Zugzwang: Pushing the Limits of General Large Language Models in Chess Using Pure Prompt Engineering Techniques

Zugzwang is a reproducible research platform that explores the capability boundaries of general large language models (LLMs) in chess tasks without fine-tuning, using techniques such as pure prompt engineering, RAG, chain-of-thought, and multi-agent orchestration.

• Original Author/Maintainer: maelrx
• Source Platform: GitHub
• Original Title: Zugzwang
• Original Link: https://github.com/maelrx/Zugzwang
• Source Publication/Update Time: 2026-05-31T08:44:42Z

"Zugzwang" (German for "forced move") is a chess term describing a situation where it's a player's turn to move, but any legal move will worsen their position. This situation perfectly metaphorizes the dilemma of large language models in complex reasoning tasks—they possess rich knowledge but often struggle to make optimal decisions under specific constraints.

Choosing chess as the test platform is no accident. This game has clear rules, verifiable results, and rich tactical strategies, making it an ideal "microscope" for evaluating AI reasoning capabilities. More importantly, chess is complex enough to challenge the model's planning and decision-making abilities, yet not as difficult to evaluate as open-domain tasks.

The Zugzwang project builds on the LLM Chess benchmark study published by Saplin et al. in 2025. This study revealed several key findings:

• Most LLMs cannot even beat a random-move opponent; the problem is not a lack of chess knowledge but an inability to follow instructions correctly
• Only reasoning-enhanced models (e.g., o3, o4-mini, Grok 3 Mini) can reliably beat random opponents
• The best-performing model (o3 low) achieves an Elo rating of approximately 758 against a calibrated engine—slightly above the average level of regular players on chess.com
• Providing move history can significantly reduce errors (o4-mini's error rate dropped from 11.2% to 1.6%)
• Mixture-of-Agents models, which combine strong reasoning and strong instruction-following abilities, can double the win rate and achieve a 100% game completion rate

However, the LLM Chess benchmark uses simple general prompts without few-shot examples, retrieval-augmented generation (RAG), structured chain-of-thought, or feedback-rich retry mechanisms. Zugzwang is designed precisely to fill these gaps.

The project's core research question is concise yet profound:

Using only LLM manipulation techniques—system prompts, RAG, few-shot learning, chain-of-thought, tool use, multi-agent orchestration—without fine-tuning any model, to what extent can a general large language model be pushed in chess?

This question has important methodological significance. It attempts to distinguish between two types of capabilities: the model's inherent "raw ability" and the potential capabilities that can be "unlocked" through carefully designed prompts and system architectures.

Zugzwang adopts a modular seven-layer architecture, where each layer can be tested independently:

Responsible for basic functions such as configuration loading, key management, and environment validation to ensure the reproducibility of experiments.

Contains components like BoardManager, game loop, LLM/random/engine players, serving as the backbone of the entire system.