Tiny but TrustedEfficient Vision-Language Reasoning for Time-Series Anomaly Detection

Xiaona Zhou♦, Muntasir Wahed♦, Tianjiao Yu♦, Constantin Brif♣, Ismini Lourentzou♦
♦University of Illinois Urbana-Champaign♣Sandia National Laboratories

NeurIPS 2026

PaperarXivCode (coming soon)Dataset (coming soon)Checkpoint (coming soon)

TL;DR

We introduce VisAnomReasoner, a compact vision-language model that detects and localizes anomalies in time-series plots and explains its decisions. Trained with visually grounded reasoning on our curated VisAnomBench dataset, VisAnomReasoner connects unusual patterns to specific time intervals and supporting evidence.

Do you see any anomalies on this plot?
Time-series plot from the paper teaser, showing the anomalous downward deflection.

Anomaly localization and explanation

VisAnomReasoner reads a time-series plot and returns an anomaly decision, a localized interval, and reasoning grounded in the visible signal.

Anomaly detected
Yes
Predicted index interval
16970–17079
Training time series
2,576
Model variants
3B / 7B
F1 on VisAnomBench · 7B
73.94%

Despite its small scale, VisAnomReasoner consistently outperforms general-purpose VLMs, specialized LLM/VLM anomaly detectors, time-series foundation models, and classical detectors, improving precision and F1 by at least 21.23 and 23.87 percentage points on VisAnomBench.

Abstract

Recent advances in Vision-Language Models (VLMs) have achieved impressive performance across many tasks, yet prior studies report unsatisfactory performance when applying large language or multimodal models to finding abnormal patterns in sequential data. Public anomaly detection benchmarks typically provide interval annotations but not natural-language rationales, making it difficult to fine-tune VLMs to produce grounded, interpretable decisions. To address this gap, we construct VisAnomBench, a curated benchmark built from public time-series datasets and augmented with high-quality anomaly explanations selected from multiple large VLMs using fine-grained, task-specific rewards. Through fine-tuning on this benchmark, we develop VisAnomReasoner, a parameter-efficient VLM for time-series anomaly detection. Experimental results on VisAnomBench show that VisAnomReasoner achieves more accurate anomaly localization and consistently outperforms all baselines, with improvements of at least 21.23 and 23.87 percentage points in precision and F1, respectively. Additional experiments on the TSB-AD-U benchmark demonstrate strong cross-benchmark generalization, with VisAnomReasoner improving precision and F1 by 9.57 and 13.39 percentage points, respectively.

Contributions

  • VisAnomBench provides explanation-augmented supervision. We introduce VisAnomBench, an explanation-augmented benchmark constructed from four public time-series anomaly detection datasets. Each plot is paired with anomaly interval annotations and temporally grounded, step-by-step explanations. Reasoning traces from multiple large VLMs are selected using fine-grained rewards for anomaly localization, visual groundedness, axis awareness, and clarity. The benchmark contains 2,576 training and 740 held-out test time series.
  • VisAnomReasoner jointly localizes and explains anomalies. We develop a parameter-efficient vision-language model by supervised fine-tuning Qwen2.5-VL-3B and Qwen2.5-VL-7B on VisAnomBench. Structured outputs separate anomaly decisions, interval localization, and reasoning, allowing the model to explain spikes, level shifts, periodic deviations, and other anomalies using evidence in the plotted time series.
  • Evaluation demonstrates accurate localization and cross-benchmark transfer. VisAnomReasoner improves precision and F1 by at least 21.23 and 23.87 percentage points on VisAnomBench. On TSB-AD-U, it improves precision and F1 by 9.57 and 13.39 percentage points, respectively, demonstrating generalization beyond the training benchmark.
  • Ablations isolate the benefit of explanation-augmented supervision. Comparisons between pretrained models, interval-only fine-tuning, and reasoning-augmented fine-tuning show that interval supervision reduces false positives, while adding grounded reasoning traces further improves precision and recall and yields more accurate temporal boundaries.

