We present THEMIS, a novel multi-task benchmark designed to comprehensively evaluate multimodal large language models (MLLMs) on visual fraud reasoning within real-world academic scenarios. Compared to existing benchmarks, THEMIS introduces three major advances. (1) Real-World Scenarios and Complexity: Our benchmark comprises over 4,000 questions spanning seven scenarios, derived from authentic retracted-paper cases and carefully curated multimodal synthetic data. With 60.47% complex-texture images, THEMIS bridges the critical gap between existing benchmarks and the complexity of real-world academic fraud. (2) Fraud-Type Diversity and Granularity: THEMIS systematically covers five challenging fraud types and introduces 16 fine-grained manipulation operations. On average, each sample undergoes multiple stacked manipulation operations, with the diversity and difficulty of these manipulations demanding a high level of visual fraud reasoning from the models. (3) Multi-Dimensional Capability Evaluation: We establish a mapping from fraud types to five core visual fraud reasoning capabilities, thereby enabling an evaluation that reveals the distinct strengths and specific weaknesses of different models across these core capabilities. Experiments on 16 leading MLLMs show that even the best-performing model, GPT-5, achieves an overall performance of only 56.15%, demonstrating that our benchmark presents a stringent test. We expect THEMIS to advance the development of MLLMs for complex, real-world fraud reasoning tasks.
Figures & tables
Figure 1: Overview of THEMIS. (1) Real-World Scenarios and Complexity , comprising over 4,000 questions across 7 representative scenarios; (2) Fraud-Type Diversity and Granularity , covering 5 challenging fraud methods with 16 fine-grained manipulation operations; (3) Multi-Dimensional Capability Evaluation , mapping fraud types to 5 core reasoning capabilities.
Table 1: Comparison of THEMIS and other related benchmarks. Mod. : Modality; Manip. : Manipulation Operations; Real : Contains real fraud cases; I : Image; T-I : Cross Text–Image; Task : BC : Binary Classification; IP-L : Image Partial Localization. SMF , CMO , and CMI denote Single-Mode Forgery, Composite Manipulation Operations, and Cross-Modal Inconsistency, respectively, with suffixes -I and -L indicating Identification and Localization ( -IL for both). Fraud Type : 1 Splicing, 2 Copy-Move, 3 AI-Generated, 4 Duplication, and 5 Text–Image Inconsistency.
Figure 2: Statistics of THEMIS. (a) Distribution of fraud methods. (b) Distribution of manipulation operations (synthetic data). See Table 5 for real cases. IIF : Image Inference Forgery; TRE : Targeted Region Editing; CT : Color Temperature; DR : Direct Reuse; HF : Horizontal Flip; VF : Vertical Flip.
Figure 3: Dataset construction pipeline of THEMIS. The dataset is built through 2 stages: Stage 1: Extraction and Parsing , where figures, captions, and related sentences are parsed from scientific PDFs and segmented into panels; Stage 2: Fraud Data Generation , where 5 major fraud types ( Splicing , Copy-Move , AI-Generated , Duplication , and Text–Image Inconsistency ) are applied to construct challenging tasks.
Figure 4: Evaluation task design of THEMIS. A principled mapping from 5 fraud types to 5 core reasoning capabilities ( Expert Knowledge Utilization , Visual Recognition , Spatial Reasoning , Region Localization , and Comparative Reasoning ). The capability distribution bars on the right of each box illustrate the reasoning skills involved and their relative emphasis, with the darkest color highlighting the primary capability being evaluated.
Model
Single-Mode Forgery Identification (Id Score)
Single-Mode Forgery Localization (Loc Score)
Composite Manipulation Operations Identification (Id Score)
Table 2: Model performance across tasks on THEMIS synthetic data. We report the Id Score (Identification Score) and Loc Score (Localization Score) for 5 fraud methods: SPL (Splicing), CM (Copy-Move), AIG (AI-Generated), DUP (Duplication), and TII (Text–Image Inconsistency), with BRI (Balanced Robustness Index). The best and second-best scores are highlighted.
Figure 5: Impact of different factors on model performance using synthetic data. SPL : Splicing; CM : Copy-Move; AIG : AI-Generated; DUP : Duplication; DUP-Id : Id Score for Duplication; TII-Id : Id Score for Text–Image Inconsistency; TII-Loc : Loc Score for Text–Image Inconsistency; SMF-Id : Average Id Score of Single-Mode Forgery Identification; SMF-Loc : Average Loc Score of Single-Mode Forgery Localization.
Model
SMF-IL
CMO-I
Id Score
Loc Score
Id Score
Qwen2.5-VL-72B
50.00
45.41
22.39
Gemini 2.5 Flash
44.55
43.07
49.58
GPT-5
43.64
28.68
23.22
OpenAI o4-mini-high
34.55
21.67
36.04
Gemma 3 27B
20.00
16.15
42.18
Table 3: Performance comparison on real cases. We report the results for Single-Mode Forgery Identification and Localization ( SMF-IL ) and Composite Manipulation Operations Identification ( CMO-I ) tasks.
