Technical Summary: Where Does the Answer Come From? Benchmarking View-Level Visual Evidence Identification in Multi-View MLLMs for Autonomous Driving

Problem Statement

Multimodal Large Language Models (MLLMs) have demonstrated strong performance on visual reasoning benchmarks. However, standard evaluation metrics focus almost exclusively on final answer correctness, failing to verify whether the model's reasoning is grounded in the correct visual evidence. This limitation is particularly critical in autonomous driving scenarios involving multi-view inputs (e.g., surround-view cameras). In these settings, a model may generate a plausible or even correct answer by relying on language priors, dataset shortcuts, or hallucinated evidence, while incorrectly attributing that answer to a specific camera view. Such "wrong-view/correct-answer" behaviors undermine trust in safety-critical systems, as reliable reasoning requires not only a correct conclusion but also a correct identification of the supporting visual source.

Methodology

Benchmark Design

The authors introduce a diagnostic benchmark for evaluating view-level visual evidence identification in multi-view MLLMs. The benchmark is built upon the NuScenes dataset, utilizing six synchronized camera views (Front, Front Left, Front Right, Back, Back Left, Back Right) per scene.

• Dataset Construction: The dataset comprises 122 conflict-centric question-answer (QA) pairs derived from 73 scenes. The construction pipeline involves:

  1. Conflict Mining: Automatically mining candidate events (pedestrian crossing, braking, cut-in, lane change, turning, and Car-to-Car Front Turn-Across-Path) from kinematic and geometric features.
  2. View Projection: Projecting 3D bounding boxes of involved actors onto the six camera planes to identify visibility.
  3. QA Generation: Creating questions targeting three reasoning types: causality, counterfactual reasoning, and intent prediction.
  4. Annotation: Each QA pair is associated with a manually verified "golden view" (the single camera channel providing the most direct evidence) or labeled as NONE if no view provides sufficient evidence. Ambiguous cases are excluded.

• Evaluation Settings: The benchmark evaluates models under three distinct protocols to isolate different failure modes:

  1. View Selection: The model receives all six views and must predict the supporting camera view (or NONE).
  2. Oracle QA: The model receives only the golden view image and must answer the question. This isolates reasoning capability when the correct visual source is provided.
  3. Joint Prediction: The model receives all six views and must simultaneously predict the supporting view and the answer in a single pass.

• Metrics:

  • View Accuracy: Exact match against the golden view label.
  • Answer Accuracy: Exact match for multiple-choice (MC) questions; semantic equivalence judged by an LLM for free-form responses.
  • Joint Success: Strict accuracy requiring both the view selection and the answer to be correct.

Experimental Setup

The authors evaluated a range of proprietary (GPT-5.4, Gemini, Claude) and open-source (Qwen2.5-VL, Qwen3-VL, InternVL3) MLLMs. All models were tested in a zero-shot setting using a unified prompting protocol without task-specific fine-tuning.

Key Results

• View Selection Performance: Proprietary models demonstrated superior view identification capabilities, with Claude achieving the highest mean accuracy (82.62%), followed by GPT-5.4 (77.54%) and Gemini (74.10%). Open-source models showed significant variance; Qwen3-VL and InternVL3 clustered around 61.5%, while Qwen2.5-VL dropped to 12.62%, performing near a random baseline.
• The Oracle-to-Joint Gap: A consistent performance drop was observed when moving from Oracle QA (where the correct view is provided) to Joint Prediction. For example, Claude's MC accuracy dropped from 89.3% (Oracle) to 73.2% (Joint). This indicates that view identification is a primary bottleneck for end-to-end grounded answering; models often fail to locate the evidence before attempting to reason.
• Wrong-View/Correct-Answer Phenomenon: Analysis revealed that a significant portion of correct answers were generated with incorrect view selections. Across all models, 192 out of 732 multiple-choice trials resulted in a correct answer despite selecting the wrong view. This suggests that answer-only evaluation metrics substantially overestimate the grounded reasoning capabilities of these models.
• Regional and Reasoning Biases: Models performed better on front and rear views than on side views, where adjacent-camera ambiguity is higher. Additionally, view identification accuracy dropped notably for intent prediction questions compared to causality or counterfactual questions, suggesting that forecasting future behavior requires more subtle evidence localization.

Significance and Claims

The paper claims that its primary contribution is shifting the evaluation paradigm from answer plausibility to evidence-grounded reasoning. By explicitly separating visual-source identification from answer correctness, the benchmark exposes grounding failures that traditional benchmarks miss.

• Diagnostic Utility: The authors position the benchmark as a "controlled diagnostic probe" rather than a large-scale training resource or a leaderboard for saturated performance. Its compact size (122 QA pairs) is intentional, designed to test whether existing MLLMs can identify evidence sources under realistic multi-view prompting.
• Safety Implications: In autonomous driving, trustworthy reasoning requires not just a correct decision but a correct attribution of why that decision was made (i.e., which sensor view supports it). The benchmark highlights that models may "hallucinate" evidence attribution, a critical failure mode for safety-critical systems.
• Limitations: The authors acknowledge that the benchmark is limited to camera inputs (excluding LiDAR/radar), focuses on conflict-centric scenarios rather than general driving, and relies on an LLM judge for free-form evaluation. They caution that because the benchmark is derived from NuScenes, models may have encountered related data during pretraining, potentially biasing absolute performance estimates.

In summary, the paper demonstrates that while MLLMs can often answer driving-related questions correctly, they frequently fail to identify the specific camera view that justifies their answers, a gap that standard evaluation methods fail to detect.