VisAnomReasoner Method

VisAnomReasoner is a parameter-efficient vision-language model for reasoning-based time-series anomaly detection from plots. Given a plotted time series and optional context, the model jointly predicts anomaly intervals and generates grounded step-by-step explanations aligned with observable visual patterns such as spikes, level shifts, and periodic deviations. Unlike prior prompt-based approaches, VisAnomReasoner is supervised using explanation-augmented training data from VisAnomBench, where each example contains both target anomaly intervals and preferred reasoning traces. The model is implemented by supervised fine-tuning of Qwen2.5-VL-3B and Qwen2.5-VL-7B using structured outputs with dedicated tags for anomaly decisions, interval localization, and reasoning generation.

  1. Input

    Time-series plot

    The plotted signal retains its shape, axes, and temporal context.

  2. Model

    Qwen2.5-VL · 3B / 7B

    Supervised fine-tuning uses anomaly intervals and grounded reasoning traces from VisAnomBench.

  3. Output

    Decision, interval, explanation

    Structured outputs identify whether an anomaly exists, where it occurs, and the visual evidence.

VisAnomBench Dataset

VisAnomBench is an explanation-augmented benchmark for vision-language time-series anomaly reasoning, constructed from four public anomaly detection datasets spanning diverse domains and anomaly types. For each time-series plot, the benchmark provides (1) anomaly interval annotations and (2) temporally grounded step-by-step explanations aligned with the labeled anomalies. To build the dataset, we segment and render time series as plots, elicit structured reasoning traces from multiple large VLMs, and select the highest-quality explanation using a reward that combines anomaly localization accuracy, visual groundedness, axis awareness, and clarity. The final benchmark contains 2,576 training and 740 held-out test time series for supervised fine-tuning and evaluation of reasoning-based anomaly detection models.

VisAnomBench datasets

VisAnomBench datasets
Benchmark #TS #Domains Avg length Anomaly rate (%) #Explanations Explanation length
KPI 160 1 1,777 4.3 640 170
GUTENTAG 810 10 5,000 2.2 3240 188
UCR-EGI 2,097 5 8,268 5.8 8388 112
UCR-TSAD 249 8 2,755 6.9 996 112

Quantitative Results

Anomaly detection on VisAnomBench

Anomaly detection and localization on the held-out VisAnomBench split, comparing vision-language models, language-model anomaly detectors, and conventional detectors.

Anomaly detection on VisAnomBench
Model TP ↑ FP ↓ FN ↓ Precision (%) ↑ Recall (%) ↑ F1 (%) ↑ Overlap (%) ↑
Vision-language models
Grok-4-Fast (314B) 242 837 517 22.43 31.88 26.33 3.26
LLaMA-4-Maverick (17B) 442 1128 317 28.15 58.23 37.96 12.92
Qwen2.5-VL-7B 517 1784 242 22.47 68.12 33.79 14.04
Idefics3-8B 123 2057 636 5.64 16.21 8.37 2.25
SmolVLM-7B 49 1416 710 3.34 6.46 4.41 1.29
LLaVA-7B 9 523 750 1.69 1.19 1.39 0.23
Time-series foundation models
TimesFM (200M) 196 543 563 26.52 25.82 26.17 2.03
Chronos (120M) 200 538 559 27.10 26.35 26.72 2.27
Language-model methods
AnomLLM (GPT-4o) 149 862 610 14.74 19.63 16.84 4.04
LLM-TSAD (GPT-4o) 477 5621 282 7.82 66.53 14.00 20.81
LLMAD (GPT-4o) 349 402 410 46.47 45.98 46.23 17.31
VLM4TS (GPT-4o) 435 642 324 40.39 57.31 47.39 17.94
Conventional detectors
Sub-PCA 329 411 430 44.46 43.35 43.90 17.41
Matrix Profile 290 450 469 39.19 38.21 38.69 8.33
IForest 362 378 397 48.92 47.69 48.30 20.53
VisAnomReasoner
VisAnomReasoner (3B) 564↑47 240↓138 195↓47 70.15↑21.23 74.30↑6.18 72.17↑23.87 27.07↑6.26
VisAnomReasoner (7B) 576↑59 223↓155 183↓59 72.09↑23.17 75.88↑7.76 73.94↑25.64 25.35↑4.54

