dsm-llm-modularization

# DSM Modularization with Large Language Models > LLM-based approach to Design Structure Matrix modularization achieving near-reference quality within 30 iterations without specialized optimization code. Introduces the semantic-alignment hypothesis governing when domain knowledge helps or hurts LLM-based optimization. ## Metadata - **Source**: arXiv:2604.28018 - **Authors**: Shuo Jiang, Jianxi Luo - **Published**: 2026-04-30 - **Subjects**: cs.CE (Computational Engineering); cs.AI ## Core Methodology ### Key Innovation DSM modularization partitions system elements into cohesive modules — a fundamental combinatorial challenge in engineering design. This paper shows that **LLMs can solve DSM modularization** through iterative prompting, without requiring specialized optimization code. ### Technical Framework **Three-stage LLM-based modularization**: 1. **Input Representation** - Convert DSM (N×N dependency matrix) into natural language description - Include element names, dependency strengths, and constraints - **Best practice**: Include raw matrix structure + semantic labels 2. **Objective Formulation** - Frame modularization as minimizing inter-module dependencies - Use standard metrics: Minimum Description Length (MDL), clustering coefficient - **Best practice**: Provide explicit objective function in natural language 3. **Solution Pool Design** - Maintain a pool of candidate solutions across iterations - LLM generates new candidates, evaluates against pool - **Best practice**: Pool size 5-10, retain top-k by objective value ### Critical Finding: Semantic-Alignment Hypothesis **Counterintuitive result**: Domain knowledge, beneficial in DSM *sequencing*, consistently **impairs** performance on DSM *modularization* for complex DSMs. **Hypothesis**: LLM effectiveness with domain knowledge depends on **semantic alignment** between: - The LLM's functional priors (what it "knows" about the domain) - The optimization objective (structural vs. functional) **When knowledge helps**: The domain semantics align with the optimization goal **When knowledge hurts**: The LLM's functional priors conflict with structural optimization objectives **Testable condition**: Before adding domain knowledge to LLM prompts, verify whether the knowledge supports or contradicts the mathematical objective. ### Implementation Guide **Step 1: DSM Construction** - Build N×N matrix of element dependencies - Weight edges by dependency strength - Validate matrix completeness **Step 2: LLM Prompt Design** ``` System: You are an expert in system architecture modularization. Task: Partition these {N} system elements into cohesive modules. Input: {DSM_description_with_element_names_and_dependencies} Objective: Minimize inter-module dependencies (total coupling between modules). Constraints: {max_modules, min_module_size, etc.} Format: Return module assignments as JSON. ``` **Step 3: Iterative Optimization** - Run 30 iterations of LLM prompting - Each iteration: present current best + ask for improvement - Track objective value across iterations - Convergence typically within 20-30 iterations **Step 4: Validation** - Compare against reference solutions (if available) - Compute modularity metrics (MDL, clustering coefficient) - Evaluate semantic coherence of modules ### Pitfalls - **Semantic misalignment**: Adding domain knowledge to prompts can degrade results if the knowledge conflicts with structural optimization. Test both with and without domain context. - **LLM variability**: Different backbone LLMs produce different quality results. Test across 2-3 models. - **Prompt sensitivity**: Input representation format significantly impacts results. Use ablation studies to find optimal format. - **Scale limits**: Method validated on DSMs up to moderate size. Very large DSMs (N > 200) may need hierarchical decomposition. ## Applications - System architecture design - Software modularization - Product family planning - Organization design - Supply chain clustering ## Related Skills - modern-systems-engineering-patterns - emergent-systems-design - agent-first-bootstrap