Innovation and Evaluation of AI-Driven Economics Experimental Course Teaching Model

To address the shortcomings of current outcome-oriented static evaluation mechanisms in economics experimental teaching, which are lagging behind and unable to accurately measure the evolution of students’ higher-order economic decision-making thinking, this paper proposes an AI (Artificial Intelligence)-driven innovation and dynamic evaluation system for economics experimental teaching, integrating a large language model and a time-series tracking algorithm. First, in terms of teaching model innovation, an LLM (Large Language Model) agent with dynamic risk preferences and memory mechanisms is introduced as a virtual market participant to reconstruct a highly realistic and adaptive economic game scenario. Second, a seamless multimodal data acquisition engine is constructed, which, in addition to extracting clickstream and latency data, utilizes an NLP (Natural Language Processing) model to deeply analyze the textual sentiment of negotiation dialogues; combined with economic theory, the fragmented interaction logs are transformed into quantitative economic features representing the degree of risk probing and information processing resistance. Finally, in terms of effect evaluation and feedback design, a deep tracking model based on the attention mechanism is constructed to accurately map the feature matrix to students’ economic cognitive state, and then calculate a dynamic decision rationality index. Simultaneously, by focusing on the extraction of key price adjustment decision nodes using the model’s attention weights, high theoretical interpretability of the process evaluation is achieved, and this is used as a trigger condition to provide students with LLM-driven real-time personalized intervention feedback. The deployment results of this system in the course “Two-Way Auctions and Oligopoly Markets” showed that the decision rationality index of the experimental group students significantly increased by 0.278 compared to the traditional group, and the dynamic comprehensive evaluation score output by the system highly matched the students’ final professional ability performance (Pearson correlation coefficient r=0.88). This study effectively addresses the interpretability concerns of the traditional teaching evaluation’s ‘algorithm black box’ by establishing a differentiable mapping between model attention weights and specific economic decision-making indicators, thereby providing a new path for the reconstruction of the economics experimental teaching paradigm in the digital age.