This article analyzes why measuring the real capability of a multimodal system is harder than benchmark rankings suggest. By reading it, you will understand grounding and language bias (and why the latter can make a model answer correctly without actually processing the image), how evaluation-data contamination artificially inflates published results, and what benchmarks such as OCRBench v2, MMAU, ZeroBench and HallusionBench reveal about the field's real limits in documents, audio, long video and spatial reasoning. The article is useful both for technical readers evaluating models and for anyone who wants to interpret the comparisons circulating in the field rigorously.
Evaluating whether a language model produces accurate and useful answers is already a complex problem, but adding the visual or auditory dimension multiplies the difficulty in two different ways.
The first is that current multimodality benchmarks have two systematic problems that lead us to overestimate real capabilities: evaluation-data contamination and the dominance of text in benchmarks.
The second is that evaluation has historically been dominated by vision-language tasks, leaving whole capabilities poorly measured: understanding documents with complex layouts, reasoning over audio, temporal coherence in video, or the quality of outputs generated in modalities other than text. OCRBench v2 and MMAU are recent reminders that this space, evaluated only superficially until now, remains difficult terrain for the best current models.
In multimodal systems, grounding is the degree to which the model's answer is supported by the actual content of the image or audio, rather than by statistical inferences about what kind of answer is likely given the text of the question. A model can correctly answer "What color is the car in the image?" without actually processing the image if the color most frequent in its training for cars in similar contexts happens to match the correct answer.
That model does not have grounding; it has a strong language bias that produces the right answer for the wrong reasons. The difference remains invisible as long as the statistical bias and the correct answer point in the same direction.
To measure grounding, benchmarks need examples where the correct answer violates statistical expectations. If every question about fruit in images has an answer matching the fruit most represented in training, there is no way to distinguish a model with real visual understanding from one that answers from probability.
The Visual Question Answering Challenge (VQA), historically one of the most widely used benchmarks, has exactly this problem Goyal et al., 2017. A 2017 analysis showed that a model that completely ignored the images and answered only from the distribution of the most frequent responses for each question type still achieved surprisingly high results. Later benchmark improvements introduced balancing techniques to reduce this bias, although they did not remove it completely.
Grounding: what it means for a model to truly "see"
The difference between answering from visual evidence and answering from the statistical distribution learned during training.
The central evaluation question
When a model answers correctly, did it do so because it processed the image, or because the answer was statistically likely regardless of the image?
Without grounding
The image does not determine the answer
image ignored
↓
Statistical prior
distribution learned during training
↓
"White"
most frequent answer in the corpus
vs
With grounding
The image determines the answer
image analyzed
↓
Visual evidence
actual image content processed
↓
"Red"
actual color of the object in the image
Reference question in both columns
"What color is the car in the image?" — The actual car is red, but cars in the training corpus are mostly white or gray.
The underlying problem
The failure is detectable only when the statistical prior and the correct answer diverge. If the image shows the most likely case, the ungrounded model and the grounded model produce the same answer, and the benchmark cannot distinguish them.
Invisible failure
Prior matches the image
Actual image
white car
+
Strongest prior
"white"
→
Result
"White" ✓
The model is correct, but it may have answered without looking at the image. The benchmark detects nothing.
Visible failure
Prior diverges from the image
Actual image
red car
+
Strongest prior
"white"
→
Result
"White" ✗
The model fails. Only this case makes it possible to detect the absence of real grounding.
Implication for benchmark design
A benchmark without cases of the second type cannot distinguish a model with real visual understanding from one that guesses from probability. Results are valid only if the evaluation set systematically includes examples where the statistical prior fails.
The solution: designed counterexamples
Benchmarks that address this problem systematically include images where the correct answer violates the most frequent statistical expectations in the training corpus.
1
Objects with atypical properties
A blue banana, a bright green car, a triangular building. If the model answers with the most common color or shape, the failure is unambiguous and cannot be attributed to chance.
