VisAnomBench

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

Xiaona Zhou, Muntasir Wahed, Tianjiao Yu, Constantin Brif, Ismini Lourentzou

🎉 Accepted to NeurIPS 2026

📄 Paper · 🌐 Project Page · 💻 Code

🤗 VisAnomReasoner-3B · 🤗 VisAnomReasoner-7B

---

Overview

VisAnomBench is an explanation-augmented benchmark for vision-language time-series anomaly reasoning. Each sample pairs a rendered time-series plot with a prompt and ground-truth anomaly intervals. Training samples also include a step-by-step explanation grounded in the plot. The benchmark is used to train and evaluate VisAnomReasoner-3B and VisAnomReasoner-7B.

Left: Given a time-series plot, VisAnomReasoner, trained on VisAnomBench, locates anomalies with details grounded in the plot. Right: VisAnomReasoner-7B outperforms the strongest baselines on VisAnomBench and TSB-AD-U.

---

Dataset Structure

Split	Samples	UCR-EGI	GutenTAG	UCR-TSAD	KPI
train	2,576	1,682	563	200	131
test	740	415	247	49	29

Data Fields

Field	Description
image	Path to the plot image, relative to the root of the VisAnom code repository, for example VisAnomBench/VisAnomBench_plot_test/test_0.png.
query	Prompt given to the model: context about the series, task instructions, and the required output format.
resp	Target response: image description, rationale, step-by-step reasoning, and a final answer of the form <anomaly>True</anomaly><index>(start, end)</index> followed by the reasoning in <think>...</think>. Empty in the test split.
anomalies	Ground-truth anomaly intervals as inclusive (start, end) pairs, for example (52000, 52620). KPI intervals use timestamps, for example (2016-11-09 03:35:00, 2016-11-09 03:45:00).

Download

To use the data with the VisAnom code, run this from the root of the code repository:

hf download PLAN-Lab/VisAnomBench \
  --repo-type dataset \
  --include "VisAnomBench_train.json" \
  --include "VisAnomBench_test.json" \
  --include "VisAnomBench_plot_train/*" \
  --include "VisAnomBench_plot_test/*" \
  --include "VisAnomBench_train_ts/*" \
  --include "VisAnomBench_test_ts/*" \
  --local-dir VisAnomBench

How the Explanations Were Created

For each time-series segment, four general-purpose VLMs (Grok-4-Fast, LLaMA-4-Maverick, Gemma-3-27B-IT, and Qwen2.5-VL-32B) generated candidate outputs, each with an anomaly decision, predicted intervals, and a reasoning trace. Candidates were ranked by a composite reward that combines anomaly localization accuracy with explanation quality (visual groundedness, axis awareness, and clarity) scored by Qwen2.5-VL-72B, and the highest-scoring candidate was kept as the target response. See the paper for details.

Source Datasets

Source	Reference	License
KPI	Xu et al., "Unsupervised Anomaly Detection via Variational Auto-Encoder for Seasonal KPIs in Web Applications," WWW 2018	MIT
GutenTAG	Wenig et al., "TimeEval: A Benchmarking Toolkit for Time Series Anomaly Detection Algorithms," PVLDB 2022	MIT
UCR-EGI	Gao, Lin, and Brif, "Ensemble Grammar Induction for Detecting Anomalies in Time Series," EDBT 2020	See source
UCR-TSAD	Wu and Keogh, "Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of Progress," TKDE 2021	See source

The source time series remain subject to the terms of their original sources. Please also cite the source datasets when using VisAnomBench.

License

VisAnomBench is released under the CC BY 4.0 license.

Citation

@inproceedings{zhou2026tiny,
  title={Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection},
  author={Zhou, Xiaona and Wahed, Muntasir and Yu, Tianjiao and Brif, Constantin and Lourentzou, Ismini},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}