MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning

Jiachun Li; Shaoping Huang; Zhuoran Jin; Chenlong Zhang; Pengfei Cao; Yubo Chen; Kang Liu; Jun Zhao

cs.CLMar 2, 2026

MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning

Authors: Jiachun Li, Shaoping Huang, Zhuoran Jin, Chenlong Zhang, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao

Organizations: School of Artificial Intelligence, University of Chinese Academy of Sciences · The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences · School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences

Abstract

Recent progress in the reasoning capabilities of multimodal large language models (MLLMs) has empowered them to address more complex tasks such as scientific analysis and mathematical reasoning. Despite their promise, MLLMs' reasoning abilities across different scenarios in real life remain largely unexplored and lack standardized benchmarks for evaluation. To address this gap, we introduce MMR-Life, a comprehensive benchmark designed to evaluate the diverse multimodal multi-image reasoning capabilities of MLLMs across real-life scenarios. MMR-Life consists of 2,646 multiple-choice questions based on 19,108 images primarily sourced from real-world contexts, comprehensively covering seven reasoning types: abductive, analogical, causal, deductive, inductive, spatial, and temporal. Unlike existing reasoning benchmarks, MMR-Life does not rely on domain-specific expertise but instead requires models to integrate information across multiple images and apply diverse reasoning abilities. The evaluation of 37 advanced models highlights the substantial challenge posed by MMR-Life. Even top models like GPT-5 achieve only 58% accuracy and display considerable variance in performance across reasoning types. Moreover, we analyze the reasoning paradigms of existing MLLMs, exploring how factors such as thinking length, reasoning method, and reasoning type affect their performance. In summary, MMR-Life establishes a comprehensive foundation for evaluating, analyzing, and improving the next generation of multimodal reasoning systems.

Figures & tables

Figure 1: Overview of the MMR-Life . Left: 7 reasoning types and 21 tasks. Middle: A typical example of multi-image reasoning in real-life scenarios. Right: Extensive evaluation reveals a gap between humans and SOTA MLLMs on some real-life reasoning tasks.

Statistics	Number
Total Questions	2,646
Total Reasoning Types/Tasks	7/21
Image Types	15
Reasoning Type
- Abductive Reasoning	307 (11.60%)
- Analogical Reasoning	568 (21.47%)

Table 1: Key statistics of MMR-Life.

Figure 2: MMR-Life examples from each reasoning type.

Dataset	Size	Images	Reason	Source	Knowledge	Avg Img.#
MME-Reasoning ( Yuan et al., 2025 )	1.2K	Symbolic	3 Types	W+T+A+E	Low	1
VisualPuzzles ( Song et al., 2025 )	1.1K	Symbolic	5 Types	W+T+A	Low	1
MMMU ( Yue et al., 2024 )	11.5K	Hybrid	-	W+T+A	High	1.05
MMLU-Reason ( Tie et al., 2025 )	1.1K	Hybrid	6 Types	W+T+E	Medium	1.85
MEGA-Bench ( Chen et al., 2025 )	8K	Hybrid	-	A	Medium	2
MME-COT ( Jiang et al., 2025 )	1.1K	Hybrid	6 Types	E	Medium	2.10

Table 2: The comparison between MMR-Life and other existing benchmarks. W (Web), T (Textbook), A (Annotated), E (Existing datasets), and Avg Img.# (average image counts each question).

Closed-source & Thinking
Model	Abd	Ana	Cau	Ded	Ind	Spa	Tem	Avg
Human	79.76	57.65	75.00	70.59	63.41	79.76	79.76	72.28
GPT-5	53.75	78.87	41.06	80.14	78.32	17.25	41.70	58.69
Gemini-2.5-Pro	54.40	73.77	36.99	79.43	73.66	25.10	35.79	56.86
Gemini-2.5-Flash	46.25	75.18	34.22	71.63	73.66	23.92	30.81	53.10
o4-mini	41.37	73.59	27.38	71.28	68.07	19.22	32.66	50.49

Table 3: Performance comparison of SOTA MLLMs on MMR-Life. The highest and lowest scores for each model type across reasoning types are highlighted in green and red , respectively. The highest performance achieved by the model in each type is indicated in bold .

