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%.
hub Baseline reference
MMBench: Is Your Multi-modal Model an All-around Player?
Baseline reference. 55% of citing Pith papers use this work as a benchmark or comparison.
83
Pith papers citing it
Baseline
55% of classified citations
abstract
Large vision-language models (VLMs) have recently achieved remarkable progress, exhibiting impressive multimodal perception and reasoning abilities. However, effectively evaluating these large VLMs remains a major challenge, hindering future development in this domain. Traditional benchmarks like VQAv2 or COCO Caption provide quantitative performance measurements but lack fine-grained ability assessment and robust evaluation metrics. Meanwhile, subjective benchmarks, such as OwlEval, offer comprehensive evaluations of a model's abilities by incorporating human labor, which is not scalable and may display significant bias. In response to these challenges, we propose MMBench, a bilingual benchmark for assessing the multi-modal capabilities of VLMs. MMBench methodically develops a comprehensive evaluation pipeline, primarily comprised of the following key features: 1. MMBench is meticulously curated with well-designed quality control schemes, surpassing existing similar benchmarks in terms of the number and variety of evaluation questions and abilities; 2. MMBench introduces a rigorous CircularEval strategy and incorporates large language models to convert free-form predictions into pre-defined choices, which helps to yield accurate evaluation results for models with limited instruction-following capabilities. 3. MMBench incorporates multiple-choice questions in both English and Chinese versions, enabling an apples-to-apples comparison of VLMs' performance under a bilingual context. To summarize, MMBench is a systematically designed objective benchmark for a robust and holistic evaluation of vision-language models. We hope MMBench will assist the research community in better evaluating their models and facilitate future progress in this area. The evalutation code of MMBench has been integrated into VLMEvalKit: https://github.com/open-compass/VLMEvalKit.
hub tools
citation-role summary
dataset 16
background 12
baseline 2
other 1
citation-polarity summary
claims ledger
- abstract Large vision-language models (VLMs) have recently achieved remarkable progress, exhibiting impressive multimodal perception and reasoning abilities. However, effectively evaluating these large VLMs remains a major challenge, hindering future development in this domain. Traditional benchmarks like VQAv2 or COCO Caption provide quantitative performance measurements but lack fine-grained ability assessment and robust evaluation metrics. Meanwhile, subjective benchmarks, such as OwlEval, offer comprehensive evaluations of a model's abilities by incorporating human labor, which is not scalable and
- dataset 2 / 82.0 69.8 62.8 76.0 49.3 53.5 Table 2. Comparison with SoTA models on 16 multimodal benchmarks. OCR-related benchmarks include: DocVQA test [82], ChartQA test [81], InfographicVQA test [83], TextVQA val [100], and OCRBench [67]. General multimodal benchmarks encompass: MME [26], RealWorldQA [125], AI2D test [39], MMMU val [135], MMBench-EN/CN test [66], CCBench dev [66], MMVet [133], SEED Image [46], and HallusionBench [30]. Additionally, the math dataset includes MathVista testmini [75]. *
- dataset MME[ 68]: MME is the first comprehensive evaluation benchmark designed for MLLMs. It assesses models' perception and cognitive abilities across 14 subtasks, including object presence, counting, position, color recognition, as well as commonsense reasoning, numerical computation, text translation, and code reasoning. We report the overall score across all tasks. MMBench[ 156]: MMBench evaluates the multimodal understanding of MLLMs through nearly 3,000 multiple- choice questions spanning 20 dimen
- dataset With 7B parameters, ShareGPT4V-7B outperforms competitors in 9 out of 11 benchmarks and ranks second on the others, despite these competitors using larger training datasets or more parameters. Benchmark names are abbreviated due to space limits. LLaV AW : LLaV A-Bench (In-the-Wild) [31]; MMEP : MME Perception [15]; MME C: MME Cognition [15]; MMB: MMBenchmark [33]; MMB CN : MMBench-Chinese [33]; SEED I: SEED-Bench (Image) [24]; MM-Vet [58]; QBench [55]; SQAI: ScienceQA-IMG [34]; VQAV 2 [17]; VizW
- dataset soning and even outperforms later approaches like Instruct- BLIP [14] on such benchmarks [ 55], while InstructBLIP excels in traditional VQA benchmarks that demands single- word or short answers. Given the significant differences in the model architecture and training data between them, the root cause of the disparity in their capabilities remains elusive, despite conjectures [37, 55]: the amount of training data, the usage of resamplers like Qformer [32], etc. To this arXiv:2310.03744v2 [cs.CV]
- dataset 1 Evaluation Setting Benchmarks.To comprehensively assess the capabilities of our models, we conduct evaluations across 42 public benchmarks, covering eight distinct categories:General VQA,STEM,OCR & Doc- ument,Visual Grounding,Spatial Reasoning,GUI Agents,Coding, andVideo Understanding. The following benchmarks are used for evaluation: • General VQA: MMBench-V1.1 [ 30], MMStar [ 7], BLINK(val) [ 11], MUIRBENCH [ 53], ZeroBench(val) [39], HallusionBench [15], GeoBench [2]; • STEM: MMMU(val) [67]
- dataset comparison results of human evaluation scores based on English prompts. For the vision-language understanding task, we assess the average scores across twelve benchmarks: SEEDBench-Img [45], OCRBench [ 59](with normalized results), MMVet [ 98], POPE [ 51], VQAv2 [ 27], GQA [ 34], TextVQA [78], ChartQA [61], AI2D [36], RealWorldQA [91], MMMU [99], and MMbench [58]. For the video generation task, we present comparison results of VBench. 1 Introduction Next-token prediction has revolutionized the f
co-cited works
representative citing papers
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 provides 11.5K heterogeneous college-level multimodal questions that current models solve at 56-59% accuracy, establishing a new standard for expert multimodal evaluation.
