GENSTRAT: Toward a Science of Strategic Reasoning in Large Language Models

In three linesGENSTRAT introduces a benchmark for evaluating strategic reasoning in LLMs using procedurally generated card games. Evaluation of 9 models (GPT-5, Claude, Gemini-3.1-Pro) across 36,000+ matches. Methodology decomposes competence across 6 axes and measures local volatility (jaggedness) to diagnose real-world deployments.

Summary generated by Claude — human-verified