2
Counts that violate the statistical norm
If "2 or 3" dominates the corpus, the benchmark includes images with 7, 11, or 14 elements of the same type. An ungrounded model falls back to the usual statistical range even when the image shows something else.
3
Uncommon spatial configurations
The sphere below the cube. The cup to the left of the plate when the training prior is "to the right". Cases where the actual configuration reverses the statistical expectation.
MMStar
Designed with counterexamples where the statistical prior systematically leads to error. Chen et al., 2024.
SEEDBench
Evaluates consistency and spatial reasoning, not only final-answer accuracy. Li et al., 2023.
The second systematic problem is contamination. Foundation models are pretrained on massive amounts of internet data, and there is no guarantee that image-description pairs or evaluation datasets do not appear in those data.
Contamination in text is already a documented problem: models that obtain exceptional results on some reasoning benchmarks can recite the correct answers when given the problem identifier, suggesting that the benchmark was present in their training data. In multimodality the problem is potentially larger because benchmark images are often publicly available photographs that may have appeared in pretraining together with their descriptions or labels.
The technical solution is to use benchmarks whose evaluation data did not exist on the internet when the model was pretrained or that are protected from indexing. In practice, however, the most useful recommendation is to interpret results skeptically when the evaluated model was pretrained at internet scale, especially if the benchmark is old.
The most rigorous labs perform contamination analyses before publishing results: they search their training data for images similar to those in the evaluation benchmark and exclude those images from the final analysis. Without that analysis, published results are an upper bound on the model's real capability on that benchmark, not a direct measurement.
Benchmark contamination: when the test has already been seen
Evaluation images appear in the pretraining data. The score reflects memorization, not real generalization capability.
The problem
Foundation models are pretrained on massive amounts of internet data. There is no guarantee that the image-description pairs used for evaluation did not appear in those data. If the model has already seen the benchmark images during training, the score does not measure generalization.
Contamination mechanism
Internet
public photographs with labels · published datasets · image-text pairs
→
Pretraining
model training data
→
Benchmark
evaluation images (possibly already seen)
→
Inflated score
does not reflect real generalization capability
Without contamination analysis, published results are an upper bound on the model's real capability on that benchmark, not a direct measurement.
The key difference from text
Text contamination is already a documented problem. In multimodality the problem is structurally larger: benchmark images are often public photographs that may have appeared in pretraining together with their descriptions, labels and surrounding textual context.
Text contamination
Contamination vector
Correct benchmark answers appear in training data
Warning sign
The model recites answers when given the problem ID
Detection
Search for text strings in the training data
Multimodal contamination
more serious
Contamination vector
The full image + its caption + surrounding text may be in pretraining
Additional problem
Image benchmarks often use public photographs that existed before the model
Detection
Search for similar images (perceptual hash) in training data — more expensive
Practical implication
Older benchmarks are more suspect: more time available on the internet means a greater chance that their images are in the pretraining data of current models. Results are more reliable when the benchmark is newer and better protected from indexing.
Mitigation in practice
The most rigorous labs run contamination analyses before publishing results. Without this analysis, the results are an upper bound on real capability, not a direct measurement.
1
Perceptual-similarity search
Compute perceptual hashes for benchmark images and search for matches in the pretraining data. Exact identity is not required: near-identical images already constitute contamination.
2
Exclusion from the final analysis
Images with matches in the training data are excluded from the result analysis. Published results report what percentage of the benchmark was excluded for this reason.
3
Use benchmarks with non-indexed data
The most robust solution is to use benchmarks created from data that did not exist on the internet when the model was pretrained, or that are protected from public indexing.
Interpretation rule
Skepticism about results should increase as the benchmark gets older, the model's pretraining becomes larger, and the benchmark images were more publicly accessible at training time.
3. Language bias: answering from probability, not evidence¶
The language bias/prior is the tendency of models to generate answers that are statistically likely given the text of the question, regardless of the image content. It is the subtlest form of missing grounding and the hardest to detect with standard benchmarks because the errors it causes are invisible when the statistical distribution matches the distribution of correct answers.
