Abstract: In recent years, large language models (LLMs) have demonstrated broad prospects in the judicial field. However, in knowledge-intensive reasoning and complex logical judgment tasks within judicial question-answering scenarios, challenges such as inadequate reasoning capabilities and imprecise application of legal knowledge persist. To address these issues, this paper proposes a decoupled collaborative reasoning framework (DCRF) that separates “thinking” from “reasoning” in a multi-stage coope- rative process. First, a fine-tuned lightweight “Thinker” generates high-level chains to guide downstream reasoning strategies. Then, an unmodified Qwen1.5-14B-chat “Reasoner”, supported by retrieval-augmented generation and relevant statutory texts, performs fine-grained logical inference. By coordi- nating strategic planning with execution, this framework significantly enhances the model’s flexibility and accuracy in invoking legal knowledge, while avoiding the high costs of fine-tuning large models and reducing overall training overhead. On the JEC-QA and DISC-Law-Eval benchmarks, DCRF achieves an average improvement of 9.77 percentage points in accuracy for single-choice questions and an average increase of 7.48 percentage points in F1-score for multiple-choice questions compared to the base models. Notably, it surpasses DeepSeek-R1-Distill-Qwen-14B in single-choice questions and performs comparably in multiple-choice questions. Experimental results indicate that DCRF effectively strengthens the judicial reasoning capabilities of large language models while reducing training costs.

Key words: multi-stage reasoning, large language models, legal reasoning, retrieval-augmented gene- ration, instruction tuning