Hazard or Anomaly? Evaluating VLMs for Understanding Dangers and Discrepancies

Technical Summary: Hazard or Anomaly? Evaluating VLMs for Understanding Dangers and Discrepancies

Problem Statement
Modern safety-critical systems, such as those used in search and rescue and industrial safety, increasingly rely on Vision-Language Models (VLMs) to interpret complex scenes and support human decision-making. However, current evaluation frameworks for VLM safety reasoning often frame danger recognition as a binary decision (Safe/Unsafe). This binary approach obscures a critical distinction: the difference between a scene that is merely unusual (an anomaly) and a scene that poses a genuine physical threat (a hazard). The authors argue that without explicitly separating these concepts, it remains unclear whether VLMs are identifying true physical hazards or merely reacting to contextual irregularities. This conflation can lead to "over-reaction," where anomalous but non-dangerous situations are misclassified as hazardous, potentially distracting operators from real risks.

Methodology
To address this limitation, the authors propose an evaluation pipeline that distinguishes between "hazard" and "anomaly" as orthogonal dimensions.

• Definitions:

  • Hazard: A source of potential physical harm under normal interaction conditions (aligned with OSHA definitions).
  • Anomaly: A violation of expected spatial, functional, or contextual relations that does not necessarily imply danger.
  • Classes: The study utilizes a four-class scheme: Safe, Anomalous, Hazardous, and Anomalous-Hazardous.

• Dataset: The authors curated a 610-image benchmark derived from public datasets covering construction, natural disasters, and hazardous objects. Images were generated from textual descriptions and labeled by researchers using a two-bit encoding (Safe, Anomaly, Hazard, Both). Borderline cases were discussed to refine definitions, with images lacking consensus excluded. The final dataset was down-sampled to ensure balanced class representation for testing.

• Experimental Design: The study evaluates several state-of-the-art VLMs (including GPT-4.1, Mistral variants, and Gemini variants) across three prompting strategies:

  1. Zero-Shot: Providing definitions and asking for classification.
  2. Few-Shot: Providing examples of correct classifications for each class.
  3. Chain-of-Thought (CoT): Guiding the model to identify hazards and anomalies independently before making a final decision.

  Additionally, the authors tested two input modalities: raw images and "dense captions" (detailed textual descriptions of the images generated by Gemini 3 Flash) to isolate the model's reasoning capabilities from its visual perception.

• Metrics: Performance was measured using per-class Precision, Recall, and F1 scores for both Anomaly and Hazard detection. Joint classification performance was evaluated using Hamming Loss to assess the error rate in the four-class taxonomy.

Key Results

• Conflation of Concepts: Under binary safety prompting, VLMs frequently misinterpret anomalousness as hazardousness. The study confirms that models often rely on contextual irregularity as a proxy for danger.
• Performance Variance:
  • Hazard vs. Anomaly: Models consistently performed better at classifying hazards (F1 scores often >0.75) than anomalies. The concept of "anomaly" proved more difficult to define and classify, likely due to its semantic ambiguity.
  • Input Modality: Contrary to the hypothesis that dense captions would isolate reasoning, models generally performed worse or similarly with dense captions compared to raw images. The authors suggest captions may lack the necessary contextual focus or introduce noise.
  • Prompting Strategies: No single prompting technique was universally superior. While Few-Shot prompting often yielded the best results for hazard classification, Chain-of-Thought or Few-Shot strategies varied by model for anomaly classification.
  • Model Size: Larger models (e.g., GPT-4.1, Gemini 3 Flash) generally outperformed smaller models (e.g., Mistral 3B), particularly in joint classification tasks. The best joint classification performance (lowest Hamming Loss of 0.23) was achieved by Gemini 3 Flash using Few-Shot prompting with image inputs.
• Baseline Comparison: When compared to a prior binary safety study (Choi et al. [5]), the authors' Zero-Shot prompt (utilizing the four-class definitions) resulted in significantly higher accuracy (p-value 0.0022) even when results were collapsed to binary categories. This suggests that forcing models to reason through the four-class taxonomy improves their safety judgments, even if the final output is binary.

Significance and Claims
The paper claims that explicitly separating anomaly from hazard provides a more informative evaluation of VLM safety reasoning than traditional binary approaches. By exposing the failure mode where models conflate "unusual" with "dangerous," the study highlights a specific area where current VLMs lack reliable situational awareness.

The authors conclude that while state-of-the-art VLMs show moderate capability in distinguishing these states (with F1 scores around 0.70 for anomalies and up to 0.85 for hazards), they still struggle with the complex reasoning required for safety-critical four-class taxonomies. The study argues that current models are not yet ready for fully autonomous deployment in high-risk environments without further refinement in reasoning capabilities and prompt engineering. The work provides a public dataset and a structured evaluation pipeline to facilitate future research into making AI systems more consistent in distinguishing unusual conditions from genuine dangers.