MME-RealWorld is the largest manually annotated high-resolution benchmark for MLLMs, where even the best models achieve less than 60% accuracy on challenging real-world tasks.
hub Mixed citations
mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality
Mixed citation behavior. Most common role is background (68%).
84
Pith papers citing it
168
external citations · Pith
Background
68% of classified citations
abstract
Large language models (LLMs) have demonstrated impressive zero-shot abilities on a variety of open-ended tasks, while recent research has also explored the use of LLMs for multi-modal generation. In this study, we introduce mPLUG-Owl, a novel training paradigm that equips LLMs with multi-modal abilities through modularized learning of foundation LLM, a visual knowledge module, and a visual abstractor module. This approach can support multiple modalities and facilitate diverse unimodal and multimodal abilities through modality collaboration. The training paradigm of mPLUG-Owl involves a two-stage method for aligning image and text, which learns visual knowledge with the assistance of LLM while maintaining and even improving the generation abilities of LLM. In the first stage, the visual knowledge module and abstractor module are trained with a frozen LLM module to align the image and text. In the second stage, language-only and multi-modal supervised datasets are used to jointly fine-tune a low-rank adaption (LoRA) module on LLM and the abstractor module by freezing the visual knowledge module. We carefully build a visually-related instruction evaluation set OwlEval. Experimental results show that our model outperforms existing multi-modal models, demonstrating mPLUG-Owl's impressive instruction and visual understanding ability, multi-turn conversation ability, and knowledge reasoning ability. Besides, we observe some unexpected and exciting abilities such as multi-image correlation and scene text understanding, which makes it possible to leverage it for harder real scenarios, such as vision-only document comprehension. Our code, pre-trained model, instruction-tuned models, and evaluation set are available at https://github.com/X-PLUG/mPLUG-Owl. The online demo is available at https://www.modelscope.cn/studios/damo/mPLUG-Owl.
hub tools
citation-role summary
background 17
baseline 3
dataset 1
method 1
citation-polarity summary
claims ledger
- abstract Large language models (LLMs) have demonstrated impressive zero-shot abilities on a variety of open-ended tasks, while recent research has also explored the use of LLMs for multi-modal generation. In this study, we introduce mPLUG-Owl, a novel training paradigm that equips LLMs with multi-modal abilities through modularized learning of foundation LLM, a visual knowledge module, and a visual abstractor module. This approach can support multiple modalities and facilitate diverse unimodal and multimodal abilities through modality collaboration. The training paradigm of mPLUG-Owl involves a two-sta
- baseline MiniGPT-4 [7] Flan-T5-XL 47.40 4 29.89 7 42.84 4 VPGTrans [40] LLaMA-7B 41.81 5 31.40 5 39.10 5 MultiModal-GPT [12] LLaMA-7B 34.54 12 29.21 10 33.15 11 Otter [11] LLaMA-7B 35.16 11 30.35 6 33.91 8 OpenFlamingo [41] LLaMA-7B 34.51 13 29.25 9 33.14 12 LLaMA-Adapter V2 [42] LLaMA-7B 35.19 10 25.75 14 32.73 13 GVT [33] Vicuna-7B 35.49 9 27.77 12 33.48 10 mPLUG-Owl [9] LLaMA-7B 37.88 7 23.02 18 34.01 7 VideoLLM VideoChat [15] Vicuna-7B 39.02 6 33.68 4 37.63 6 Video-ChatGPT [16] LLaMA-7B 33.88 14 23.4
- method Contrastive Decodinge.g.VCD [94], IBD [226], ICD [166] Guided Decoding e.g.MARINE [212], GCD [38], DeCo [158] Visual Amplificatione.g.M3ID [41], IBD [226], AGLA [1] Others e.g.OPERA [66], Skip'\n' [56] Visual Prompting e.g.SoM-LLaVA [179] RAG e.g.ARA [133], FilterRAG [141] Ensembling e.g.RITUAL [169], MAD [107], MVP [134] Post-hoc Correction e.g.Woodpecker [188], Volcano [93], LURE [224], VFC [45] Fig. 1. The main content flow and categorization of this survey. Preprint, Vol. 1, No. 1, Article .
