5월에 예정되어있는 콜로퀴움을 안내해드립니다. 

* 일시 : 5. 14. (목), 17시-18시

* 장소 :종A317호

* 연사 : 박예찬 (세종대학교)

* 강연 분야: 응용수학(인공지능)

* 강연 제목 : Do Language Models Understand Math?

 * 초록 : The recent success of large language models (LLMs), such as ChatGPT, marks a major milestone in artificial intelligence. Despite their ability to generate fluent and contextually appropriate text, it remains unclear whether these models possess genuine reasoning capabilities. This limitation is especially evident in mathematics, where precise logical operations are required, making it a natural testbed for evaluating whether LLMs can truly learn and apply exact rules rather than relying on statistical patterns. In this talk, we investigate whether language models can internalize mathematical rules purely from data through the lens of the grokking phenomenon—a striking training behavior in which models abruptly transition from memorization to near-perfect generalization after prolonged optimization.

By analyzing the internal training dynamics, we propose that generalization is driven by the acquisition of intrinsic symmetries underlying the task. We identify a consistent three-stage progression: (i) memorization, (ii) symmetry acquisition, and (iii) geometric organization, and show that generalization emerges during the symmetry acquisition phase, followed by a reorganization of representations into structured, task-aligned geometries. We validate this symmetry-driven perspective across diverse algorithmic domains, including algebraic, structural, and relational reasoning tasks. Building on these insights, we introduce a symmetry-based diagnostic that predicts the onset of generalization and propose principled strategies to accelerate it, offering a new perspective on how learning systems move beyond memorization toward rule-based reasoning.

*대상: 수학과 학부생, 복수전공자 및 대학원생

*학부 재학 중 특강 4회 이상 참석은 수학과 졸업 요건입니다. (복수전공생도 동일 요건)