Microservice architectures dominate large‑scale cloud applications, offering rapid deployment and elastic scaling. Yet system reliability hinges on effective anomaly detection because a single service failure or performance glitch can quickly cascade across the call graph, degrading end‑to‑end behavior. Existing multimodal anomaly detectors suffer from two fundamental drawbacks: (i) they prioritize cross‑modality fusion at the expense of modality‑specific characteristics; (ii) they fall into an “identity‑shortcut” where reconstruction merely copies the input.
To overcome these issues we introduce LLMMAD, an autoregressive multimodal anomaly detection framework built on Large Language Models (LLMs). LLMMAD first embeds each modality and recombines them into a unified token sequence for the LLM, enabling the model to learn inter‑modality relationships. At the same time, separate reconstruction heads are attached to each modality, preserving modality‑specific information. The training objective is prediction‑driven autoregression: the model forecasts future states rather than reproducing current inputs, thereby eliminating the identity‑shortcut.
Extensive experiments on three real‑world microservice datasets show that LLMMAD consistently outperforms state‑of‑the‑art baselines, achieving F1‑score gains of 3.1%, 1.4% and 0.9% respectively. These results demonstrate a more reliable solution for microservice monitoring.
Review: LLMMAD’s combination of unified sequence modeling and modality‑wise reconstruction balances cross‑modal interaction with detail preservation, while the autoregressive prediction objective prevents trivial copying, making it a practical advance for real‑world microservice anomaly detection.