Technical Summary: RiskCueBench

Problem Statement

The rapid proliferation of video-centric social media has created a critical need for systems capable of anticipating risky events from visual data to ensure public safety. While advanced Vision-Language Models (VLMs) have demonstrated strong capabilities in general video understanding, existing benchmarks largely focus on post hoc analysis—describing events after they have fully unfolded. This approach fails to reflect real-world safety-critical scenarios where models must predict future risks based on partial, ambiguous, and early visual signals. Furthermore, current datasets often provide access to the full video sequence, including the accident itself, which artificially reduces task difficulty and masks a model's inability to perform genuine anticipatory reasoning.

Methodology: RiskCueBench Construction

To address these gaps, the authors introduce RiskCueBench, a benchmark designed to evaluate VLMs' ability to infer potential danger from early risk cues. The construction of the benchmark follows a four-stage, reproducible, and domain-agnostic pipeline:

1. Data Collection: The authors curated a large-scale candidate set of YouTube videos using structured queries derived from real-world events across diverse regions and time periods (specifically focusing on protests and traffic incidents). Queries were generated using GPT-4V for year-location pairs and verified via manual search to ensure news coverage.
2. Dataset Filtering: An automated filtering process retained only videos depicting active risk events with high visual quality.
   • Text-based Relevance: An LLM (GPT-4o) scored videos based on the likelihood of containing in-action risk events versus commentary or summaries.
   • Visual Quality: A model (Gemini 2.5 Flash Lite) evaluated 12 visual dimensions (e.g., absence of overlays, subtitles, or heavy editing) to ensure models rely on authentic visual perception rather than spurious shortcuts.
   • Hard Example Mining: Five VLMs were used to identify ambiguous cases. Videos where more than three models disagreed on the presence of a clear risk event were retained as "hard examples" to test model resilience in uncertain conditions.
3. Annotation: Trained human annotators labeled the filtered videos using a temporal protocol. Key annotations included:
   • Risk Signal Start/End: The earliest moment a risk signal appears and when it ceases.
   • Visual Indicators: Specific cues used to judge risk.
   • Accident Start/End: Timestamps for the actual incident.
   • Risk Label: Binary classification ("yes" for clear risk, "no" for normal activity).
4. Evaluation Framework: The benchmark employs a Question-Reasoning-Answer (QRA) framework. Models are presented only with the "risk signal clip" (the segment before the accident) and must output a prediction and reasoning trace.

Evaluation Metrics

Beyond standard binary classification (F1 score), the authors propose specific metrics to analyze reasoning quality:

• Reasoning Grounding Accuracy (RGA): Measures the semantic alignment between the model's extracted decision items and human-annotated visual indicators using sentence embeddings.
• Temporal Reasoning Difference (TRD): Evaluates sensitivity to temporal ordering by shuffling, swapping, or reversing video frames. A low TRD indicates reliance on static cues rather than temporal dynamics.
• Self-Correction Degradation (SCD): Quantifies the impact of model hesitation (self-correction markers like "wait..." or "actually...") on prediction accuracy.

Key Contributions

1. RiskCueBench Benchmark: A challenging dataset requiring VLMs to infer danger from subtle, early visual cues in two distinct domains: crowd dynamics (protests) and vehicular incidents (car crashes).
2. Automated Curation Pipeline: An efficient workflow leveraging model disagreement and LLM filtering to automatically identify difficult cases and construct a high-quality, temporally grounded dataset.
3. Comprehensive Reasoning Analysis: The introduction of custom metrics (RGA, TRD, SCD) to dissect model behavior, moving beyond simple accuracy to evaluate the faithfulness, grounding, and temporal coherence of model reasoning.

Experimental Results

The authors evaluated 16 state-of-the-art models, including standard VLMs and reasoning-enhanced variants (e.g., Qwen-VL, MiMo-RL, GLM-4.1V-Thinking).

• Performance Gap: Models achieved moderate success on car crash prediction (F1: 0.27–0.71) but struggled significantly with protest scenarios (F1: 0.26–0.46), highlighting a difficulty in interpreting nuanced social and behavioral cues compared to structured vehicular patterns.
• Reasoning Grounding: Correct predictions exhibited significantly higher grounding scores than incorrect ones. Incorrect predictions contained approximately 50% more ungrounded decision items, indicating that failure to reference relevant visual objects is a primary cause of error.
• The "Overthinking" Phenomenon: Contrary to expectations, self-correction behaviors consistently degraded performance by 15–26 percentage points. Models often overrode correct initial judgments with speculative reasoning or failed to correct initial misperceptions.
• Lack of Temporal Reasoning: The TRD metric revealed that models perform similarly on original, shuffled, and reversed videos. This suggests VLMs rely heavily on static frame content rather than genuine temporal progression or causal ordering.
• Temporal Distance Sensitivity: As the temporal gap between the risk signal and the incident increased (1 to 20 seconds), model accuracy systematically declined.

Significance and Claims

The paper claims that RiskCueBench exposes a critical limitation in current VLMs: their inability to perform anticipatory reasoning from early signals. The results indicate that:

• State-of-the-art models lag significantly behind human performance, particularly in socially embedded scenarios.
• Current "reasoning-enhanced" models do not consistently outperform standard VLMs in this domain, suggesting that explicit reasoning mechanisms are not yet effectively calibrated for early risk detection.
• The field requires a shift from retrospective analysis to predictive reasoning that is temporally coherent and visually grounded.

The authors position RiskCueBench as a foundational tool for evaluating and improving VLMs in safety-critical applications, emphasizing that accurate risk prediction requires more than just pattern recognition; it demands the ability to interpret evolving situations from incomplete information.