Chain-of-thought prompting improves complex reasoning, yet the resulting long intermediate traces inflate inference cost and impede deployment in financial contexts. To address this, we introduce the Hierarchical Semantic Distillation Network (HSDN), which compresses reasoning chains while preserving answer accuracy and logical coherence.
The framework proceeds through several stages: semantic segmentation splits the full chain into coherent segments; a directed dependency graph is built over these segments; a dual encoder assigns importance scores to each segment; constrained segment selection respects a target compression ratio; and local boundary rewriting restores context for the trimmed portions. Feature extraction and final answer generation rely solely on a frozen Qwen3‑4B model, keeping the compression pipeline structured and interpretable.
On the AFAC2025 benchmark, HSDN attains 91.0% accuracy with a 68.4% reduction in chain length, outperforming strong baselines in overall score and reasoning coherence. These results demonstrate that graph‑guided compression is highly effective for high‑stakes financial reasoning tasks.
Review: The graph‑driven compression approach achieves substantial cost savings without sacrificing reasoning quality, offering a practical pathway for deploying large language models in finance.