GazeVLM introduces internal gaze tokens that allow VLMs to dynamically suppress irrelevant visual features and simulate foveal attention for improved high-resolution multimodal reasoning.
Medical VLMs frequently select negated options that contradict visible chest X-ray findings, achieving only ~30% accuracy on direct presence probes, but a post-hoc consistency verifier raises accuracy above 95%.
XTC-Bench reveals that strong performance on generation or understanding tasks in unified multimodal models does not guarantee cross-task semantic consistency, which instead depends on how tightly coupled the learning objectives are across modalities.
ShredBench shows state-of-the-art MLLMs perform well on intact documents but suffer sharp drops in restoration accuracy as fragmentation increases to 8-16 pieces, indicating insufficient cross-modal semantic reasoning for VRDU.
PlotChain benchmark reports top MLLMs reaching ~80% field-level accuracy on engineering plot reading under human-like tolerances, but with persistent failures on frequency-domain tasks like bandpass and FFT spectra.
Derives exact guidance transition rates for discrete flow matching models that require only one model evaluation per sampling step and unify prior approximation-based methods.
FLARE is a vision-language model family using text-guided vision encoding, context-aware alignment decoding, dual-semantic mapping loss, and text-driven VQA synthesis to achieve deep cross-modal integration, outperforming larger models with only 630 vision tokens at 3B scale.
DualToken disentangles semantics and appearance via separate codebooks in one tokenizer, reporting 0.25 rFID, 82% ImageNet zero-shot accuracy, and gains over VILA-U on understanding and generation benchmarks.
PyramidDrop accelerates LVLMs by staged, similarity-based dropping of visual tokens that become redundant in deeper layers, delivering 40% faster training and 55% lower inference cost with comparable accuracy.
Janus decouples visual encoding into task-specific pathways inside a single autoregressive transformer to unify multimodal understanding and generation while outperforming earlier unified models.
WE-MATH benchmark reveals most LMMs rely on rote memorization for visual math while GPT-4o has shifted toward knowledge generalization.
Cambrian-1 is a vision-centric multimodal LLM family that evaluates over 20 vision encoders, introduces CV-Bench and the Spatial Vision Aggregator, and releases open models, code, and data achieving strong performance on visual grounding tasks.
MuirBench is a new benchmark showing that top multimodal LLMs struggle with robust multi-image understanding, with GPT-4o at 68% and open-source models below 33% accuracy.
LVBench is a new benchmark for extreme long video understanding that evaluates multimodal large language models on hour-scale videos using tasks designed to probe extended memory and comprehension.
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.
MathVerse is a benchmark that tests multi-modal LLMs on visual math by providing each problem in six versions with progressively less diagram and text information to measure true visual understanding.
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%.
SEED-Bench is a new benchmark of 19K multiple-choice questions for evaluating generative comprehension in multimodal LLMs across 12 image and video dimensions.
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.
FuRA uses block tensor-train factorization with fixed pretrained SVD basis to achieve full-rank spectral preconditioning, outperforming Full FT by +1.37 on LLaMA-3-8B commonsense reasoning and surpassing QLoRA in quantized settings.
GRIP-VLM applies group-relative policy optimization via reinforcement learning to prune visual tokens in VLMs, yielding up to 15% inference speedup at matched accuracy over prior methods.
citing papers explorer
-
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%.
-
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.