- background [46] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. [47] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024. [48] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu,
- background weights, training data, code, and hyperparam- eters are proprietary. This limits the academic community's ability to conduct research on au- toregressive vision-language models, e.g., to un- derstand how web-scraped image-text data affects models' performance and safety. Open-source al- ternatives, such as LLaVA [25], LLaMA-Adapter [41], BLIP-2 [ 23], and mPLUG-Owl [ 39], only take in single images, and they often directly train on curated datasets like COCO [ 24] rather than web data. In this t
- background proficiency in image-text dialogues through pre-training alignment and instruction fine-tuning. Subsequent research [3, 6, 10, 25, 31, 43, 57] has further enhanced LMMs by fo- cusing on the quality and diversity of pretraining and fine- tuning data. For instance, LLaV A [31] and InstructBLIP [10], with improved instruction fine-tuning, have advanced the understanding of complex prompts. mPLUG-Owl [57], Shikra [6], and KOSMOS-2 [43] have introduced new data types and training techniques, like gro
- background of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421, 2023. 6 [53] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, Qin Jin, Liang He, Xin Alex Lin, and Fei Huang. Ureader: Universal ocr-free visually-situated language understanding with multimodal large language model, 2023. 2 [54] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et
co-cited works
representative citing papers
MMMU provides 11.5K heterogeneous college-level multimodal questions that current models solve at 56-59% accuracy, establishing a new standard for expert multimodal evaluation.
MMMU-Pro is a stricter multimodal benchmark that removes text-only solvable questions, augments options, and requires reading text from images, yielding substantially lower model scores of 16.8-26.9%.
NEST is a new benchmark dataset for narrative event structures in long videos, with baselines reporting ETD below 8%, EL under 6%, EAE below 11%, and ERE at 35-44% F1.
AffectVerse improves multimodal emotion recognition by at least 2.57% on nine benchmarks through an Emotion World Module that performs short-horizon latent affective prediction via cross-modal temporal imagination and belief aggregation.
HumanVBench provides a 16-task benchmark for human-centric video understanding in MLLMs, created through automated annotation and distractor synthesis pipelines, and shows top models lag human performance on emotion perception and cross-modal alignment.
VidHal is a new benchmark that evaluates VLLM temporal hallucinations through a caption ordering task on videos with varying hallucination levels.
PolyMATH is a new 5,000-image benchmark where top MLLMs reach at most 41 percent accuracy on multi-modal mathematical reasoning, with ablation showing minimal gain from text over images.
WE-MATH benchmark reveals most LMMs rely on rote memorization for visual math while GPT-4o has shifted toward knowledge generalization.
MLVU is a new benchmark for long video understanding that uses extended videos across diverse genres and multi-task evaluations, revealing that current MLLMs struggle significantly and degrade sharply with longer durations.
HallusionBench shows GPT-4V reaches only 31.42% accuracy on paired questions testing language hallucination and visual illusion in LVLMs, with other models below 16%.
LLaMA-Adapter turns frozen LLaMA 7B into a capable instruction follower using only 1.2M new parameters and zero-init attention, matching Alpaca while extending to image-conditioned reasoning on ScienceQA and COCO.
AffectGPT-RL applies reinforcement learning to optimize non-differentiable emotion wheel metrics in open-vocabulary multimodal emotion recognition, yielding performance gains and state-of-the-art results on basic emotion recognition benchmarks.
DualComp uses a lightweight router to split visual token compression into a semantic stream with size-adaptive clustering and a geometric stream with path-tracing recovery, enabling low-cost high-fidelity UHR remote sensing interpretation.
SEED-Bench is a new benchmark of 19K multiple-choice questions for evaluating generative comprehension in multimodal LLMs across 12 image and video dimensions.
Large vision-language models exhibit severe object hallucination that varies with training instructions, and the proposed POPE polling method evaluates it more stably and flexibly than prior approaches.
MER-R1 uses dual-objective RL to optimize fast-thinking recall and slow-thinking precision separately in multimodal emotion recognition, with calibration to align them, yielding SOTA results on two benchmarks.
TOPS formulates visual token pruning as constructing Token Optimal Preservation Sets using three information-theoretic principles and demonstrates superior performance on MLLM benchmarks.
A cue-coverage reward plus a modality-token KL penalty makes multimodal emotion-reasoning models cite more real visual/audio evidence and hallucinate less, with reported SoTA on emotion benchmarks.
HPP decouples perception from reasoning in long-video VLMs by having an LLM run iterative programmatic probes on hierarchically segmented video, reporting gains on LongVideoBench, EgoSchema, VideoMME, and MLVU.
