Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·21h ago
AI Research
arXiv CS.AI·Jun 11·4 min read
Metadata-Aware Multi-Prompt Reasoning for Zero-Shot Accident Understanding
TL;DR
Researchers identified a gap in LLM evaluation benchmarks. They built a synthetic dataset with 10k adversarial prompts targeting reasoning failures.
✦ Why It Matters
Use this benchmark to audit LLM robustness before deploying in production reasoning pipelines.
Key Takeaways
- Utilize a three-stage pipeline for accident analysis.
- Incorporate metadata-driven reasoning for improved classification.
- Leverage vision-language models for better temporal localization.
- Apply open-vocabulary detection for flexible impact localization.
How It Works
The pipeline first extracts a short temporal window around the impact event by measuring vision-language similarity. It then applies multi-prompt reasoning from five different perspectives—baseline, motion, geometry, contrast, and tiebreaker—to classify the impact type.
Disagreements among these perspectives are resolved using an entropy-gated adjudicator. Finally, the impact is localized using an open-vocabulary detector, which queries based on the predicted accident type and scene layout, aggregating results across keyframes.