Ablation experiments are the standard tool for measuring it: present the model with the question without the image and observe whether the answer distribution changes significantly. When the model without the image obtains results similar to the model with the image, language bias is dominating the answer.
The effect is especially strong in categories where training distributions are skewed: questions about the usual color of certain objects, an animal species when only one animal is visible, or the number of elements in scenes where two or three is the dominant frequency. In all these cases there is a highly skewed answer distribution that the model learns during training and uses as a shortcut, ignoring the image when the bias is strong enough.
Designing benchmarks that resist language bias requires active techniques: counterexamples where an object has an unusual color, scenes where the number of elements violates expectations, and spatial configurations that are uncommon in training. MMStar and SEEDBench are examples of benchmarks designed with explicit attention to this problem.
Language prior: answering from probability, not evidence
The tendency of models to generate statistically likely answers from the text of the question, regardless of the actual content of the image.
What the language prior is
During pretraining, the model learns answer distributions for each type of question. Those distributions act as shortcuts: for questions whose answer is highly predictable, the model generates it directly from the prior without needing the image.
What color is the car?
Learned answer distribution (without seeing the image)
white
42%
gray
26%
black
17%
red
8%
other
7%
Result
The model answers "white" even though the car in the image is red. When the prior is strong enough, the model does not need to consult the image to generate the most likely answer.
Why it is difficult to detect
The error is invisible when the statistical distribution matches the distribution of correct answers. In benchmarks where most cars are white or gray, a model without real vision can score highly using only the prior.
The diagnostic tool
The ablation experiment presents the same question to the model with and without the image. If the answer distribution does not change significantly, the language prior dominatestes generation and the image is not contributing to the answer.
Real scenario
red car in the image
"What color is the car?"
Model + image
image active
↓
"White"
answers from the prior, not the image
prior dominates
Model without image
image removed
↓
"White"
same answer without seeing anything
prior confirmed
Interpreting the result
If the answer is the same with and without the image, the image is not contributing to generation. The model is not "seeing": it is generating from the statistical distribution learned during training. A different result between the two conditions is necessary (but not sufficient) evidence that the image influences the answer.
Where the prior is strongest
The effect is especially pronounced in categories where training distributions are skewed. In those categories, the statistically most likely answer is so dominant that the model disregards visual evidence even when it is present.
Colors of common objects
"What color is the car?"
Cars in everyday-image datasets are predominantly white, gray and black. The "white" prior is so strong that it can override evidence for unusual colors.
Robust benchmark: include objects with deliberately unusual colors for that type of object.
Animal species
"What animal is in the image?"
If only one animal is visible, the prior for "which species is most represented in similar scenes" is very strong. Dogs dominate domestic images, followed by cats. Rare species are lost.
Robust benchmark: include species that are uncommon in the corpus in domestic or urban contexts.
Counting elements
"How many objects are in the image?"
The 2–4 range dominates everyday-image datasets. When asked for the number of elements, the model tends to answer within that range regardless of what the image shows.
Robust benchmark: include images with quantities that violate the usual statistical range (0, 1, 7, 12, 20).
Final-answer accuracy does not capture all the relevant information about a multimodal model's capabilities. More rigorous evaluators include three additional dimensions that reveal different aspects of visual understanding.
Consistency. A genuinely capable model should answer paraphrases of the same question consistently. When the answer changes drastically under a semantically equivalent formulation, the model does not have robust understanding of the visual content; it is sensitive to the surface form of the question.
Localization when relevant. For tasks where the answer depends on the location of elements in the image, evaluation should verify not only whether the final answer is correct but also whether the model can indicate where the relevant element is in the image. A model that correctly answers "there are three cars" but cannot delimit where they are has a different kind of understanding from a model that can, and that difference matters in applications where localization is part of the expected result Hu et al., 2024.