Panel of Figure 5 — Figure 5: Performance comparison under different thinking budgets.

Model	Method	Abd	Ana	Cau	Ded	Ind	Spa	Tem	Avg ( $\Delta$ )
Qwen2.5-VL-7B	CoT	26.06	35.74	20.53	20.92	38.93	9.41	12.18	24.68
	SC@8	28.01	39.44	23.57	25.18	45.45	10.98	13.10	27.85 (+3.17)
	BoN@8	27.64	44.72	22.81	25.53	48.02	13.33	13.10	29.54 (+4.86)
	GRPO	30.62	40.49	21.29	28.72	43.59	13.73	11.81	28.23 (+3.55)
Qwen2.5-VL-32B	CoT	23.45	42.78	21.29	50.00	27.27	15.69	16.24	28.61
	SC@8	26.06	45.42	23.95	51.77	28.67	16.47	17.90	30.57 (+1.96)

Table 4: Performance across different methods. Scores higher and lower than the base model’s CoT performance are marked in green and red . The highest score in each column is in bold .

Panel of Figure 8 — Figure 8: Error distribution over 140 errors for each model on MMR-Life.

Appendix figures & tables55 assets

Supplementary material from the paper’s appendix.

Appendix

Reasoning Type	Task Type	Image Type	Data Source
Abductive	Human Activity Attribution	Domestic Life	TVbench ( Cores et al., 2025 )
	Character Interaction Attribution	Human Animation	Tom & Jerry Cartoon ( Parmar et al., 2024 )
	Multi-Hop Collision Attribution	Physical Phenomenon	CLEVRER-Humans ( Mao et al., 2022 )
Analogical	Animal Relation Inference	Natural Creatures	Kaggle ( sanadalali, 2025 )
	Product Similarity Inference	Product Shots	Kaggle ( Gégénava, 2025 )
	Artwork Style Inference	Human Artwork	Kaggle ( ikarus777, 2019 )

Appendix

Table 5: Data sources and image types of different tasks in MMR-Life

Model	Abd	Ana	Cau	Ded	Ind	Spa	Tem	Avg
gpt-5	53.75	78.87	41.06	80.14	78.32	17.25	41.70	58.69
gemini-2.5-pro	54.40	73.77	36.99	79.43	73.66	25.10	35.79	56.86
gemini-2.5-flash	46.25	75.18	34.22	71.63	73.66	23.92	30.81	53.10
o4-mini	41.37	73.59	27.38	71.28	68.07	19.22	32.66	50.49
gpt-5-mini	44.95	69.72	32.32	75.18	68.76	12.16	29.52	49.77
gpt-4.1	44.30	71.30	22.43	67.38	70.16	13.73	27.31	48.15

Appendix

Table 6: Full performance comparison of SOTA MLLMs on MMR-Life.

Model	Abd	Ana	Cau	Ded	Ind	Spa	Tem	Avg
Human	79.76	57.65	75.00	70.59	63.41	79.76	79.76	72.28
gpt-5	63.33	66.67	53.33	73.33	66.67	30.00	50.00	57.62
gemini-2.5-pro	63.33	80.00	60.00	76.67	50.00	30.00	36.67	56.67
gemini-2.5-flash	53.33	63.33	50.00	53.33	50.00	50.00	53.33	53.33
o4-mini	46.67	73.33	43.33	73.33	50.00	23.33	43.33	50.48
gpt-5-mini	50.00	56.67	40.00	83.33	50.00	6.67	33.33	45.71

Appendix

Table 7: Performance comparison of SOTA MLLMs on MMR-Life mini set.

