Large Language Models (LLMs) have shown remarkable capabilities in complex reasoning tasks. However, as the number of generated tokens increases, they tend to accumulate small errors that compound over time, often leading the model further down incorrect reasoning paths. In this work, we introduce Dynamic Sampling and Multi-Validation on Scratch Policy Optimization (ASPO), a novel framework designed to enhance the reasoning robustness of LLMs. ASPO leverages scratchpads and specialized attention masks to dynamically mask previous context during inference, allowing the model to remain resilient to earlier mistakes, explore alternative reasoning paths, and identify potential inconsistencies. Extensive experiments on four benchmark datasets and across two model architectures demonstrate that ASPO significantly improves reasoning accuracy. Our findings highlight a promising direction for improving LLM performance on complex reasoning tasks.