameplay. To support this, the team created VideoGlitchBench, the first benchmark of its kind, containing 5,238 annotated gameplay clips from 120 different games, each with detailed glitch descriptions and precise timestamps.

GliDe's architecture employs three key components for robust detection. First, a game-aware contextual memory provides the AI with necessary background knowledge for informed reasoning. Second, a debate-based reflector uses multi-agent discussion to verify potential glitches from multiple perspectives, reducing false positives. Finally, an event-level grounding module pieces together fragmented evidence to recover complete glitch intervals. Experiments show current general multimodal models struggle significantly with this task, but GliDe achieves substantially stronger performance by jointly optimizing for semantic fidelity and temporal accuracy in its evaluations.