TriGen: A Semantic-Feedback Collaborative LLM Test Case Generation Model

Abstract

Regarding the problems of semantic understanding bias and uncontrollable generation in the generation of test cases driven by natural language requirements, this paper proposes TriGen - a controllable and traceable test case generation model for Chinese requirements. To break through the limitations of end-to-end opaque generation, TriGen is based on DeepSeek-7B and adopts a modular decoupling architecture, dividing the generation process into five stages: semantic enhancement, test type mapping, test procedure generation, step-by-step feedback optimization, and structured output. It realizes the explicit modeling of the generation logic and process tracing. Through efficient parameter fine-tuning of DeepSeek-7B using LoRA, the aim is to adapt to the multi-stage generation mechanism of TriGen, clarify the semantic goals and output formats of each stage, and achieve model specialization adaptation in low-resource environments. Further, a multi-dimensional closed-loop feedback mechanism is introduced to support local correction and quality iteration. To support research in Chinese scenarios, a high-quality Chinese software requirement dataset containing 2120 labeled samples is constructed, covering scenarios such as functional descriptions and interaction logic. Experimental results show that TriGen significantly outperforms the baseline method in both automatic evaluation and human evaluation. The semantic-level precision, recall, and F1 score reach 0.825, 0.819, and 0.801 respectively, with an hallucination rate controlled at 0.183, and the human rating is 0.94. It provides an interpretable, iterative, and deployable technical path for the automatic generation of test cases in Chinese requirements.