Trend-Aware Predictive Generative Inference Caching: The Second Experimental Study on GPT–Assisted Discovery of Frequently Asked Generative AI Questions and Predictive Inference Cost Optimization

The increasing adoption of large language models (LLMs) has created substantial computational demands associated with repeated information requests. Although conventional caching mechanisms can reduce redundant inference, their effectiveness generally depends on the availability of previously generated responses or accumulated request histories. An important research opportunity is to anticipate future information demands and prepare reusable generative responses before corresponding user requests consume additional computational resources. This study presents the second experimental investigation of Trend-Aware Predictive Generative Inference Caching (TPGIC), introducing a GPT-6 mini–assisted workload construction methodology and evaluating its computational advantages over conventional inference and reactive caching strategies. A central methodological contribution is the integration of LLM-assisted question identification into predictive generative inference workload construction. GPT-6 mini is employed to identify and construct a collection of frequently asked, self-contained generative AI questions across everyday knowledge, education, programming, artificial intelligence, health-related information, professional communication, and personal development. A second GPT-6 mini elicitation procedure estimates the potential monthly global demand associated with each question, providing an exploratory workload-prior distribution. Importantly, the proposed TPGIC framework is not restricted to GPT-based question identification. Its future-demand prediction architecture can incorporate multiple complementary information sources and prediction mechanisms, including LLM-based candidate question generation, search-engine query suggestions and autocomplete patterns, emerging news and technological announcements, social media discussions and engagement dynamics, web search trends, public forum discussions, historical generative AI request logs, anonymized user interaction sequences, temporal demand patterns, seasonal and event-driven forecasting, statistical time-series models, machine-learning-based demand prediction, topic modeling, semantic clustering, and hybrid predictive models combining multiple signals. These complementary approaches can be used individually or jointly to identify emerging information requirements, estimate question-level demand probabilities, and prioritize candidate questions for advance response generation. In the present experiment, GPT-6 mini is used to construct ten representative questions, which are subsequently evaluated under two repeated-request scenarios comprising three and five requests per question. The experiment compares three inference-serving strategies—Baseline, Demand_Cache, and TPGIC—across three generation models: Gemma 3 1B, Qwen3 4B, and Gemma 3 4B. An independent Qwen3 1.7B model evaluates response quality. The complete experiment produces 720 request records, 60 independent-evaluator calls, and 180 method-level quality records. The results demonstrate that TPGIC achieves a normalized computational cost of 3 units compared with 10 units for Demand_Cache, representing a 70% reduction under the experimental cost-accounting framework. This advantage is observed in both repetition scenarios. Meanwhile, the actual generation-call counts of TPGIC and Demand_Cache are identical, indicating that the measured cost advantage is associated with the normalized computational-cost model rather than a reduction in the number of complete response-generation calls. The independent quality evaluation reports identical mean scores for all three methods within each generation model: 9.7 for Gemma 3 1B, 9.7 for Qwen3 4B, and 9.4 for Gemma 3 4B. No statistically detectable quality differences are observed in the current evaluation. These findings provide initial empirical evidence that TPGIC can achieve lower normalized computational expenditure than reactive Demand_Cache while maintaining identical observed mean response-quality scores under the tested workload. More broadly, the study establishes GPT-assisted representative question construction as an initial implementation of a more general multi-source future-demand identification methodology. By integrating potential external trend signals, LLM-based question generation, demand estimation, predictive response preparation, and computational-cost optimization within a unified research direction, TPGIC provides a foundation for developing proactive generative inference systems capable of anticipating not only what users may ask, but also which responses and computational resources should be prepared before future information demand materializes. Subsequent research will investigate larger question collections, multi-source demand prediction, temporally realistic request distributions, direct hardware-level cost measurements, and coordinated response and computational-state caching.