2026-10-14T11:47:27Z https://www.scidb.cn/oai

10.57760/sciencedb.013zp 2026-09-29T16:15:36Z

2026-09-29 When Does Guidance Still Help? Separating Scaffolding Effects from Answer-Leakage Effects in Higher-Education Cryptography Instruction - A Simulation Study doi:10.57760/sciencedb.013zp en LLM tutors default to giving the answer. A tutor that withholds it faces an attribution problem: leak less while learners improve — two things vary at once; no system we reviewed separates them. In SAGE scaffolding and withholding act through independently switchable channels — separable by design; withholding enters the objective as a scalar intensity whose cost is a boundary at λ* = 0.25. In a university cryptography simulation, leakage falls from 0.86 to 0.000, earned solves rise then fall, and at least 85% of the improvement is attributable to scaffolding under the stipulation that leaked answers do not teach (leak_durable = 0.012). Results are properties of the apparatus under a stipulated learner model, not students. intelligent tutoring systems; scaffolding; answer leakage; attribution; simulation study Yuting Zuo Minjie Zhang Hao Huang RESTRICTED dataset Science Data Bank