-
GazeVLM: Active Vision via Internal Attention Control for Multimodal Reasoning
GazeVLM introduces internal gaze tokens that allow VLMs to dynamically suppress irrelevant visual features and simulate foveal attention for improved high-resolution multimodal reasoning.
-
CXR-ContraBench: Benchmarking Negated-Option Attraction in Medical VLMs
Medical VLMs frequently select negated options that contradict visible chest X-ray findings, achieving only ~30% accuracy on direct presence probes, but a post-hoc consistency verifier raises accuracy above 95%.
-
Beyond Accuracy: Benchmarking Cross-Task Consistency in Unified Multimodal Models
XTC-Bench reveals that strong performance on generation or understanding tasks in unified multimodal models does not guarantee cross-task semantic consistency, which instead depends on how tightly coupled the learning objectives are across modalities.
-
ShredBench: Evaluating the Semantic Reasoning Capabilities of Multimodal LLMs in Document Reconstruction
ShredBench shows state-of-the-art MLLMs perform well on intact documents but suffer sharp drops in restoration accuracy as fragmentation increases to 8-16 pieces, indicating insufficient cross-modal semantic reasoning for VRDU.
-
PlotChain: Deterministic Checkpointed Evaluation of Multimodal LLMs on Engineering Plot Reading
PlotChain benchmark reports top MLLMs reaching ~80% field-level accuracy on engineering plot reading under human-like tolerances, but with persistent failures on frequency-domain tasks like bandpass and FFT spectra.
-
Discrete Guidance Matching: Exact Guidance for Discrete Flow Matching
Derives exact guidance transition rates for discrete flow matching models that require only one model evaluation per sampling step and unify prior approximation-based methods.
-
FLARE: Fully Integration of Vision-Language Representations for Deep Cross-Modal Understanding
FLARE is a vision-language model family using text-guided vision encoding, context-aware alignment decoding, dual-semantic mapping loss, and text-driven VQA synthesis to achieve deep cross-modal integration, outperforming larger models with only 630 vision tokens at 3B scale.
-
DualToken: Towards Unifying Visual Understanding and Generation with Dual Visual Vocabularies
DualToken disentangles semantics and appearance via separate codebooks in one tokenizer, reporting 0.25 rFID, 82% ImageNet zero-shot accuracy, and gains over VILA-U on understanding and generation benchmarks.
-
PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction
PyramidDrop accelerates LVLMs by staged, similarity-based dropping of visual tokens that become redundant in deeper layers, delivering 40% faster training and 55% lower inference cost with comparable accuracy.
-
Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation
Janus decouples visual encoding into task-specific pathways inside a single autoregressive transformer to unify multimodal understanding and generation while outperforming earlier unified models.
-
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.
-
Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs
Cambrian-1 is a vision-centric multimodal LLM family that evaluates over 20 vision encoders, introduces CV-Bench and the Spatial Vision Aggregator, and releases open models, code, and data achieving strong performance on visual grounding tasks.
-
MuirBench: A Comprehensive Benchmark for Robust Multi-image Understanding
MuirBench is a new benchmark showing that top multimodal LLMs struggle with robust multi-image understanding, with GPT-4o at 68% and open-source models below 33% accuracy.
-
LVBench: An Extreme Long Video Understanding Benchmark
LVBench is a new benchmark for extreme long video understanding that evaluates multimodal large language models on hour-scale videos using tasks designed to probe extended memory and comprehension.
-
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.
-
MathVerse: Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems?
MathVerse is a benchmark that tests multi-modal LLMs on visual math by providing each problem in six versions with progressively less diagram and text information to measure true visual understanding.
-
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%.
-
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.
-
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.
-
FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning
FuRA uses block tensor-train factorization with fixed pretrained SVD basis to achieve full-rank spectral preconditioning, outperforming Full FT by +1.37 on LLaMA-3-8B commonsense reasoning and surpassing QLoRA in quantized settings.
-
GRIP-VLM: Group-Relative Importance Pruning for Efficient Vision-Language Models
GRIP-VLM applies group-relative policy optimization via reinforcement learning to prune visual tokens in VLMs, yielding up to 15% inference speedup at matched accuracy over prior methods.
-
Vocabulary Hijacking in LVLMs: Unveiling Critical Attention Heads by Excluding Inert Tokens to Mitigate Hallucination
LVLMs show vocabulary hijacking by inert tokens that decode to hijacking anchors; HABI locates them, NHAR finds resilient heads, and HAVAE boosts those heads to cut hallucinations.
-
Mirror, Mirror on the Wall: Can VLM Agents Tell Who They Are at All?