Calibration. Models should be able to express uncertainty when visual content is ambiguous or when the question does not have a clear answer given the available content. A model that always generates a high-confidence answer, even for ambiguous images or questions that cannot be answered without additional information, is not properly calibrated. In production, this becomes falsely definitive answers where the system should abstain or ask for clarification.
Four dimensions accuracy does not capture
A model can get benchmark answers right for the wrong reasons. These four metrics reveal whether visual understanding is real or apparent.
⊙
Grounding
Is the answer grounded in the actual content of the image?
What it measures
Whether the model's answer depends on the actual image content or on the statistical prior learned during training. A grounded model produces different answers for different images even when the question is the same.
Grounding failure
↓
"white"
answers with the most frequent color in the corpus and ignores the image
With grounding
↓
"orange"
answers from the actual visual evidence
Diagnostic test
Ablation experiment: present the same question without the image. If the answer does not change, the prior dominates and the grounding is apparent rather than real.
⇄
Consistency
Does it answer the same way to semantically equivalent paraphrases?
What it measures
A model with robust understanding should produce the same answer for different phrasings of the same question about the same image. When the answer changes sharply under a semantically equivalent reformulation, the model is sensitive to the surface form of the text rather than the visual content.
Same image · same question · different phrasings
"How many people are in the image?"
"Three"
"How many people appear in the scene?"
"Three people"
"In the image, can you count how many people there are?"
"Yes, there are four"
inconsistency
Failure signal
The answer changes when the wording changes even though the content does not. The model does not have a stable representation of the visual content: it answers according to how the question is phrased, not according to what the image shows.
SEEDBench
Includes consistency evaluation on the same image using rephrased questions. Li et al., 2023.
◎
Localization
Can it point to where the relevant element is in the image?
What it measures
For tasks where the answer depends on the position of elements in the image, evaluation should verify not only whether the final answer is correct but also whether the model can delimit where the relevant element is located. These are distinct kinds of understanding, and both matter in real applications.
Partial understanding
"There are three cars"
correct answer · no localization
Understanding with localization
"Three cars: left, upper-center, right"
answer + spatial localization
Why it matters in production
An object-detection, scene-analysis, or real-time assistance system needs to know where elements are, not only that they exist. Localization is the dimension that distinguishes "knowing what is there" from "knowing where it is".
≈
Calibration
Does it express uncertainty for ambiguous images or questions that cannot be answered?
What it measures
A well-calibrated model expresses uncertainty when the visual content is ambiguous or when the question has no clear answer given the available content. In production, a model that always answers with high confidence produces falsely definitive claims where it should abstain or ask for clarification.
Poorly calibrated model
"What make is that car?" [blurry image, unrecognizable logo]
"It is a 2019 Toyota Corolla."
high confidence · no evidence · hallucination
Well-calibrated model
"What make is that car?" [blurry image, unrecognizable logo]
"The image does not have enough resolution to identify the make with certainty. The vehicle appears to be a midsize sedan, but I cannot be more specific."
uncertainty expressed · answer matched to the evidence
Production failure signal
A model that always generates a high-confidence answer, even for ambiguous images or questions that cannot be answered from the available content, is not calibrated. In production, that becomes falsely definitive answers where the system should abstain or surface the uncertainty to the user.
Multimodality evaluation has been dominated by VQA and visual-grounding tasks because they are the easiest to automate and turn into benchmarks with single-choice answers. That has created a systematic blind spot: the domains where evaluation is hardest to automate are precisely those that reveal the most about current model limitations.
Documents with complex layouts. OCRBench v2, published in 2024, evaluated advanced multimodal models on text localization, handwriting recognition and logical reasoning over documents Liu et al., 2024. The results showed that even models with high VQA scores struggle in real-document scenarios: text in non-standard orientations, tables with merged cells, mathematical formulas embedded in text flow, or questions requiring information to be crossed between several regions of the same document. OmniDocBench, presented at CVPR 2025, extended this evaluation to documents with non-standard layouts: multiple columns, floating figures and elements with non-linear alignment. The evaluation of 13 SOTA models showed the same collapse: systems that reach 80–90% accuracy on standard text fall to 36.9% on complex-layout reconstruction, confirming that the limit is not visual recognition but the integration of structure and semantics in scenes that are not linear prose Ouyang et al., 2025.
