In this paper, we propose Financial Evolution (FEVO), a multi-stage augmentation framework for improving the performance of large-scale language models (LLMs) in finance. FEVO extends the financial domain knowledge of LLMs, injects structured inference patterns, and integrates the learned inference ability with domain knowledge through three stages: continuous pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL). For each stage, we construct a high-quality dataset FEVO-Train, and train C32B, S32B, and R32B models based on Qwen2.5-32B. Through the evaluation of seven benchmarks, FEVO-R32B achieves the previous state-of-the-art performance on five financial benchmarks, and significantly outperforms FEVO-R32B-0, which only uses RL, demonstrating the effectiveness of financial domain knowledge augmentation and structured inference.
