SML-CML: Building a Two-Layer Framework for Structured Reasoning and Value Evaluation of Large Language Models

SML-CML is an innovative two-layer reasoning framework proposed to address the deficiencies of Large Language Models (LLMs) in deep contextual understanding and value judgment. It consists of the Semantic Model Layer (SML) and the Cosmological Model Layer (CML). This article introduces its design principles, practical application scenarios (such as medical diagnosis evaluation, cross-cultural education, ethical reasoning, etc.), and practical value, aiming to help AI systems achieve deep contextual reasoning based on worldviews.

Current LLMs perform well in pattern matching and text generation, but in scenarios requiring deep contextual understanding and value judgment—such as medical diagnosis, ethical reasoning, and cross-cultural dialogue—they often struggle to understand "why it is correct" and "for whom it is correct", showing obvious limitations.

The SML-CML framework is divided into two complementary layers:

Semantic Model Layer (SML)

Responsible for constructing problem frameworks, focusing on meaning generation, interpretive framework construction, and comparison of different interpretations. For example, when facing the symptom of "fatigue", it guides thinking about interpretations from different medical perspectives and structurally handles abductive reasoning and interpretive frameworks.

Cosmological Model Layer (CML)

Encodes the value systems or worldviews that support interpretations, answering questions such as "Should health prioritize natural harmony or pain elimination?" This grounds normative evaluations in traditions like ethics and culture, enabling deep adequacy assessment.

SML-CML demonstrates application value in multiple domains:

• AI Diagnosis Evaluation: Distinguish between semantic coherence and worldview-sensitive adequacy, going beyond traditional correctness judgments that rely on training data.
• Cross-Cultural Education: Demonstrate how worldviews shape reasoning, helping students understand the meaning differences of the same symptom due to different interpretive frameworks.
• Ethical Reasoning and Dialogue: Explicitly reveal the value assumptions behind moral judgments, facilitating constructive communication in contexts of value pluralism.
• Clinical Communication Improvement: Help healthcare providers understand patients' interpretations of diseases and support patient-respectful communication.

The framework's supporting practical materials follow three principles:

1. Respect Patients' Worldviews: Do not belittle disease concepts based on religion, culture, etc.;
2. Privacy First: Only use anonymized and generalized case summaries; do not upload original patient forms;
3. Transparent Use: Provide explicitly reviewable support; avoid opaque AI usage. Materials include questionnaire templates, anonymization rules, benchmark test templates, etc., with no identifiable patient information.

Develop AI-assisted evaluation benchmark materials based on the framework, including worldview-sensitive clinical questionnaires, anonymization conversion rules, structured comparison prompts, and hierarchical scoring tools. The evaluation goal is not model ranking, but to test AI's performance in interpretation tasks, weaknesses in worldview-sensitive reasoning, discrepancies between semantic and cosmological adequacy, and evaluation differences among human assessors.

SML-CML is not an automated moral authority or a replacement for human judgment; instead, it helps clarify the meaning layers of AI outputs and the worldview-insensitive reasoning masked by semantic coherence, emphasizing the necessity of human interpretation in complex dialogues such as ethics and medicine. It is both a conceptual model of layered meaning and a practical framework for evaluating AI's assistive role and over-delegation.

With the in-depth application of LLMs, it is urgent to make AI reasoning more context-sensitive and value-aware. By separating meaning generation and value evaluation, SML-CML provides a direction for AI to go beyond pattern matching and interact deeply with human thought, which is worth exploring by practitioners and researchers in the fields of AI ethics, medical AI, and cross-cultural computing.