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

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

Problem Statement
Detecting abnormal patterns in time-series data is critical for applications in industrial monitoring, healthcare, and cyber-physical systems. While recent Vision-Language Models (VLMs) have demonstrated strong multimodal reasoning capabilities in domains like image captioning and chart understanding, their application to time-series anomaly detection remains unsatisfactory. Existing approaches face three primary challenges:

1. Lack of Interpretability: Traditional methods output numeric scores or binary labels without explaining why a region is anomalous, limiting their utility for diagnosing failure modes.
2. Data Representation Mismatch: Time-series plots lack explicit object boundaries and spatial structures found in natural images, requiring models to reason over temporal patterns, periodicity, and deviations rather than discrete visual entities. General-purpose VLMs often produce coarse, unstable predictions when applied "out of the box."
3. Supervision Gap: Public anomaly detection benchmarks typically provide only interval or point-wise labels, lacking natural-language rationales. This absence of explanatory supervision hinders the fine-tuning of VLMs to produce grounded, interpretable decisions.

Methodology
The authors propose a framework that reframes time-series anomaly detection as a plot-grounded vision-language reasoning task, jointly requiring interval-level localization and structured explanation generation.

1. VisAnomBench (Benchmark Construction):
   To address the lack of explanatory supervision, the authors constructed VisAnomBench, the first explanation-augmented benchmark for vision-language time-series anomaly reasoning. It aggregates data from four public datasets (KPI, GutenTAG, UCR-EGI, and UCR-TSAD) covering diverse domains and anomaly types.

   • Data Augmentation: For each time-series plot, the benchmark pairs annotated anomaly intervals with high-quality, temporally grounded natural-language explanations.
   • Reward-Guided Selection: Since public datasets lack explanations, the authors used multiple large VLMs to generate candidate reasoning traces. A composite reward function selected the highest-quality candidates based on four criteria:
     • Anomaly Accuracy: Temporal overlap with ground-truth intervals.
     • Visual Groundedness: Explicit reference to observable plot patterns (e.g., spikes, level shifts).
     • Axis Awareness: Consistency of timestamps and values with the plot axes.
     • Clarity: Logical ordering and coherence of the reasoning.

2. VisAnomReasoner (Model):
   The authors developed VisAnomReasoner, a parameter-efficient VLM fine-tuned on VisAnomBench.

   • Architecture: Built upon the Qwen2.5-VL base models (3B and 7B parameters).
   • Training Objective: The model is supervised to generate structured outputs containing predicted anomaly intervals (enclosed in <anomaly> and <index> tags) and step-by-step reasoning traces (enclosed in <thought> tags).
   • Efficiency: Unlike prior approaches relying on extensive prompt engineering or proprietary models (e.g., GPT-4o), VisAnomReasoner is a compact, fine-tuned model designed to directly predict intervals and explanations from plot images.

Key Contributions

1. Task Formulation: The paper formulates anomaly detection as a unified vision-language reasoning task that evaluates a model's ability to align visual evidence, temporal boundaries, and natural-language reasoning in a single prediction.
2. VisAnomBench: The introduction of a benchmark that pairs anomaly intervals with temporally grounded rationales, enabling supervised learning for both localization and explanation generation.
3. VisAnomReasoner: A compact VLM that, despite its small scale, outperforms general-purpose VLMs, specialized large models, time-series foundation models, and classical detectors.

Experimental Results
The authors evaluated VisAnomReasoner on VisAnomBench and the TSB-AD-U benchmark against 15 baselines, including general large VLMs (e.g., Grok-4-Fast, LLaMA 4 Maverick), small VLMs, foundation models (TimesFM, Chronos), specialized LLM/VLM detectors, and classical methods (e.g., Isolation Forest).

• Performance on VisAnomBench: VisAnomReasoner (7B variant) achieved a Precision of 72.09% and an F1 score of 73.94%. This represents improvements of at least 23.17 percentage points in Precision and 25.64 percentage points in F1 over the strongest baselines. It also achieved the highest Overlap score (25.35%), indicating superior temporal boundary localization.
• Cross-Benchmark Generalization (TSB-AD-U): The model demonstrated strong generalization to unseen data characteristics, improving Precision by 9.57 and F1 by 13.39 percentage points over the second-best baseline (LLaMA 4 Maverick).
• Ablation Studies:
  • Supervised Fine-Tuning (SFT): SFT significantly reduced False Positives (FPs), improving Precision by 180% (3B) and 220% (7B) over base models.
  • Reasoning Supervision: Adding reasoning traces during fine-tuning improved both Precision and Recall compared to interval-only supervision, confirming that the gains stem from explanation-augmented supervision rather than task-specific tuning alone.
  • Explanation Quality: In a paired comparison, GPT-4o preferred VisAnomReasoner's explanations over the base model's in 69.6% of cases, citing better visual groundedness and clarity.

Significance
The paper argues that explanation-augmented supervision is a practical path toward interpretable, reasoning-centric time-series anomaly detection. By demonstrating that a compact, fine-tuned VLM can outperform much larger general-purpose models and specialized detectors, the work challenges the notion that large-scale proprietary models are necessary for high-fidelity anomaly detection. The results suggest that grounding models in visual evidence and requiring structured reasoning traces significantly enhances both detection accuracy and the reliability of temporal localization, addressing the critical need for interpretable decisions in real-world monitoring systems.