Stronger VLM agents use mirror reflections for self-identification in controlled 3D tests, while weaker ones inspect but fail to extract or correctly attribute self-relevant information.
-
VisMMOE: Exploiting Visual-Expert Affinity for Efficient Visual-Language MoE Offloading
VisMMoE exploits visual-expert affinity via token pruning to achieve up to 2.68x faster VL-MoE inference on memory-constrained hardware while keeping accuracy competitive.
-
Seeing Isn't Believing: Uncovering Blind Spots in Evaluator Vision-Language Models
Evaluator VLMs frequently fail to detect quality-degrading perturbations in I2T and T2I outputs, with failure rates exceeding 50% in some cases.
-
Modality-Native Routing in Agent-to-Agent Networks: A Multimodal A2A Protocol Extension
Modality-native routing in A2A networks raises task accuracy from 32% to 52% over text-bottleneck baselines on a 50-task benchmark, but only when paired with capable downstream reasoning.
-
See Fair, Speak Truth: Equitable Attention Improves Grounding and Reduces Hallucination in Vision-Language Alignment
Equitable attention via Dominant Object Penalty and Outlier Boost Coefficient reduces object hallucinations in multimodal LLMs without retraining.
-
CLEAR: Unlocking Generative Potential for Degraded Image Understanding in Unified Multimodal Models
CLEAR uses degradation-aware fine-tuning, a latent representation bridge, and interleaved reinforcement learning to connect generative and reasoning capabilities in multimodal models for better degraded image understanding.
-
Same Content, Different Answers: Cross-Modal Inconsistency in MLLMs
State-of-the-art MLLMs show substantial inconsistency when reasoning over the same information presented in image, text, or mixed modalities, even after accounting for OCR errors, with inconsistency linked to visual factors and modality gap.
-
Routing-Based Continual Learning for Multimodal Large Language Models
Routing architecture for MLLMs enables continual learning with constant compute, matching multi-task learning performance and supporting cross-modal transfer.
-
InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
InternVL3.5 advances open-source multimodal models with Cascade RL for +16% reasoning gains and ViR for 4x inference speedup, with the 241B model reaching SOTA among open-source MLLMs on multimodal, reasoning, and agentic tasks.
-
GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning
GLM-4.5V reaches state-of-the-art results on 42 multimodal benchmarks among open-source models of similar size by applying reinforcement learning with curriculum sampling to a strong vision foundation model.
-
Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models
Multi-SpatialMLLM integrates depth perception, visual correspondence, and dynamic perception into MLLMs via a 27M-sample MultiSPA dataset and benchmark, yielding gains on multi-frame spatial tasks.
-
Mogao: An Omni Foundation Model for Interleaved Multi-Modal Generation
Mogao presents a causal unified model with deep fusion, dual encoders, and interleaved position embeddings that achieves strong performance on multi-modal understanding, text-to-image generation, and coherent interleaved outputs including zero-shot editing.
-
InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models
InternVL3-78B sets a new open-source SOTA of 72.2 on MMMU via native joint multimodal pre-training, V2PE, MPO, and test-time scaling while remaining competitive with proprietary models.
-
Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling
InternVL 2.5 is the first open-source MLLM to surpass 70% on the MMMU benchmark via model, data, and test-time scaling, with a 3.7-point gain from chain-of-thought reasoning.
-
LLaVA-CoT: Let Vision Language Models Reason Step-by-Step
LLaVA-CoT adds autonomous multistage reasoning to vision-language models, delivering 9.4% gains over its base model and outperforming larger models like Gemini-1.5-pro on reasoning benchmarks via a 100k annotated dataset and SWIRES test-time scaling.
-
ChatSearch: a Dataset and a Generative Retrieval Model for General Conversational Image Retrieval
Presents ChatSearch dataset and ChatSearcher generative model for conversational image retrieval on open-domain images, claiming superior performance on the new dataset and competitive results elsewhere.
-
Emu3: Next-Token Prediction is All You Need
Emu3 shows that next-token prediction on a unified discrete token space for text, images, and video lets a single transformer outperform task-specific models such as SDXL and LLaVA-1.6 in multimodal generation and perception.
-
SEED-X: Multimodal Models with Unified Multi-granularity Comprehension and Generation
SEED-X is a unified multimodal foundation model that handles multi-granularity visual semantics for both comprehension and generation across arbitrary image sizes and ratios.
-
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.
-
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.
-
DoRA: Weight-Decomposed Low-Rank Adaptation
DoRA improves LoRA by decomposing weights into magnitude and direction and updating only direction with low-rank matrices, closing much of the gap to full fine-tuning.
-
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%.
-
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.