Research map / Vision-language and multimodal / Multimodal LLMs

What’s in Common? Multimodal Models Hallucinate When Reasoning Across Scenes

NeurIPS 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

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 Common-O Bench and Common-O Complex, multi-image benchmarks of in-the-wild and synthetic scenes testing whether multimodal models can identify objects common to two scenes, finding low accuracy and pervasive hallucination.

Problem

Multimodal models saturate perception benchmarks but struggle to reason across multiple real-world scenes and tend to hallucinate objects not present, with benchmarks suffering from web-data contamination and simplified visuals.

Approach

A contamination-free benchmark of two-image scenes (real photos taken by authors plus Unreal Engine synthetic renders) where models must select objects present in both images, with a single-image perception control and a hallucination-rate metric.

Key results

• GPT-4o achieves 35.11% accuracy on Common-O Bench multi-image reasoning versus 77.25% on single-image perception.
• On Common-O Complex (8-16 objects), no model exceeds ~1% accuracy, with average hallucination of 76% for 1+ objects.
• Models trained with multi-image inputs achieve roughly 3x higher accuracy on Common-O Bench than single-image-trained models.

Measured effects reported in the paper

Surprising findings

• Chain-of-thought boosts single-image perception but has a mixed effect on cross-scene reasoning.
• Llama 4 Scout is the exception where hallucination did not rise in the multi-image setting.

Limitations

• Real images were all taken by the authors, possibly reflecting bias in locations, backgrounds, and objects.
• Multiple-choice setup is brittle to prompt and option-order changes; only English text is included.
• Only two-image scenes with up to 16 objects; open-ended generation not evaluated.

Experimental setup

Datasets: Aria Digital Twin Catalog

Benchmarks: Common-O Bench Common-O Complex

Metrics: accuracy hallucination rate

Baselines: GPT-4o LLaVA-OneVision Qwen2.5-VL DeepSeek-VL2 LlamaV-o1 Llama 3.2 Instruct Llama 4 Scout PerceptionLM QVQ-72B-Preview

Builds on: MMIU Aria Digital Twin Catalog Segment Anything

Closest papers by meaning

• Multi-Object Hallucination in Vision Language Models (NeurIPS 2024 · similarity 0.91)
• Fine-Grained Multi Image Object Hallucination Benchmark (CVPR 2026 · similarity 0.91)
• More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models (NeurIPS 2025 · similarity 0.90)
• Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.90)
• Grounded Chain-of-Thought for Multimodal Large Language Models (CVPR 2026 · similarity 0.89)
• When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding (NeurIPS 2025 · similarity 0.89)