OCRBench v2: "reading" a document ≠ "reasoning over" it
Models that read clean text accurately still struggle with real document layouts. OCRBench v2 measures the gap between those two regimes.
Solved regime
Standard horizontally oriented text
Input document
Standard font · horizontal orientation · no complex structure
Model output
Text extraction
Text extracted with high fidelity. Occasional minor transcription errors on uncommon characters.
Typical accuracy
~90%
Tasks handled well
Transcribe paragraphs of continuous text
Read numbers and dates in standard formats
Identify titles and sections with clear formatting
Why it works well here
Horizontally oriented text in standard fonts is well represented in model pretraining data. The model can apply pattern recognition directly without needing to reason over structure.
Difficulty zone
Complex layout: structure that breaks linear parsing
Input document
Merged cells · rotated text · formula embedded in the flow
Model output
Frequent problems
Values are often assigned to the wrong columns when cells are merged. Rotated text may be ignored or mistranscribed. Formulas can be detached from the data flow they belong to.
Typical accuracy
~54%
Degraded performance
Read text in non-standard orientations
Assign values to rows in tables with merged cells
Extract formulas embedded in narrative text
Why it fails here
The model processes the document as a text sequence rather than as a structure with spatial relationships between cells. Merged cells, rotated text, and formulas require treating layout as semantic structure, not merely extracting character strings.
Main limit
Reasoning across document regions
Input document
Two document regions pages apart
The question that fails
Cross-region reasoning question
"Does the VAT declared in section 5.4 equal 21% of the taxable base reported in section 2.1?"
Required steps
1
Locate "taxable base" in region A → €1,458.53
2
Locate "VAT" in region B → €388.97
3
Calculate: €1,458.53 × 0.21 = €306.29
!
€388.97 ≠ €306.29 → discrepancy. The model often fails to reach this step.
Critical performance
Link information from distant regions of the document
Verify numerical consistency across sections
Reason over the document's logical structure
The gap revealed by OCRBench v2
The gap between "reading" and "reasoning over" a document remains substantial. A model may extract text accurately and still fail on questions that require combining information from different regions, because that requires maintaining logical coherence across the document structure rather than merely recognizing characters.
Expert audio. MMAU, published by Adobe Research in 2024, evaluated audio understanding and reasoning in three categories: speech, non-verbal environmental sounds and music Sakshi et al., 2024. The results showed that even the strongest models remain significantly below expert human performance on the hardest tasks in each category, with especially marked degradation when the task requires reasoning about the cause of a sound rather than merely identifying it, inferring context from multiple simultaneous sound sources, or distinguishing musical variants that share superficial structure. These limits are especially relevant for native-audio systems such as Gemini 2.5 or Qwen2.5-Omni, where capability expectations often exceed what available benchmarks can confirm.
Expert reasoning. MMMU, published in 2023, evaluated models' ability to reason over visual content in 30 university subjects grouped into 6 disciplines: Art and Design, Business, Science, Health and Medicine, Humanities and Social Sciences, and Engineering and Technology. Unlike image-description benchmarks, MMMU requires integrating domain knowledge with visual understanding: reading the image correctly is not enough; the model has to know what it sees means. The results showed a persistent gap between the best models and expert human performance, especially in disciplines where the image is not illustration but contains the decisive evidence: circuit diagrams, laboratory graphs and radiographs Yue et al., 2023.
