A new AI model, detailed in the paper "Thinking Without Words," introduces Abstract-CoT, a method that employs 64 unique abstract symbols to improve mathematical problem-solving efficiency. These symbols, unrelated to any human language, allow the model to perform preliminary reasoning before providing answers, significantly reducing the token count needed for reasoning on the MATH-500 benchmark by up to 11.6 times without sacrificing accuracy. The approach also enhances performance on AlpacaEval tests and is effective across multiple model families, including Qwen3-8B, Qwen3-4B, and IBM Granite 4.0 Micro. The symbols developed usage patterns similar to natural language, with frequent reuse of a few symbols, indicating structured reasoning rather than random output.