This paper presents Input-Time Scaling, a novel scaling paradigm that complements existing large-scale language models (LLMs) scaling methods, such as data and training scale scaling and inference time scaling. This method leverages meta-knowledge to improve inputs with various strategies, and discovers a phenomenon called "train-test co-design," where strategies are applied during both training and testing. Interestingly, we find that low-quality datasets sometimes perform better, and that peak performance can be achieved with as few as 1,000 randomly selected examples. This finding contradicts the common assumption of "garbage in, garbage out." Training with more high-quality data does not always lead to improved performance, and is consistent with the "Less is More" phenomenon, where high-dimensional inference capabilities can be achieved with as few as 1,000 examples. Experimental results using the Qwen2.5-32B-Instruct model achieved state-of-the-art performance (76.7%) on AIME24 and AIME25, and combining the three models via majority vote achieved 80% performance on AIME25. Using the DeepSeek-R1-Distill-Qwen-32B model, we achieved 86.7% performance on AIME24 and 76.7% performance on AIME25. We plan to open-source the dataset, data pipeline, evaluation results, and checkpoints.
