Skip to content

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

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

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

Authors: Xiaona Zhou, Muntasir Wahed, Tianjiao Yu, Constantin Brif, Ismini Lourentzou Date: 2026-05-28 Paper ID: openalex:2605.30344

Summary

This paper introduces VisAnomReasoner, a parameter-efficient vision-language model designed to improve interpretability and performance in time-series anomaly detection. To overcome the lack of existing labeled rationales, the authors curate VisAnomBench, which incorporates high-quality natural-language anomaly explanations generated by large models. Experimental results confirm that VisAnomReasoner achieves significant improvements in anomaly localization accuracy and F1 scores, outperforming existing baselines and demonstrating robust generalization to the TSB-AD-U benchmark.

Key Contributions

  • Introduced VisAnomBench, the first time-series anomaly detection benchmark providing natural-language rationales for anomaly intervals.
  • Developed VisAnomReasoner, a parameter-efficient VLM that improves precision and F1 scores by >21% on VisAnomBench compared to existing baselines.
  • Demonstrated superior cross-benchmark generalization on the TSB-AD-U dataset with double-digit percentage gains in F1 score.

Key Concepts

  • visanombench: A curated benchmark for time-series anomaly detection that provides explicit natural-language rationales for anomaly intervals.

Archivist Review

The review focused on identifying the core infrastructure contribution of the paper. VisAnomBench was approved as a critical new resource for enabling grounded VLM evaluation in time-series anomaly detection, whereas the model (VisAnomReasoner) was rejected as a paper-local instantiation of existing parameter-efficient fine-tuning techniques. No open questions were identified that rose to the level of substantial, unresolved research bottlenecks.

Approved Concepts

  • VisAnomBench: It addresses the fundamental lack of natural-language rationales in existing time-series anomaly detection benchmarks, which is a major barrier for evaluating and training VLM-based interpretability in this domain.

Rejected Candidates

  • [concept] VisAnomReasoner (visanomreasoner) - subcomponent_of_broader_mechanism: The model is a specific architecture implementation rather than a general mechanism; it is best described as an instance of VLM adaptation for time-series.

Datasets

Metadata & Links

url
https://arxiv.org/abs/2605.30344
paper_id
2605.30344
paper_source
openalex
domain
nlp
tags
time-series-question-answering-tsqa
architectures
datasets
visanombench
concept_slugs
visanombench
dataset_slugs
visanombench
skill
TimeSeriesSkill
created_at
2026-05-31T06:31:43Z