Qwen-RobotWorld is a language-conditioned video world model using Double-Stream MMDiT, an 8.6M-frame embodied corpus, and progressive curriculum training that ranks first on EWMBench and DreamGen Bench.
Chat-Scene++ improves 3D scene understanding in multimodal LLMs by representing scenes as context-rich object sequences with identifier tokens and grounded chain-of-thought reasoning, reaching state-of-the-art on five benchmarks using pre-trained encoders.
An adaptive, multi-tool verifier for multimodal RL that scores final answers, spatial grounding, and reasoning quality improves agentic models on spatial, hallucination, embodied, and robotic benchmarks.
ORCA is an agentic reasoning framework that enhances factual accuracy and adversarial robustness of pretrained LVLMs via an Observe-Reason-Critique-Act loop with small vision models, reporting accuracy gains of up to 40% on hallucination benchmarks and 20% under adversarial perturbations.
citing papers explorer
-
MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans?
MME-RealWorld is the largest manually annotated high-resolution benchmark for MLLMs, where even the best models achieve less than 60% accuracy on challenging real-world tasks.
-
MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI
MMMU provides 11.5K heterogeneous college-level multimodal questions that current models solve at 56-59% accuracy, establishing a new standard for expert multimodal evaluation.
-
MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark
MMMU-Pro is a stricter multimodal benchmark that removes text-only solvable questions, augments options, and requires reading text from images, yielding substantially lower model scores of 16.8-26.9%.
-
NEST: Narrative Event Structures in Time for Long Video Understanding
NEST is a new benchmark dataset for narrative event structures in long videos, with baselines reporting ETD below 8%, EL under 6%, EAE below 11%, and ERE at 35-44% F1.
-
AffectVerse: Emotional World Models for Multimodal Affective Computing
AffectVerse improves multimodal emotion recognition by at least 2.57% on nine benchmarks through an Emotion World Module that performs short-horizon latent affective prediction via cross-modal temporal imagination and belief aggregation.
-
HumanVBench: Probing Human-Centric Video Understanding in MLLMs with Automatically Synthesized Benchmarks
HumanVBench provides a 16-task benchmark for human-centric video understanding in MLLMs, created through automated annotation and distractor synthesis pipelines, and shows top models lag human performance on emotion perception and cross-modal alignment.
-
VidHal: Benchmarking Temporal Hallucinations in Vision LLMs
VidHal is a new benchmark that evaluates VLLM temporal hallucinations through a caption ordering task on videos with varying hallucination levels.
-
Polymath: A Challenging Multi-modal Mathematical Reasoning Benchmark
PolyMATH is a new 5,000-image benchmark where top MLLMs reach at most 41 percent accuracy on multi-modal mathematical reasoning, with ablation showing minimal gain from text over images.
-
We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning?
WE-MATH benchmark reveals most LMMs rely on rote memorization for visual math while GPT-4o has shifted toward knowledge generalization.
-
MLVU: Benchmarking Multi-task Long Video Understanding
MLVU is a new benchmark for long video understanding that uses extended videos across diverse genres and multi-task evaluations, revealing that current MLLMs struggle significantly and degrade sharply with longer durations.
-
HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models
HallusionBench shows GPT-4V reaches only 31.42% accuracy on paired questions testing language hallucination and visual illusion in LVLMs, with other models below 16%.
-
LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention
LLaMA-Adapter turns frozen LLaMA 7B into a capable instruction follower using only 1.2M new parameters and zero-init attention, matching Alpaca while extending to image-conditioned reasoning on ScienceQA and COCO.
-
AffectGPT-RL: Revealing Roles of Reinforcement Learning in Open-Vocabulary Emotion Recognition
AffectGPT-RL applies reinforcement learning to optimize non-differentiable emotion wheel metrics in open-vocabulary multimodal emotion recognition, yielding performance gains and state-of-the-art results on basic emotion recognition benchmarks.
-
Semantic-Geometric Dual Compression: Training-Free Visual Token Reduction for Ultra-High-Resolution Remote Sensing Understanding
DualComp uses a lightweight router to split visual token compression into a semantic stream with size-adaptive clustering and a geometric stream with path-tracing recovery, enabling low-cost high-fidelity UHR remote sensing interpretation.
-
SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehension
SEED-Bench is a new benchmark of 19K multiple-choice questions for evaluating generative comprehension in multimodal LLMs across 12 image and video dimensions.
