Benchmarking Small LLMs and Jev on Real-Time ViZDoom Control

A developer benchmarked several small models—including TypeSafe's Jev API, Laya, fine-tuned ModernCE-base-nli, and fine-tuned Qwen3.5-4B—to control a player in ViZDoom in real-time using text representations of the game state. The test evaluated performance across two Doom scenarios ('Defend the Center' and 'Health Gathering') targeting five decisions per second while the game clock continued running. The experiment demonstrates the feasibility and latency trade-offs of using language and structured decision models for real-time game and robotic control loops. It highlights how small generative models like Qwen3.5-4B can achieve higher gameplay performance at the cost of execution speed, while specialized decision engines provide ultra-low latency. Qwen3.5-4B (LoRA) achieved the highest average kills (3.63) in 'Defend the Center' with a median latency of 146.8 ms, whereas ModernCE-base-nli achieved a median latency of 7.6 ms with 1.25 mean kills. Local models ran on an NVIDIA GB10 DGX Spark system using a deterministic Python adapter that converted ViZDoom object labels, bounding boxes, and HUD data into text inputs.