Research map / Vision-language and multimodal / Multimodal LLMs

MVVBench: Benchmarking 4D Reasoning in Vision-Language Models

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MVVBench, a human-authored benchmark of 1,323 multiple-choice questions over 608 multi-camera scenes testing 4D (view + time) reasoning in VLMs, plus inference-time methods (MV-Guided CoT, TADSG) and a preliminary RLVR study.

Problem

Existing video and multi-view benchmarks test temporal or spatial reasoning in isolation, and VLM-generated QA caps difficulty; there is no rigorous benchmark for joint multi-view and temporal reasoning where questions are unanswerable from any single view.

Approach

Human-authored QA under a monocular-ambiguity constraint (unanswerable from any single selected view), converted to multiple choice via a generator-judge VLM loop, with structured multi-view prompting and a tool-augmented, confidence-gated query-conditioned dynamic scene graph for cross-view identity.

Key results

• Best model GPT-5-mini reaches 54.20% overall accuracy versus 80.27% human performance; random baseline is 35.81%.
• MV-Guided CoT raises Qwen3-VL-4B-Instruct from 42.78% to 44.82%, while naive CoT slightly lowers it to 42.55%.
• RLVR with DAPO and LoRA on Qwen3-VL-4B-Instruct raises held-out accuracy from 43.67% to 47.67%.

Surprising findings

• Naive chain-of-thought prompting slightly lowers accuracy on multi-view questions, and the thinking variant underperforms the instruct model.
• Providing all videos rather than the selected views barely changes accuracy, and text-only input stays near chance.

Limitations

• Focuses on human-centric scenes from source datasets, limiting generalization to other object types.
• TADSG depends on person detection and re-identification, restricting it to scenes with people.
• Fixed set of six categories may miss fine-grained action or causal reasoning across views.

Experimental setup

Datasets: Ego-Exo4D Panoptic Studio MMPTRACK MVVBench

Benchmarks: MVVBench

Metrics: accuracy

Baselines: Gemini-2.5-pro GPT-5-mini LLaVA-OneVision-72B MiniCPM-o-2.6-8B Qwen3-VL-4B-Instruct CoT M3CoT

Builds on: chain-of-thought prompting DAPO VMCBench M3CoT YOLOv8 OSNet

Closest papers by meaning

• MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.90)
• Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs (AAAI 2026 · similarity 0.89)
• SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.89)
• VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.89)
• Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events (CVPR 2025 · similarity 0.89)
• TOMATO: Assessing Visual Temporal Reasoning Capabilities in Multimodal Foundation Models (ICLR 2025 · similarity 0.89)