VisAnomReasoner-3B

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

🤗 VisAnomBench · 🤗 VisAnomReasoner-7B

---

Overview

VisAnomReasoner-3B is a parameter-efficient vision-language model for time-series anomaly detection. It is a LoRA adapter for Qwen2.5-VL-3B-Instruct, fine-tuned on VisAnomBench to jointly localize anomalies and explain them from time-series plots.

Given a time-series plot, VisAnomReasoner locates the anomalous intervals and explains the visual evidence behind each decision.

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

---

Key Idea

Public time-series anomaly detection benchmarks typically provide interval annotations but not natural-language rationales, which makes it difficult to fine-tune VLMs to produce grounded, interpretable decisions.

VisAnomBench fills this gap by augmenting four public benchmarks with temporally grounded explanations, selected from multiple large VLMs using fine-grained, task-specific rewards. Fine-tuning on this explanation-augmented supervision teaches a compact VLM to couple interval-level localization with step-by-step reasoning grounded in what is visible in the plot.

---

Highlights

• Explanation-augmented supervision: Trained on VisAnomBench, which pairs anomaly intervals with step-by-step explanations for time series from KPI, GutenTAG, UCR-EGI, and UCR-TSAD.

• Joint localization and explanation: Outputs anomaly intervals in a parseable format together with a numbered reasoning trace.

• Parameter-efficient: A rank-32 LoRA adapter with 74.3M trainable parameters on top of Qwen2.5-VL-3B-Instruct.

• Strong localization: On VisAnomBench, both VisAnomReasoner variants achieve the highest precision, recall, F1, and overlap against 15 baselines, including general-purpose VLMs, LLM/VLM anomaly detectors, time-series foundation models, and classical detectors.

• Tighter intervals: VisAnomReasoner-3B exceeds the overlap score of every baseline by at least 6.26 percentage points.

---

Resources

• Paper: Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection
• Project Page: plan-lab.github.io/projects/VisAnom
• Code: github.com/PLAN-Lab/VisAnom
• Dataset: PLAN-Lab/VisAnomBench

---

License

Subject to the Qwen Research License (non-commercial use only). Built with Qwen.

---

Citation

If you find VisAnomReasoner useful in your research, please consider citing our work:

@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}
}