-
Evaluating Object Hallucination in Large Vision-Language Models
Large vision-language models exhibit severe object hallucination that varies with training instructions, and the proposed POPE polling method evaluates it more stably and flexibly than prior approaches.
-
MER-R1: Multimodal Emotion Reasoning via Slow-Fast Thinking Synergy
MER-R1 uses dual-objective RL to optimize fast-thinking recall and slow-thinking precision separately in multimodal emotion recognition, with calibration to align them, yielding SOTA results on two benchmarks.
-
TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference
TOPS formulates visual token pruning as constructing Token Optimal Preservation Sets using three information-theoretic principles and demonstrates superior performance on MLLM benchmarks.
-
Omni-Perception Policy Optimization for Multimodal Emotion Reasoning
A cue-coverage reward plus a modality-token KL penalty makes multimodal emotion-reasoning models cite more real visual/audio evidence and hallucinate less, with reported SoTA on emotion benchmarks.
-
HPP: Hierarchical Programmatic Probing for Long Video Understanding by Decoupling Perception and Reasoning
HPP decouples perception from reasoning in long-video VLMs by having an LLM run iterative programmatic probes on hierarchically segmented video, reporting gains on LongVideoBench, EgoSchema, VideoMME, and MLVU.
-
Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation
Qwen-RobotWorld is a language-conditioned video world model using Double-Stream MMDiT, an 8.6M-frame embodied corpus, and progressive curriculum training that ranks first on EWMBench and DreamGen Bench.
-
Chat-Scene++: Exploiting Context-Rich Object Identification for 3D LLM
Chat-Scene++ improves 3D scene understanding in multimodal LLMs by representing scenes as context-rich object sequences with identifier tokens and grounded chain-of-thought reasoning, reaching state-of-the-art on five benchmarks using pre-trained encoders.
-
Multimodal Reinforcement Learning with Adaptive Verifier for AI Agents
An adaptive, multi-tool verifier for multimodal RL that scores final answers, spatial grounding, and reasoning quality improves agentic models on spatial, hallucination, embodied, and robotic benchmarks.
-
ORCA: An Agentic Reasoning Framework for Hallucination and Adversarial Robustness in Vision-Language Models
ORCA is an agentic reasoning framework that enhances factual accuracy and adversarial robustness of pretrained LVLMs via an Observe-Reason-Critique-Act loop with small vision models, reporting accuracy gains of up to 40% on hallucination benchmarks and 20% under adversarial perturbations.
-
LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL
A two-stage RL framework first boosts text reasoning in 3B LMMs then adapts it to multimodal inputs, producing modest benchmark gains of 4.5-4.8%.
-
When Large Vision-Language Models Meet Person Re-Identification
LVLM-ReID guides LVLMs to produce refined semantic tokens as pedestrian identity features for ReID, achieving competitive benchmark results without additional image-text data.
-
LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding
LongVU adaptively compresses long video tokens using DINOv2-based frame deduplication, text-guided cross-modal selection, and temporal spatial reduction to improve video-language understanding in MLLMs with minimal detail loss.
-
SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference
SparseVLM uses text-guided attention to prune and recycle visual tokens in VLMs, delivering 54% FLOPs reduction and 37% lower latency with 97% accuracy retention on LLaVA.
-
VideoPhy: Evaluating Physical Commonsense for Video Generation
VideoPhy benchmark shows state-of-the-art text-to-video models follow physical commonsense and text prompts in only 39.6% of cases for the best model.
-
MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training
MM1 models achieve state-of-the-art few-shot multimodal results by pre-training on a careful mix of image-caption, interleaved, and text-only data with optimized image encoders.
-
TempCompass: Do Video LLMs Really Understand Videos?
TempCompass benchmark reveals that state-of-the-art Video LLMs have poor ability to perceive temporal aspects such as speed, direction, and ordering in videos.
-
Mobile-Agent: Autonomous Multi-Modal Mobile Device Agent with Visual Perception
Mobile-Agent is a vision-centric autonomous agent that uses MLLMs to perceive UI elements, plan complex multi-step tasks, and operate mobile apps without relying on XML or system metadata, showing strong results on the introduced Mobile-Eval benchmark.
-
MoE-LLaVA: Mixture of Experts for Large Vision-Language Models
MoE-LLaVA applies mixture-of-experts sparsity to LVLMs via MoE-Tuning, delivering LLaVA-1.5-7B level visual understanding and better hallucination resistance with only ~3B active parameters.