Long-duration video. Video-MME, published in 2024, evaluated video understanding over durations ranging from minutes to hours, with questions requiring temporal tracking, analysis of changes between segments and synthesis of information distributed throughout the video. The evaluation revealed a pronounced quality drop as duration increases: models that understand short videos well fail on long versions of the same tasks because the attention mechanism loses temporal coherence at the scale of minutes or hours, a limitation that image or short-video benchmarks do not capture Fu et al., 2024. ZeroBench, published in February 2025, highlighted another angle of the problem: it evaluated 20 frontier models on one hundred visual-spatial cognition tasks over static images, and all obtained 0.0% accuracy Roberts et al., 2025. This is not temporal coherence but something more basic: pure spatial reasoning in scenes that any three-year-old can solve effortlessly systematically exceeds what any current model can do. LVOmniBench, introduced in 2026, confirmed the pattern for long-duration real-world video between 10 and 90 minutes: every open-source model remains below 35% accuracy, with the best evaluated commercial model reaching only 65%.
Video-MME: video understanding does not scale with duration
Models that understand short videos well fail on longer versions of the same tasks. Attention coherence degrades over time scales of minutes or hours.
Accuracy on temporal questions
78%
best Video-MME result (short segment)
0:001:30
active attention window
Temporal attention density
Situation
All information fits in the active context
The model can attend to any frame in the video with similar effectiveness. Questions that require linking the beginning and end of the clip remain manageable because the token distance is small.
Tasks handled well
Identify the object that appears first
Count actions in the clip
Describe the sequence of events
Check whether something changes between beginning and end
Accuracy on temporal questions
54%
24 pp drop versus short video
0:007:3015:00
active context (latest frames)
Temporal attention density
Situation
The context does not cover the entire video
The model has to compress or discard early frames to keep the video within context. Questions that link information from the beginning and end of the video fail because one of the references has been compressed or discarded.
Degraded performance
Describe the second half of the video
Compare something from the beginning with the end
Count how many times X appears throughout the video
Check whether the initial state was preserved
Accuracy on temporal questions
38%
40 pp drop versus short video · barely above chance level
0:0030 min60 min
end only
Temporal attention density
Situation
Most of the video is inaccessible
The model only has coherent access to the last few minutes of the video. Questions about earlier events are answered from prior assumptions or fabricated. The model does not know what it has lost: it produces answers with the same confidence as on short videos.
Critical performance
Summarize the entire video
Identify the character who appears at minute 12
Check whether a statement is consistent with the entire video
Describe how a topic evolves over the course of the hour
Implication for production systems
Systems that analyze long meetings, films, or full conferences operate in this regime. Short-video benchmarks do not predict their real performance. A model scoring 78% on short clips can be close to chance level on hour-long videos.
Visual hallucinations. HallusionBench, published in 2023, was designed to detect hallucinations specific to vision-language systems: cases where the model claims to see absent elements, denies the presence of visible elements, or assigns incorrect spatial relations to objects that it can identify individually. The results showed that visual hallucination is a consistent pattern across all evaluated models rather than a marginal phenomenon, and that frequency varies by task type—counting, spatial reasoning, existence—in such a way that no model is robust across all categories at once Liu et al., 2023.
HallusionBench: three patterns of visual hallucination
Visual hallucination is not a marginal phenomenon. HallusionBench identified three distinct, consistent patterns across all evaluated models, including the most capable ones.
Actual image
There is only one glass in the image
Question
"Is there a book in the image?"
Hallucinating model
"Yes. On the left side of the image there is a red-covered book leaning against the wall."
hallucination · nonexistent element
Correct answer
"No. There is no book in the image. Only a glass appears on the table."
correct · genuine grounding
Underlying mechanism
The model generates an object that is plausible for the described scene (table + glass → book is coherent with a domestic setting) using its prior over typical scenes, without checking the visual evidence. The answer is fluent and convincing.
Actual image
Document with clearly visible figures
Question
"Does any VAT figure appear in this document?"
