SEER: Self-Enhancing Chain-of-Thought Compression for Reasoning Models
SEER is a framework that compresses long Chain-of-Thought (CoT) reasoning traces by learning concise patterns from self-generated outputs. It uses Best-of-N sampling to select shorter correct paths and an adaptive filter to control length based on data distribution. This results in an average 41.6% reduction in CoT length while maintaining or improving performance across software engineering tasks.