Generalization to TSB-AD-U

Cross-benchmark anomaly detection on unseen time series, evaluated with standard point-wise metrics and event-level affiliation metrics.

Generalization to TSB-AD-U
Method Standard metrics Affiliation metrics
Precision (%) ↑ Recall (%) ↑ F1 (%) ↑ Precision (%) ↑ Recall (%) ↑ F1 (%) ↑
Grok-4-Fast (314B) 27.95 15.53 16.78 32.43 33.24 32.20
LLaMA-4-Maverick (17B) 63.35 46.49 48.89 76.07 79.18 76.75
Qwen2.5-VL-7B 66.18 48.85 49.52 75.90 74.93 74.36
Idefics3-8B 45.12 35.11 34.64 50.83 59.02 53.03
SmolVLM-7B 35.07 12.48 15.80 37.57 35.84 35.75
LLaVA-7B 14.06 0.72 1.35 14.06 8.16 10.15
TimesFM (200M) 32.32 43.21 32.19 71.09 57.23 60.08
Chronos (120M) 47.18 55.37 47.23 74.36 75.61 73.31
AnomLLM (GPT-4o) 12.97 10.54 11.63 13.05 7.08 8.42
LLM-TSAD (GPT-4o) 31.74 28.40 29.98 66.69 52.76 53.69
LLMAD (GPT-4o) 32.96 37.54 22.91 67.50 73.37 64.16
VLM4TS (GPT-4o) 29.78 40.10 19.84 72.52 51.38 54.06
Sub-PCA 11.80 16.82 12.28 48.22 31.71 34.99
Matrix Profile 9.40 15.84 9.91 44.90 41.44 41.90
IForest 12.44 17.89 13.05 50.44 41.51 42.67
VisAnomReasoner (3B) 77.93 40.46 49.59 95.05 71.69 78.36
VisAnomReasoner (7B) 75.75 60.91 62.91 95.27 81.51 85.58

Ablation Studies

Consistent gains across model sizes

Both 3B and 7B models improve over their Qwen2.5-VL baselines. Filled gray bars show base scores; blue and orange outlines show VisAnomReasoner scores, with arrows and percentages marking relative improvements.

Base and supervised fine-tuned Qwen2.5-VL scores for 3B and 7B models on a 0–1 axis. Filled gray bars show base scores; blue and orange outlines show SFT scores. Relative gains for 3B and 7B are 180% and 220% in precision, 2% and 11% in recall, 94% and 119% in F1, and 134% and 81% in overlap.

Effect of reasoning-augmented fine-tuning

Ablation studies compare the pretrained base model, interval-only fine-tuning, and reasoning-augmented fine-tuning. While interval-only supervision substantially reduces false positives, adding reasoning traces further improves both precision and recall, showing that explanation-augmented supervision contributes directly to improved anomaly localization.

Effect of reasoning-augmented fine-tuning
Mode TP FP FN Precision Recall F1
Base 517 1784 242 22.47 68.12 33.79
No reasoning 516 393 243 56.76 67.98 61.87
With reasoning 576 223 183 72.09 75.88 73.94

Qualitative Results

VisAnomReasoner precisely localizes the anomalous interval with visually grounded, structured reasoning; whereas other methods exhibit coarser localization or produce numerous spurious intervals with less grounded explanations.

Qualitative anomaly reasoning comparisons on VisAnomBench.

BibTeX

@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}
}