-
SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents
SeeClick improves visual GUI agents via GUI grounding pre-training on automatically curated data and introduces the ScreenSpot benchmark, with results indicating that stronger grounding boosts downstream task performance.
-
MVBench: A Comprehensive Multi-modal Video Understanding Benchmark
MVBench is a benchmark of 20 temporal video understanding tasks built by transforming static tasks into dynamic ones, with VideoChat2 outperforming prior MLLMs by over 15%.
-
An Embodied Generalist Agent in 3D World
LEO is an embodied generalist agent that performs 3D captioning, question answering, reasoning, navigation, and manipulation after 3D vision-language alignment followed by vision-language-action instruction tuning on large-scale object- and scene-level datasets.
-
Video-LLaVA: Learning United Visual Representation by Alignment Before Projection
Video-LLaVA creates a unified visual representation for images and videos via pre-projection alignment, enabling mutual enhancement from joint training and strong results on image and video benchmarks.
-
LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment
LanguageBind aligns video, infrared, depth, and audio to a frozen language encoder via contrastive learning on the new VIDAL-10M dataset, extending video-language pretraining to N modalities.
-
Analyzing and Mitigating Object Hallucination in Large Vision-Language Models
LURE reduces object hallucination in LVLMs by 23% via post-hoc revision informed by co-occurrence, uncertainty, and text position analysis.
-
Aligning Large Multimodal Models with Factually Augmented RLHF
Factually Augmented RLHF aligns large multimodal models to reduce hallucinations, reaching 94% of GPT-4 on LLaVA-Bench and 60% improvement on the new MMHAL-BENCH.
-
Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning
A new dataset of 400k visual instructions including negative examples at three semantic levels reduces hallucinations in models like MiniGPT-4 when used for fine-tuning while improving benchmark performance.
-
Mitigating Action-Relation Hallucinations in LVLMs via Relation-aware Visual Enhancement
A new attention-enhancement method using ARS scores and RVE reduces action-relation hallucinations in LVLMs while generalizing to spatial and object hallucinations.
-
LLaVA-UHD v4: What Makes Efficient Visual Encoding in MLLMs?
LLaVA-UHD v4 reduces visual-encoding FLOPs by 55.8% for high-resolution images in MLLMs via slice-based encoding plus intra-ViT early compression while matching or exceeding baseline performance on document, OCR, and VQA benchmarks.
-
ReasonEdit: Towards Interpretable Image Editing Evaluation via Reinforcement Learning
ReasonEdit uses a new CoT dataset and reinforcement learning to produce interpretable, human-aligned evaluations of text-guided image edits.
-
ICU-Bench:Benchmarking Continual Unlearning in Multimodal Large Language Models
Existing continual unlearning methods fail to maintain forgetting, retained knowledge, and utility over 100 sequential privacy-deletion tasks in multimodal document models.
-
SSL-R1: Self-Supervised Visual Reinforcement Post-Training for Multimodal Large Language Models
SSL-R1 reformulates visual SSL tasks into verifiable puzzles to supply rewards for RL post-training of MLLMs, yielding gains on multimodal benchmarks without external supervision.
-
R-CoV: Region-Aware Chain-of-Verification for Alleviating Object Hallucinations in LVLMs
R-CoV is a six-step region-aware chain-of-verification technique that elicits coordinate and description outputs from LVLMs themselves to detect and reduce object hallucinations without external models or retraining.
-
CoME-VL: Scaling Complementary Multi-Encoder Vision-Language Learning
CoME-VL fuses contrastive and self-supervised vision encoders via entropy-guided multi-layer aggregation and RoPE cross-attention to improve vision-language model performance on benchmarks.
-
Are We on the Right Way for Evaluating Large Vision-Language Models?
Current LVLM benchmarks overestimate capabilities because many questions can be answered without images due to design flaws or data leakage; MMStar is a human-curated set of 1,500 vision-indispensable samples across 6 capabilities and 18 axes with new metrics for leakage and true multi-modal gain.
-
ShareGPT4V: Improving Large Multi-Modal Models with Better Captions
A new 1.2M-caption dataset generated via GPT-4V improves LMMs on MME and MMBench by 222.8/22.0/22.3 and 2.7/1.3/1.5 points respectively when used for supervised fine-tuning.