Princeton Researchers Developed Chess-Playing AI Model

The 4-billion-parameter model achieved an expert-level Elo rating of 2697 after seven rounds of training.

Updated on Oct. 6, 2026 in Artificial Intelligence

Isometric editorial illustration featuring a geometric data center tower and a single stone knight chess piece on a plinth. — Princeton researchers have developed a 4-billion-parameter AI model named Queen, which reached an expert Elo rating of 2697 after seven training rounds. AI Illustration. Upload story photo >

Live Poll

Do you trust AI agents more when they provide clear explanations for their decisions?

Researchers at Princeton have introduced Queen, a 4-billion-parameter language model that demonstrated significant skill in chess. The model reached a 2697 Elo rating on the ChessBench scale after seven training rounds.

Why it matters

The development aims to create language models capable of explaining their decision-making processes by generalizing expert system knowledge. This architecture could eventually influence fields such as robotics and computer-use agents.

The Queen model contains 4 billion parameters and utilizes a chess-specific encoder linked to an instruction-tuned decoder. It achieved a 2697 Elo rating, gaining over 900 points from its initial 1782 baseline.

The players

Princeton

This prestigious research university located in New Jersey is home to the team that developed the Queen model.

The details

The training process employs a natural-language version of the Bellman update for self-distillation. Researchers reported no performance plateau following the seventh round of training.

The research paper describing the Queen model was posted to arXiv on October 2, 2026.

The Tech Race

This development follows the trajectory of using high-stakes strategy games to refine the reasoning capabilities of artificial intelligence. By benchmarking against the ChessBench standard, researchers are measuring how new architectures compete with established models like Gemini 3.1 Pro Preview.

While the current model focuses on chess, the underlying self-distillation technique could eventually lead to more transparent AI assistants that can explain their logic to users. These advancements represent a step toward agents that are more reliable for complex tasks like computer-use automation.

The takeaway

This project demonstrates that specialized training methods can dramatically improve the reasoning capabilities of smaller parameter-count models. The research suggests a future where AI systems provide clearer insight into the steps taken to reach a conclusion.