TianyuCodings/NanoJev

What it solves

NanoJev is a small-scale replica of the Jev model, designed to handle decision-making tasks without the need for traditional token-by-token decoding. It allows a model to take in states and questions and directly output complete probability distributions over a set of candidates, enabling faster and more structured decision-making in environments like mazes and the game of Snake.

How it works

Built on a 0.6B parameter Qwen3 backbone, NanoJev uses specialized decision heads instead of a standard language model head. It processes multiple states and questions in a single forward pass, allowing it to batch independent decisions. It supports three types of decisions:

• Dynamic Choice: Provides probabilities for 2 to 255 candidates.
• Boolean decisions: Returns the probability that a proposition is true.
• Ordered Score: Returns a probability-weighted expectation across 2 to 10 levels.

Who it’s for

This project is for researchers and developers interested in "System One" models, high-efficiency decision models, and the integration of local judgments with code-based planning in AI agents.

Highlights

• Zero output-token decoding: Decisions are read directly from a forward pass, removing the overhead of autoregressive generation.
• 0.6B LLM backbone: Efficiently leverages Qwen3-0.6B for structured outputs.
• Parallel decision processing: Can handle multiple queries (e.g., 6 states and 18 questions) in one forward pass.
• Game-based evaluation: Demonstrated success in 50x50 mazes and Snake, often outperforming untuned base models.
• Calibrated decisions: Includes implementations for paired proper-reward learning and Brier score training to improve probability quality.