Are LLMs Ready for Anti-Pattern Detection in Microservice Architectures?

2026-02-24 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

A study investigated the readiness of Large Language Models (LLMs) for detecting architectural anti-patterns (APs) in microservice architectures (MSAs) using a prompt-based analysis pipeline over static repository artifacts. Researchers evaluated three general-purpose LLMs—ChatGPT 5.2, Gemini 3 Pro, and Qwen 3.5 Plus—on a benchmark of 13 microservice repositories annotated with 16 architectural anti-patterns. Their performance was compared against MARS, a static-analysis tool, using precision and recall metrics. Results indicate that LLMs offer useful support, achieving competitive performance on several anti-patterns, particularly those inferred from local, heterogeneous, or semantically rich evidence like "No API Versioning" (NAV) and "Hardcoded Endpoint" (HE). However, LLMs showed limitations on anti-patterns requiring explicit structural or cross-service dependency evidence, such as "Cyclic Dependency" (CD) and "Shared Persistence" (SP), where static analysis remains more reliable. The findings suggest LLMs are a promising complementary aid, not a replacement, for traditional analyzers in architectural assessment.

Key takeaway

For AI Architects evaluating microservice system quality, you should integrate LLM-based tools as a complementary layer to existing static analyzers. While your traditional tools like MARS will reliably detect structural anti-patterns such as cyclic dependencies, LLMs can effectively identify issues requiring semantic interpretation, like "No API Versioning" or "Hardcoded Endpoint." Consider a hybrid approach, combining LLMs with richer structural inputs like dependency graphs, to enhance detection coverage and reduce architectural decay.

Key insights

LLMs complement static analysis for microservice anti-pattern detection, excelling with semantic cues but struggling with structural dependencies.

Principles

• LLMs excel with semantic, local evidence.
• Static analysis is superior for structural proof.
• Interpretive flexibility risks false positives.

Method

The study used a four-step pipeline: repository flattening (Repomix), LLM prompt execution (parametric, AP-specific), MARS execution, and data analysis (precision/recall against ground truth).

In practice

• Use LLMs for qualitative design judgment.
• Combine LLMs with structural analysis tools.
• Prioritize static analysis for explicit dependencies.

Topics

• Microservice Architectures
• Anti-Pattern Detection
• Large Language Models
• Static Analysis
• Architectural Assessment
• Prompt Engineering
• MARS Tool

Code references

• yamadashy/repomix
• apolloconfig/apollo
• benwilcock/cqrs-microservice-sampler
• DescartesResearch/TeaStore
• ewolff/microservice

Best for: AI Scientist, AI Architect, Research Scientist

Related on AIssential

• FM-Agent: Scaling Formal Methods to Large Systems via LLM-Based Hoare-Style Reasoning — LLMs can automate formal specification generation and Hoare-style reasoning for large-scale software systems.
• Analyzing the Evolution of Structural Communities within Microservice Architecture — Temporal community detection identifies stable microservice structures and reveals cross-cutting services indicating potential anti-patterns.
• Evaluating Agentic Configuration Repair for Computer Networks — Agentic LLM architectures, augmented with verification and context tools, significantly improve network misconfiguration repair.
• Why SDD Breaks Down in Microservices—Part 3: Distributed Systems Need Distributed Context — Microservices SDD with LLMs requires architecture-level context via machine-readable contracts and a plan-first workflow to ensure cross-service rule adherence.
• Top Anti-Patterns to Avoid in Service Architecture — Seemingly good architectural decisions can accumulate into problematic service architectures.
• Finding Bugs with Claude and Property-based Testing — An AI agent using Claude and property-based testing effectively identifies bugs in major Python libraries by inferring code properties.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.