Hallucinating model
"The image does not contain numerical information clear enough to identify VAT figures with certainty."
hallucination · present element denied
Correct answer
"Yes. The document shows VAT of €388.97 on a taxable base of €1,458.53."
correct · genuine grounding
Underlying mechanism
The model expresses uncertainty about content that is actually present, possibly because the document uses typography or a layout outside its training distribution. The hallucination is inverted: instead of generating something that does not exist, it denies something that does.
Actual image
The cube is above the sphere
Question
"Which object is on top: the cube or the sphere?"
Hallucinating model
"The sphere is on top of the cube. The blue sphere rests on the upper surface of the green cube."
hallucination · inverted spatial relation
Correct answer
"The cube is on top. The sphere is below, and the cube rests on it."
correct · genuine grounding
Underlying mechanism
Spheres are lighter and are associated with being on top in the training corpus (balloons, balls on surfaces). The statistical prior for "sphere above" is stronger than the visual evidence of the actual arrangement, especially when both objects are individually recognizable.
Pattern shared by all three types
1
Claim absent
2
Deny present
3
Incorrect spatial relation
In every case, the language prior outweighs the visual evidence. The answer sounds confident and fluent even when it is factually wrong. Standard VQA benchmarks do not detect these patterns because they are not designed to do so.
Multimodal outputs. There are no established benchmarks that adequately measure real-time spoken-response quality, consistency between text and voice generated simultaneously, or the accuracy of images generated conditioned on both text and an input image. This absence of metrics means that we do not know precisely where the current limits of systems operating in that space lie.
Next chapter
Chapter 5 — Risks → — Which risks are specific to multimodality, why poor grounding has different consequences depending on the modality, and how the risk profile changes when perception and action are coupled in the same system.
Benchmark for text localization, handwriting and logical reasoning over documents.
R6
Sakshi et al. (2024) — MMAU: A Massive Multi-Task Audio Understanding and Reasoning Benchmark (Adobe Research)
Audio-understanding and reasoning benchmark covering speech, non-verbal sounds and music.
R7
Yue et al. (2023) — MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI (arXiv)
Visual-reasoning benchmark across 57 university disciplines with a persistent gap from expert humans.
R8
Fu et al. (2024) — Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis (arXiv)
Video-understanding benchmark spanning durations from minutes to hours.
R9
Liu et al. (2023) — HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models (arXiv)
Benchmark specifically designed to detect visual hallucinations in vision-language systems.
R10
Ouyang et al. (2024) — OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations (arXiv)
PDF-parsing benchmark with complex layouts; presented at CVPR 2025.
R11
Roberts et al. (2025) — ZeroBench: An Impossible Visual Benchmark for Contemporary Large Multimodal Models (arXiv)
Impossible benchmark: 20 frontier models evaluated, all at 0.0% on visual-spatial cognition.
Why can a model score highly on a visual benchmark without actually processing the image?
Because of language bias: the model answers from the statistical distribution of likely responses given the question text, not from the visual content. The effect is invisible when that statistical bias matches the correct answer because the model is right for the wrong reasons and an accuracy benchmark cannot distinguish it from a model that did process the image.
How can you detect whether a system really reasons over the temporal sequence of a video?
With what the article describes as the shuffling test: pass the video frames to the model in random order before asking the question. If the score does not change or even improves after shuffling, the system does not have dynamic temporal reasoning and is answering from semantic cues present in individual frames rather than from the sequence.
What makes MMMU more demanding than image-description benchmarks?
MMMU uses university-exam questions across 30 subjects where the image is not an illustration but contains the decisive evidence: circuit diagrams, laboratory graphs and radiographs. Reading the image correctly is not enough; the model has to know what it sees means. The best current models still remain significantly below expert human performance in the most technical categories.
Why does model performance fall so sharply when moving from standard text to documents with complex layouts?
Because their representations are optimized for natural photographs, not for integrating spatial structure and semantics at the same time. OCRBench v2 and OmniDocBench show that systems with 80–90% accuracy on standard text fall to 36.9% on reconstruction of layouts with multiple columns, tables with merged cells or formulas embedded in text flow.