AccidentBench: Benchmarking Multimodal Understanding and Reasoning in Vehicle Accidents and Beyond
AccidentBench: Benchmarking Multimodal Understanding and Reasoning in Vehicle Accidents and Beyond [101.2] AccidentBenchは、自動車事故シナリオとBeyondドメインを組み合わせた大規模なベンチマークである。 このベンチマークには、約2000のビデオと19000以上の人間による質問応答ペアが含まれている。 論文参考訳（メタデータ） (Tue, 30 Sep 2025 17:59:13 GMT)
事故シナリオのベンチマーク、「AccidentBench targets understanding and reasoning across diverse vehicle accident scenarios (83.0%), while also encompassing airspace (10.2%) and waterway (6.8%) domains, in which safety, perception, and decision-making are deeply interdependent. Unlike benchmarks that emphasize isolated skills or single domains, AccidentBench systematically challenges models across several critical understanding and reasoning capabilities: temporal understanding and reasoning (tracking event sequences and causality over extended periods); spatial understanding and reasoning (understanding dynamic spatial relationships and multi-agent trajectories); and intent and goal reasoning (inferring agent intentions and planning goals), which further includes complex strategic and counterfactual reasoning (evaluating higher-order strategies, action implications, and “what-if” scenarios).」