New Method Compresses Chain-of-Thought for Financial AI

Chain-of-Thought (CoT) prompting is a powerful technique for enhancing the reasoning capabilities of large language models, particularly for complex tasks. However, the extended intermediate reasoning traces generated by CoT can lead to substantial inference costs, posing a significant barrier to its practical deployment, especially in high-stakes environments like financial analysis. To address this, researchers have introduced the Hierarchical Semantic Distillation Network (HSDN), a novel framework designed to compress these reasoning chains. HSDN aims to drastically reduce the length of the traces while meticulously preserving both the accuracy of the final answer and the logical coherence of the reasoning process. The framework operates through several integrated steps: semantic segmentation, construction of dependency graphs, importance scoring using a dual encoder, constrained segment selection, and local boundary rewriting. By using a frozen Qwen3 4B model solely for feature extraction and final answer generation, the compression process remains structured and interpretable. On the AFAC2025 benchmark, HSDN achieved 91.0% accuracy with an impressive 68.4% compression, outperforming existing compression baselines in both overall score and reasoning coherence, demonstrating the effectiveness of graph-guided compression for critical financial reasoning.