Search for a command to run...
본 논문은 대규모 언어 모델(LLM)에서 반복적으로 발생하는 추론 붕괴, 논리적 공백 메움, 결심 위치 불명확성 문제를 구조적 결함의 관점에서 재정식화한다. 기존 접근이 성능 향상이나 사후 보정에 의존했다면, 본 연구는 실패의 원인을 "공백(Blank)"과 "결심 경로의 분기"라는 구조적 문제로 규정한다. 이에 따라 공백을 봉인하고 단일한 결심 경로를 강제하는 실행 아키텍처인 구조적 닫힘(Structural Closure)을 제안한다. 이 아키텍처는 최소 허용 필드, 승인 게이트, 단일 경로 전이 규칙을 통해 탐색과 결심을 분리하고, 추론 과정의 안정성과 재현성을 확보한다. 또한 사건 단위 추적 구조를 도입하여 판정 가능성과 책임성을 강화한다. 구조적 닫힘은 모델 규모 확장이 아닌 실행 구조의 재설계를 통해 LLM의 반복 실패를 억제하고, 고위험 도메인에서의 예측 가능성과 결정 무결성을 향상시키는 것을 목표로 한다. This paper reformulates recurrent inference collapse and decision ambiguity in large language models (LLMs) as structural defects rather than performance limitations. Instead of relying on scaling or post-hoc correction, the study identifies logical blanks and branching decision paths as the primary causes of instability. To address this, it proposes Structural Closure, an execution architecture that seals blanks and enforces a single admissible decision path. Through admissibility gates, minimal required fields, and one-way transition rules, the architecture separates exploration from commitment and stabilizes decision formation at the structural level. By introducing event-level traceability, it enhances reproducibility and accountability. Structural Closure aims to mitigate recurrent LLM failures not by increasing model size, but by redesigning the execution structure to ensure decision integrity in high-stakes domains.