TianyuCodings/NanoJev

What it solves

NanoJev addresses the challenge of making real-time, high-performance decisions in interactive environments (like games) without the latency and computational overhead of traditional LLM text generation. Instead of generating tokens, it treats decision-making as a classification or scoring problem, outputting probability distributions directly.

How it works

It uses a small 0.6B parameter backbone (Qwen3) equipped with specialized decision heads. When provided with a state, a question, and a set of candidate actions, the model performs a single forward pass to produce probabilities for those candidates. This allows it to rank, select, or sample actions instantly. It is trained on a mixed-task dataset covering four different game environments, allowing a single model checkpoint to handle diverse tasks like navigation and target tracking.

Who it’s for

This project is for researchers and developers interested in "System 1" AI models—systems that make rapid, intuitive decisions rather than slow, deliberative reasoning. It is particularly useful for those building AI agents for games or real-time simulation environments where low-latency action selection is critical.

Highlights

• Zero-token decoding: Eliminates the need for autoregressive text generation, resulting in near-instant decision-making.
• Unified architecture: A single model checkpoint handles four distinct tasks (ViZDoom Basic, Predict Position, Maze, and Snake).
• High performance: Demonstrates superior success rates in game benchmarks compared to baseline models like Jev and untuned Qwen.
• Interactive evaluation: Includes a web-based interface for visualizing side-by-side performance comparisons between models in real-time.