Appendix figures & tables37 assets
Supplementary material from the paper’s appendix.
Appendix
Table 9
Figure 6: Distribution and visualization of 7 scenarios.
Scope
Fraud Type
Manipulation Operation
# Samples
Intra-Element
Copy-Move
Copy-Move
92
AI-Generated
Image Inference Forgery
74
Targeted Region Editing
550
Inter-Element
Splicing
Splicing
534
Duplication
Global Direct Reuse
104
Global 180° Rotation
283
Appendix
Table 6: Detailed distribution of fraud types and manipulation operations across the synthetic data.
Figure 7: Stage 1: Extraction and Parsing. The regions highlighted with red boxes correspond to figures , while those highlighted with blue boxes correspond to panels . Green boxes are used to verify whether the visualization is effective.
Provider
Version
Parameters
Access Links
Proprietary MLLMs
OpenAI
GPT-5
N/A
https://openrouter.ai/openai/gpt-5
OpenAI o4-mini-high
N/A
https://openrouter.ai/openai/o4-mini-high
Google
Gemini 2.5 Flash
N/A
https://openrouter.ai/google/gemini-2.5-flash
Gemini 2.5 Pro
N/A
https://openrouter.ai/google/gemini-2.5-pro
Alibaba
Qwen-VL-Max
N/A
https://openrouter.ai/qwen/qwen-vl-max
Appendix
Table 7: Detailed information on benchmarked MLLMs.
Model
λ=0.10
λ=0.25
λ=0.30
λ=0.40
GPT-5
64.56
56.15
53.35
47.75
OpenAI o4-mini-high
63.57
52.34
48.60
41.11
Qwen-VL-Max
57.31
49.83
47.34
42.35
Gemini 2.5 Flash
57.67
44.70
40.36
31.71
Doubao-Seed-1.6-thinking
48.43
37.14
33.37
25.85
Doubao-Seed-1.6-vision
46.17
33.47
29.24
20.78
Appendix
Table 8: Sensitivity of the BRI under different values of λ .
Model
Splicing
ACC
IoU
Set-based F1
Proprietary MLLMs
GPT-5
43.51
16.67
20.27
OpenAI o4-mini-high
41.49
10.44
12.81
Qwen-VL-Max
30.37
40.07
52.84
Gemini 2.5 Flash
63.39
56.72
65.08
Appendix
Table 9: Experimental results on THEMIS synthetic data for Splicing. We report the ACC , IoU , and Set-based F1 . The best and second-best scores are highlighted.
Model
Copy-Move
ACC
IoU
Set-based F1
Proprietary MLLMs
GPT-5
72.73
36.41
46.11
OpenAI o4-mini-high
77.89
29.78
39.32
Qwen-VL-Max
87.40
48.34
60.94
Gemini 2.5 Flash
67.56
46.98
57.90
Appendix
Table 10: Experimental results on THEMIS synthetic data for Copy-Move. We report the ACC , IoU , and Set-based F1 . The best and second-best scores are highlighted.
Model
AI-Generated
ACC
IoU
Set-based F1
Proprietary MLLMs
GPT-5
44.26
19.33
22.21
OpenAI o4-mini-high
35.67
19.79
22.86
Qwen-VL-Max
35.43
35.85
43.87
Gemini 2.5 Flash
35.45
38.87
44.76
Appendix
Table 11: Experimental results on THEMIS synthetic data for AI-Generated. We report the ACC , IoU , and Set-based F1 . The best and second-best scores are highlighted.
Model
Duplication
Set-based F1
EM
SPC
Proprietary MLLMs
GPT-5
33.32
19.05
22.72
OpenAI o4-mini-high
30.34
17.12
21.35
Qwen-VL-Max
23.33
12.13
15.29
Gemini 2.5 Flash
24.96
15.78
17.41
Appendix
Table 12: Experimental results on THEMIS synthetic data for Duplication. We report the Set-based F1 , EM (Exact Match), and SPC (Subset Partial Credit). EM requires strict equality with the ground truth; SPC awards credit for correct subsets only if no incorrect elements (false positives) are predicted. The best and second-best scores are highlighted.
Model
Text–Image Inconsistency
ACC
Macro-F1
Text-based F1
Proprietary MLLMs
GPT-5
60.67
44.95
27.44
OpenAI o4-mini-high
66.33
48.58
32.22
Qwen-VL-Max
56.00
33.74
15.36
Gemini 2.5 Flash
36.33
25.30
28.24
Appendix
Table 13: Experimental results on THEMIS synthetic data for Text–Image Inconsistency. We report the ACC , Macro-F1 , and Text-based F1 . The best and second-best scores are highlighted.
Figure 8: Success case by GPT-5 on Splicing.
Figure 9: Failure case by GPT-5 on Splicing. The model correctly identified the forgery type but mislocalized it.
Figure 10: Failure case by GPT-5 on Splicing. The model misidentified and mislocalized.
