Source-linked AI summary
Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection
Xiaona Zhou, Muntasir Wahed, Tianjiao Yu, Constantin Brif, Ismini Lourentzou
TL;DR
Time-series anomaly detection lacks explanation-rich supervision, limiting grounded VLM reasoning. The paper introduces VisAnomBench and fine-tunes VisAnomReasoner on it; the resulting compact model improves localization and F1 across benchmarks, while its scope remains dependent on domain-specific anomaly definitions.
Problem
Public anomaly benchmarks generally provide anomaly locations without explanations, limiting supervised adaptation of vision-language models for grounded anomaly reasoning.
Method
The paper constructs VisAnomBench with temporally grounded explanations and fine-tunes VisAnomReasoner, a parameter-efficient VLM that predicts intervals and grounded explanations.
Results
21.23 and 23.87 percentage points are the minimum precision and F1 improvements achieved on VisAnomBench, with strong generalization reported on TSB-AD-U.
Takeaways & Limitations
Explanation-augmented supervision supports joint anomaly localization and plot-grounded explanation generation with a compact vision-language model.
Takeaways & Limitations
Applicability is bounded because anomaly definitions vary across domains.
Abstract
from arXiv · showhide
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.
1. Introduction
Time-series anomaly detection requires interpretable interval-level decisions, but existing methods and benchmarks provide limited reasoning supervision. The paper addresses this gap with VisAnomBench and VisAnomReasoner, which jointly support anomaly localization and explanation generation.
- Existing anomaly detectors typically output numeric scores or binary labels, offering limited explanation for why an interval is anomalous.
- Public datasets usually provide anomaly intervals or point-wise labels without natural-language rationales, limiting supervised fine-tuning of VLMs.
- VisAnomBench augments multiple public benchmarks with temporally grounded explanations aligned with anomaly intervals.
- VisAnomReasoner is a compact, parameter-efficient VLM that improves anomaly localization and temporal alignment across multiple evaluation metrics.
- The paper formulates anomaly detection as plot-grounded vision-language reasoning requiring interval localization and structured explanation generation.
- 21.23 and 23.87 percentage points are the minimum precision and F1 improvements reported for VisAnomReasoner on VisAnomBench.
2. Related Work
Prior time-series reasoning methods commonly encode sequences as text or tokens, while visual approaches use plots to expose temporal patterns. Existing methods remain constrained by context windows and lack benchmarks supporting joint localization with plot-consistent explanations.
- LLM-based time-series methods encode sequences as text or symbolic representations for reasoning, question answering, forecasting, or classification.
- Context-window limits restrict many language-model approaches to short sequences and can cause information loss or fragmented temporal context.
- Visual time-series approaches use plots to expose trends, periodicity, and deviations for anomaly detection through visual reasoning.
- Existing benchmarks do not support jointly localizing anomaly intervals and producing plot-consistent explanations.
3. VisAnomBench Time-Series Anomaly Reasoning Benchmark
VisAnomBench augments public anomaly-detection benchmarks with temporally grounded explanations and structures VLM outputs around interval predictions and reasoning. Candidate explanations are selected using rewards for anomaly accuracy, visual groundedness, axis awareness, and clarity.
- VisAnomBench augments four public benchmarks with natural-language explanations aligned with annotated anomaly intervals.
- Each VLM output contains predicted anomaly intervals A and a numbered, step-by-step explanation E enclosed by structured tags.
- Ground-truth intervals guide explanations because anomaly semantics vary across domains and application contexts.
- The benchmark construction segments series, renders plot images, elicits reasoning traces, and selects candidates with a composite reward.
- The reward combines range-based anomaly-overlap F1 with visual groundedness, axis awareness, and explanation clarity.
4 VISANOMREASONER MODEL
VisAnomBench retains the highest-reward candidate as supervision and spans diverse public time-series data. Its dataset includes separate training and held-out test collections.
- The highest-reward candidate for each time series becomes the supervised fine-tuning target for VisAnomReasoner.
- VisAnomBench draws from four public benchmarks containing real-world and synthetic series across diverse domains, anomaly types, and temporal characteristics.
- The dataset contains 2,576 training time series and 740 held-out test time series, with a strictly disjoint test set.
4. VisAnomReasoner Model
VisAnomReasoner formulates anomaly detection from plots as joint interval localization and explanation generation. It is fine-tuned with explanation-augmented supervision on Qwen2.5-VL base models.
- VisAnomReasoner directly predicts anomaly intervals and generates grounded explanations instead of relying on prompt engineering alone.
- The task defines a univariate series x with ground-truth anomaly intervals and a predicted interval set A.
- Given a time-series plot I(x) and optional context C, the model outputs predicted intervals A and a natural-language explanation E.
- Explanations are grounded in observable visual evidence such as spikes, level shifts, or periodic deviations.