Figure 9: Image type distribution in MMR-Life.

Figure 10: Image counts distributions in MMR-Life.

Figure 11: Negative option generation prompt.

Figure 12: Negative option generation prompt.

Figure 13: Negative option generation prompt.

Figure 14: Negative option generation prompt.

Figure 15: Negative option generation prompt.

Figure 16: Negative option generation prompt.

Figure 17: Negative option generation prompt.

Figure 18: Prompt used in the main experiment.

Figure 19: A correct example of Human Activity Attribution task.

Figure 20: An error example of Human Activity Attribution task.

Figure 21: A correct example of Character Interaction Attribution task.

Figure 22: An error example of Character Interaction Attribution task.

Figure 23: A correct example of Multi-Hop Collision Attribution task.

Figure 24: An error example of Multi-Hop Collision Attribution task.

Figure 25: A correct example of Animal Relation Inference task.

Figure 26: An error example of Animal Relation Inference task.

Figure 27: A correct example of Product Similarity Inference task.

Figure 28: An error example of Product Similarity Inference task.

Figure 29: A correct example of Artwork Style Inference task.

Figure 30: An error example of Artwork Style Inference task.

Figure 31: A correct example of Character Interaction Prediction task.

Figure 32: An error example of Character Interaction Prediction task.

Figure 33: A correct example of Multi-Hop Collision Prediction task.

Figure 34: An error example of Multi-Hop Collision Prediction task.

Figure 35: A correct example of Counterfactual Fluid Prediction task.

Figure 36: An error example of Counterfactual Fluid Prediction task.

Figure 37: A correct example of Material Composition Deduction task.

Figure 38: An error example of Material Composition Deduction task.

Figure 39: A correct example of Card Winner Deduction task.

Figure 40: An error example of Card Winner Deduction task.

Figure 41: A correct example of Recipe Step Deduction task.

Figure 42: An error example of Recipe Step Deduction task.

Figure 43: A correct example of Bird Migration Induction task.

Figure 44: An error example of Bird Migration Induction task.

Figure 45: A correct example of Plant Disease Induction task.

Figure 46: An error example of Plant Disease Induction task.

Figure 47: A correct example of Sport Feature Induction task.

Figure 48: An error example of Sport Feature Induction task.

Figure 49: A correct example of Relative Position Estimation task.

Figure 50: An error example of Relative Position Estimation task.

Figure 51: A correct example of Camera Rotation Estimation task.

Figure 52: An error example of Camera Rotation Estimation task.

Figure 53: A correct example of Navigation Route Planning task.

Figure 54: An error example of Navigation Route Planning task.

Figure 55: A correct example of Crowd Timeline Reconstruction task.

Figure 56: An error example of Crowd Timeline Reconstruction task.

Figure 57: A correct example of Driving Sequence Prediction task.

Figure 58: An error example of Driving Sequence Prediction task.

Figure 59: A correct example of Human Activity Localization task.

Figure 60: An error example of Human Activity Localization task.

Multi-Image Reasoning

See all 15 papers

May 11, 2026

SciVQR: A Multidisciplinary Multimodal Benchmark for Advanced Scientific Reasoning Evaluation

Scientific reasoning is a key aspect of human intelligence, requiring the integration of multimodal inputs, domain expertise, and multi-step inference across various subjects. Existing benchmarks for multimodal large language models (MLLMs) often fail to capture the complexity and traceability of reasoning processes necessary for rigorous evaluation. To fill this gap, we introduce SciVQR, a multimodal benchmark covering 54 subfields in mathematics, physics, chemistry, geography, astronomy, and biology. SciVQR includes domain-specific visuals, such as equations, charts, and diagrams, and challenges models to combine visual comprehension with reasoning. The tasks range from basic factual recall to complex, multi-step inferences, with 46% including expert-authored solutions. SciVQR not only evaluates final answers but also examines the reasoning process, providing insights into how models reach their conclusions. Our evaluation of leading MLLMs, including both proprietary and open-source models, reveals significant limitations in handling complex multimodal reasoning tasks, underscoring the need for improved multi-step reasoning and better integration of interdisciplinary knowledge in advancing MLLMs toward true scientific intelligence. The dataset and evaluation code are publicly available at https://github.com/CASIA-IVA-Lab/SciVQR.

