<![CDATA[Purpose of the study: This study investigates the multidisciplinary integration of artificial intelligence, learning analytics, cognitive psychology, and science education to evaluate three Large Language Model (LLM) configurations. It aims to optimize an adaptive digital scaffolding framework for primary computational thinking (CT) and scientific reasoning under tight latency constraints. Methodology: Deployed via Python 3.11 within an automated benchmarking ecosystem, OpenAI gpt-5-mini, Google gemini-3.5-flash, and Meta llama-3.3-70b-versatile (Groq LPU) were evaluated across 15 Bebras tasks (135 structured API interactions). The multidisciplinary validation applied content and psycholinguistic triage to analyze the interface between technical inference latency and the continuity of students' scientific inquiry processes. Main Findings: Meta Llama-3.3-70b achieved optimal performance with a 0.2687s latency, maximizing the Student Waiting Threshold (SWT) compliance margin to support uninterrupted scientific schema construction. OpenAI GPT-5 Mini exhibited superior Socratic instruction adherence (6.9% failure) but introduced a 2.3764s latency overhead. Gemini 3.5 Flash truncated crucial pedagogical contexts due to its constrained 3-token output distribution. Novelty/Originality of this study: This work introduces a multidisciplinary engineering blueprint that bridges hardware-level computing optimization with technology-enhanced science education. By formalizing a latency-constrained routing protocol, it establishes a theoretical model demonstrating how infrastructure responsiveness directly safeguards the cognitive sustainability of scientific reasoning and problem-solving sequences in primary STEM learning contexts.]]>
Copyrights © 2026
