Where Does the Answer Come From? Benchmarking View-Level Visual Evidence Identification in Multi-View MLLMs for Autonomous Driving
2026-06-08 • Computation and Language
Computation and LanguageComputer Vision and Pattern Recognition
AI summaryⓘ
The authors created a new test to check if multimodal language models used the right camera view when answering questions about driving scenes. Instead of just checking if the answer is correct, their benchmark also checks if the model points to the correct camera among six views from a driving dataset. They included questions that involve tricky reasoning like causes and intentions. Their tests show that models can give correct answers but sometimes rely on the wrong visual source, which regular accuracy scores don't catch. This helps better evaluate how well models understand and use visual information.
multimodal large language modelsvisual question answeringmulti-view camerasNuScenes datasetevidence groundingcausal reasoningcounterfactual reasoningintent predictionbenchmark evaluationcamera view selection
Authors
Yimu Wang, Yee Man Choi, Barry Zhang, Mozhgan Nasr Azadani, Sean Sedwards, Krzysztof Czarnecki
Abstract
Multimodal large language models (MLLMs) achieve strong results on visual reasoning benchmarks, but answer accuracy alone does not indicate whether a model relied on the correct visual evidence. This gap is particularly important in multi-view driving scenes used for autonomous driving, where a model can produce a plausible answer while grounding it in the wrong camera view. We introduce a multi-view visual question answering benchmark for evaluating evidence-source identification: given six synchronized NuScenes views and a question, the model must identify the supporting camera view and answer the question. The benchmark contains 122 conflict-centric question-answer pairs from 73 scenes, spanning causality, counterfactual reasoning, and intent prediction. View labels are proposed by an automatic conflict-mining pipeline and manually verified by annotators. We evaluate three settings: camera-view selection, oracle QA given the golden view, and joint prediction in which the model selects a view and answers in one pass. Answers are evaluated in both multiple-choice and free-form formats, using exact match for structured predictions and an LLM judge for free-form responses. By explicitly separating visual-source identification from answer correctness, the benchmark exposes grounding failures that answer-only evaluation misses.