Architectural inference of behavior-driven patterns by LLMs

Establish whether large language models can reliably infer and synthesize behavior-driven architectural patterns—particularly the Observer pattern—from event-driven, complex requirements without explicit pattern naming, in order to achieve dependable software design synthesis.

Background

In analyzing performance on a high-complexity, event-driven sensor network scenario, the study reports that no evaluated model inferred the need for the Observer pattern from the requirements, indicating difficulties with behavior-driven architectural reasoning.

The authors highlight that this shortcoming reflects a broader limitation: current models struggle to infer architectural solutions from dynamic behavior rather than explicit structural cues, leading them to state that architectural inference—especially for behavior-driven patterns—remains an open challenge for LLM-based design systems.

References

Key Insight: Architectural inference — particularly for behavior-driven patterns — remains a major open challenge for LLM-based design systems.

— Reliability of Large Language Models for Design Synthesis: An Empirical Study of Variance, Prompt Sensitivity, and Method Scaffolding  (2604.00851 - Iftikhar et al., 1 Apr 2026) in Section 5, Answer to RQ4 (Results and Discussion)