Let AI think in symbols humans can't understand, answering just as accurately but 12 times faster

CryptoWorld News reports that OneMillion_AI has published a paper titled “Thinking Without Words,” which proposes the abstract-cot method. By introducing 64 entirely new “abstract symbols” into the model’s vocabulary—symbols that do not correspond to any human language—the model outputs a small string of these symbols as a draft before answering questions, directly providing the answer and skipping the traditional natural language reasoning process. In the math-500 problem experiment, the number of tokens in the reasoning process was compressed from hundreds to dozens, with token usage reduced by up to 11.6 times, while maintaining the same accuracy rate. The experiment covered three model families: qwen3-8b, qwen3-4b, and ibm granite 4.0 micro, with consistent results. These 64 symbols spontaneously develop usage patterns similar to natural language during training—some symbols are reused frequently, while most appear occasionally, with a distribution curve similar to common words in human language. When the order of the symbols is shuffled, the accuracy of answering questions drops significantly, indicating that the model has learned to perform structured reasoning using this “cipher.”