Longteng Guo, Xuanxu Lin, Dongze Hao +5

Institute of Automation, Chinese Academy of Sciences (CASIA) · School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) · OPPO AI Center

Multimodal Large Language Models Scientific QA Multimodal Reasoning Multimodal Model Evaluation

Ask PDF

Mar 3, 2026

Evaluating Cross-Modal Reasoning Ability and Problem Characteristics with Multimodal Item Response Theory

Multimodal Large Language Models (MLLMs) have recently emerged as general architectures capable of reasoning over diverse modalities. Benchmarks for MLLMs should measure their ability for cross-modal integration. However, current benchmarks are filled with shortcut questions, which can be solved using only a single modality, thereby yielding unreliable rankings. For example, in vision-language cases, we can find the correct answer without either the image or the text. These low-quality questions unnecessarily increase the size and computational requirements of benchmarks. We introduce a multi-modal and multidimensional item response theory framework (M3IRT) that extends classical IRT by decomposing both model ability and item difficulty into image-only, text-only, and cross-modal components. M3IRT estimates cross-modal ability of MLLMs and each question's cross-modal difficulty, enabling compact, high-quality subsets that better reflect multimodal reasoning. Across 24 VLMs on three benchmarks, M3IRT prioritizes genuinely cross-modal questions over shortcuts and preserves ranking fidelity even when 50% of items are artificially generated low-quality questions, thereby reducing evaluation cost while improving reliability. M3IRT thus offers a practical tool for assessing cross-modal reasoning and refining multimodal benchmarks.

Shunki Uebayashi, Kento Masui, Kyohei Atarashi +5

Kyoto University · CyberAgent. · The Institute of Statistical Mathematics +1

Multimodal Model Evaluation Multimodal Reasoning Item Response Theory

Ask PDF

Jun 4, 2026

WorldBench: A Challenging and Visually Diverse Multimodal Reasoning Benchmark

In real-world applications, models are expected to perform reliably across diverse settings. Yet, many existing multimodal benchmarks expand task types without capturing the visual diversity needed to handle open-ended visual inputs. We present WorldBench, a challenging and visually diverse reasoning benchmark to evaluate Multimodal Large Language Models (MLLMs). We build a taxonomy of thousands of visual concepts across multiple domains (e.g., living things). Guided by this taxonomy, we curate a broad collection of images from search engines and existing datasets to comprehensively represent the visual world. Through structured trial-and-error, we manually design challenging questions that frontier MLLMs fail to answer. On quantitative and human evaluations, WorldBench achieves higher visual diversity than any existing diverse benchmark. Evaluating 15 MLLMs on WorldBench reveals weaknesses in visual understanding: even the strongest model reaches only 64.0% accuracy, while some models perform marginally above chance-level. We hope our work highlights the importance of visual diversity in building multimodal benchmarks.

Yida Yin, Harish Krishnakumar, Chung Peng Lee +9

Princeton University · NYU · University of Waterloo +1

Multimodal Large Language Models Multimodal Reasoning Visually Grounded Reasoning Multimodal Model Evaluation Multimodal Understanding

Ask PDF

Abstract

Multi-Image Reasoning

Explore similar work

SciVQR: A Multidisciplinary Multimodal Benchmark for Advanced Scientific Reasoning Evaluation

Evaluating Cross-Modal Reasoning Ability and Problem Characteristics with Multimodal Item Response Theory

WorldBench: A Challenging and Visually Diverse Multimodal Reasoning Benchmark