The Strongest Teacher Is Not Always the Best Teacher: Student-Centric Answer Selection

A new machine-learning framework challenges the common practice of using the highest-performing model to generate training data, showing that a weaker teacher can sometimes provide better supervision for a specific student model. The framework, called Student-Centric Answer Sampling (SCAS), was detailed in a paper submitted to arXiv on 26 May 2026 [1]. The authors argue that the field’s reliance on teacher-generated supervision—from synthetic responses to reasoning traces—has an implicit flaw: teacher test performance is treated as a proxy for teaching quality [1]. Their work demonstrates that even when multiple teachers provide correct answers to the same question, the answer from the strongest teacher is not necessarily the best supervision for a given student [1]. To address this, SCAS selects from verified teacher-generated answers according to an estimated student-centric learning cost [1]. The method is motivated by a token-wise gradient decomposition, from which the researchers derive an efficient forward-only proxy for this cost and use it to guide answer selection during training [1]. The paper reports experiments across 30 teacher models, 6 student base models, and 8 tasks, finding that SCAS consistently improves student performance [1]. The concept of questioning a dominant metric in a system has parallels in other domains. The 20th-century philosopher Karl Popper, for instance, rejected classical inductivist views of science in favor of empirical falsification, arguing that a theory can never be proven, only scrutinized with decisive experiments [3]. The SCAS framework similarly subjects the assumption of teacher superiority to empirical test, finding it wanting. In education history, the mismatch between a standardized system and the needs of specific learners has prompted the creation of alternative institutions. In 1973, land rights activist Eddie Koiki Mabo and Burnum Burnum founded the Black Community School in Townsville, Australia, to provide an education better suited to Indigenous children’s needs and culture, operating outside the state system for twelve years before closing due to funding and site lease issues [5]. The SCAS framework addresses a comparable mismatch inside machine-learning training pipelines, tailoring supervision to the student rather than defaulting to the highest-scoring model. The paper’s authors conclude that effective distillation should prioritize supervision matched to the current student rather than teacher strength alone [1]. The work was developed within arXivLabs, a framework that allows collaborators to develop and share new arXiv features directly on the website, with both individuals and organizations embracing values of openness, community, and user data privacy [1].