- Supervised fine-tuning of Qwen2.5-VL-3B and Qwen2.5-VL-7B produces the 3B and 7B VisAnomReasoner variants.
5. Experiments
Experiments evaluate VisAnomReasoner on VisAnomBench and TSB-AD-U against VLM, foundation-model, specialized, and classical baselines. The model consistently leads on VisAnomBench and generalizes strongly to TSB-AD-U.
- 5. Experiments: Experiments use the VisAnomBench test set and the TSB-AD-U benchmark, with interval metrics on VisAnomBench and standard plus affiliation metrics on TSB-AD-U.
- 5. Experiments: Baselines span general large VLMs, small VLMs, foundation models, specialized large models, and classical anomaly detectors.
- 5.1.1. Results on VisAnomBench: VisAnomReasoner consistently outperforms the compared baselines across VisAnomBench evaluation metrics.
- 5.1.2. Results on TSB-AD-U: VisAnomReasoner performs best on all six TSB-AD-U metrics against LLaMA-4-Maverick, despite its substantially smaller scale.
- 5.1.2. Results on TSB-AD-U: VisAnomReasoner improves standard precision, recall, and F1 over Qwen2.5-VL-7B by 9.57 pp, 12.06 pp, and 13.39 pp, respectively.
- 5.1.2. Results on TSB-AD-U: Chronos has standard and affiliation precision approximately 30 pp and 20 pp lower than VisAnomReasoner, indicating systematic over-flagging.
6 QUALITATIVE ANALYSIS
VisAnomReasoner consistently outperforms the evaluated baselines, while qualitative examples show tighter localization and more grounded reasoning than competing approaches.
- VisAnomReasoner outperforms all baselines across six evaluation metrics.
- Classical anomaly detectors perform worst overall on TSB-AD-U, reflecting limited robustness to diverse time-series characteristics.
- VisAnomReasoner provides balanced standard and affiliation-metric performance, supporting more reliable event-level anomaly detection.
- VisAnomReasoner produces a tighter ground-truth-aligned interval and coherent reasoning linking amplitude and pattern changes to signal deviations.
7 ABLATION STUDIES
Ablations indicate that supervised fine-tuning reduces false positives, while adding reasoning traces improves both localization metrics and explanation quality.
- Qualitative Examples: VisAnomReasoner produces tighter anomaly intervals and more grounded reasoning than competing models in qualitative examples.
- Supervised Fine-Tuning: 180% and 220% precision gains occur for the 3B and 7B variants, respectively, after supervised fine-tuning.
- Supervised Fine-Tuning: Supervised fine-tuning improves all metrics, while marginal recall gains indicate that its main effect is reducing false positives.
- Reasoning Supervision: Interval-only fine-tuning substantially reduces false positives but does not recover additional missed anomalies.
- Reasoning Supervision: Adding reasoning traces improves both precision and recall beyond interval-only supervision.
8 CONCLUSION
The paper concludes that explanation-augmented supervision enables compact vision-language models to localize anomalies and generate grounded explanations across diverse time-series benchmarks.
- VisAnomReasoner jointly localizes anomaly intervals and generates visual-temporal explanations from time-series plots.
- VisAnomBench combines four public benchmarks spanning diverse domains, anomaly types, sequence lengths, and temporal patterns.
- Candidate reasoning traces are generated by four large VLMs and ranked using rewards for interval accuracy, visual groundedness, axis awareness, and clarity.
- The prompt enforces axis-aware interval localization and concise, visually grounded reasoning under a parsable output schema.
- A judge scores generated reasoning for visual groundedness, axis awareness, and clarity using the time-series plot.
A.1.2. Explanation Quality Evaluation
VisAnomBench explanations are selected for visual grounding and consistency with anomaly intervals, while VisAnomReasoner produces more precise, better-grounded anomaly reasoning than several baselines. Human validation remains limited to a random subset because full-scale expert verification is impractical.
- Validation: Human validation covers only a random subset because VisAnomBench contains thousands of time series across multiple domains.Full-scale verification would require domain experts.
- Explanation selection: Using four VLM generators and a separate judge mitigates concerns that synthetic supervision reflects a single generator’s bias.Training on VisAnomBench consistently outperforms training on single-model-generated data.
- Explanation selection: Preferred explanations provide detailed, axis-aware justifications tied to scale, baseline behavior, and visual deviations supporting selected intervals.Lower-scoring alternatives are more generic or weakly supported by the plot.
- Baseline comparison: Deep-learning baselines achieve substantially lower interval-level precision, recall, and F1 than VisAnomReasoner on VisAnomBench.AE is the strongest deep-learning baseline by F1, while Anomaly-Transformer performs worst in this setting.
- Limitations: VisAnomReasoner is limited to univariate plots and depends on visualizations that expose trends or deviations clearly.Ambiguous or compressed anomalies can produce weakly grounded or hallucinated explanations.