Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection

Problem
This paper addresses the inadequacy of existing Vision-Language Models (VLMs) in effectively detecting anomalies in time-series data. Prior research has shown that large language and multimodal models struggle with this task, particularly due to the lack of natural-language rationales accompanying interval annotations in public anomaly detection benchmarks. This gap hinders the fine-tuning of VLMs for producing grounded and interpretable decisions in anomaly detection. The authors present a preprint work, indicating that the findings have not yet undergone peer review.

Method
The authors introduce VisAnomBench, a curated benchmark specifically designed for time-series anomaly detection. This benchmark is constructed from existing public time-series datasets and is enhanced with high-quality anomaly explanations derived from multiple large VLMs, utilizing fine-grained, task-specific rewards. The core technical contribution is the development of VisAnomReasoner, a parameter-efficient VLM fine-tuned on VisAnomBench. The architecture details are not explicitly disclosed, but the model is designed to leverage the curated explanations to improve anomaly localization. The training process involves optimizing for precision and F1 score, although specific compute resources are not mentioned.

Results
VisAnomReasoner demonstrates significant improvements over baseline models on the VisAnomBench. The model achieves enhancements of at least 21.23 percentage points in precision and 23.87 percentage points in F1 score compared to named baselines. Furthermore, when evaluated on the TSB-AD-U benchmark, VisAnomReasoner shows strong cross-benchmark generalization, with improvements of 9.57 percentage points in precision and 13.39 percentage points in F1 score. These results indicate that the proposed model not only excels in anomaly detection but also maintains robustness across different datasets.

Limitations
The authors acknowledge that while VisAnomReasoner shows promising results, the reliance on high-quality anomaly explanations may limit its applicability to datasets where such annotations are not available. Additionally, the paper does not discuss the scalability of the model to larger datasets or its performance in real-time anomaly detection scenarios. The lack of detailed architecture and training compute specifications may also hinder reproducibility and further optimization by other researchers.

Why it matters
This work has significant implications for the field of anomaly detection in time-series data, particularly in applications where interpretability is crucial, such as finance, healthcare, and industrial monitoring. By bridging the gap between VLMs and time-series anomaly detection, the authors provide a framework that can enhance the interpretability of model decisions, potentially leading to more trustworthy AI systems. The introduction of VisAnomBench sets a precedent for future research to build upon, encouraging the development of more sophisticated models that can leverage natural language for improved anomaly detection.

Authors: Xiaona Zhou, Muntasir Wahed, Tianjiao Yu, Constantin Brif, Ismini Lourentzou
Source: arXiv:2605.30344
URL: https://arxiv.org/abs/2605.30344v1