Figure 11: Success case by GPT-5 on Copy-Move.
Figure 12: Failure case by GPT-5 on Copy-Move. The model correctly identified the forgery type but mislocalized it.
Figure 13: Failure case by GPT-5 on Copy-Move. The model correctly localized the forgery region but misidentified it.
Figure 14: Failure case by GPT-5 on Copy-Move. The model misidentified and mislocalized.
Figure 15: Success case by GPT-5 on AI-Generated. The panel is entirely forged via Image Inference Forgery.
Figure 16: Failure case by GPT-5 on AI-Generated. The model misidentified the forgery type. The panel is entirely forged via Image Inference Forgery.
Figure 17: Success case by GPT-5 on AI-Generated. The panel is partially forged via Targeted Region Editing.
Figure 18: Failure case by GPT-5 on AI-Generated. The model correctly identified the forgery type but mislocalized it. The panel is partially forged via Targeted Region Editing.
Figure 19: Failure case by GPT-5 on AI-Generated. The model correctly localized the forgery region but misidentified it. The panel is partially forged via Targeted Region Editing.
Figure 20: Failure case by GPT-5 on AI-Generated. The model misidentified and mislocalized. The panel is partially forged via Targeted Region Editing.
Figure 21: Success case by GPT-5 on Duplication. These panels resulted from Global Duplication with flip.
Figure 22: Failure case by GPT-5 on Duplication. These panels resulted from Global Duplication with rotation and parameter modification.
Figure 23: Failure case by GPT-5 on Duplication. These panels resulted from Local Duplication with rotation. The duplicated region is indicated by the yellow boxes.
Figure 24: Failure case by GPT-5 on Duplication. These panels resulted from Local Duplication with parameter modification. The duplicated region is indicated by the yellow boxes.
Figure 25: Success case by GPT-5 on Text–Image Numerical Inconsistency.
Figure 26: Failure case by GPT-5 on Text–Image Numerical Inconsistency. The model identified but mislocalized.
Figure 27: Failure case by GPT-5 on Text–Image Numerical Inconsistency. The model misidentified and mislocalized.
Figure 28: Failure case by GPT-5 on Text–Image Trend Inconsistency. The model identified but mislocalized.
Figure 29: Failure case by GPT-5 on Text–Image Trend Inconsistency. The model misidentified and mislocalized.
Figure 30: Left: Extraction and Parsing by MinerU. The green highlighted regions indicate MinerU’s panel-level parsing outputs, while the blue regions represent its caption parsing results. MinerU fails to extract complete figures and frequently mismatches image–caption pairs. Right: Extraction and Parsing by our pipeline. The regions highlighted with red boxes correspond to figures , while those highlighted with blue boxes correspond to panels . ( Green boxes are used solely to verify the effectiveness of the visualization.) Our extraction pipeline correctly identifies both full figures and their corresponding panels without structural or visual degradation.
Figure 31: One missing panel highlighted with a yellow box.
Figure 32: Two missing panels highlighted with yellow boxes. Two extra panels highlighted with red boxes.
Figure 33: Three missing panels highlighted with yellow boxes. One extra panel highlighted with a red box.
Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.
Humans develop perception through a bottom-up hierarchy: from basic primitives and Gestalt principles to high-level semantics. Current Multimodal Large Language Models (MLLMs) are trained directly on complex downstream tasks. Does it guarantee their performance on these foundational visual capabilities? To systematically investigate this gap, we introduce VisFacor, a benchmark that digitizes 20 vision-centric subtests from FRCT, a well-established cognitive psychology assessment spanning four domains of human visual cognition. Furthermore, we design algorithms to automatically construct and validate unlimited test cases with controllable difficulty. Using VisFacor, we evaluate 39 frontier MLLMs, including both proprietary (e.g., GPT, Gemini) and open-source (e.g., LLaMA, Qwen) models. The best model achieves a score of only 55.9%. Analysis reveals good internal consistency (Cronbach's alpha = 0.94) and construct validity (compared to existing vision benchmarks). Models consistently fail on tasks such as mental rotation, spatial relation inference, and figure-ground discrimination, regardless of model size or prompting strategy. These findings suggest that performance improvements on existing general benchmarks might represent castles in the air instead of a genuine mastery of human-like visual cognition.
Large Multimodal Models (LMMs) exhibit shortfalls when interpreting images and, by some measures, have poorer spatial cognition than young children or animals. Despite this, they attain high scores on many popular visual benchmarks, with headroom rapidly eroded by model progress. This creates a need for difficult benchmarks that remain relevant for longer. We introduce ZeroBench - a lightweight visual reasoning benchmark curated using adversarial filtering to be "impossible" for frontier LMMs at its original release, with initial SotA scores of 0% pass@1 and pass^5. We track progress on ZeroBench over the subsequent year, observing SotA reaching 6% pass^5 and 19% pass@5, indicating the potential longevity of the benchmark. We evaluate 46 LMMs on ZeroBench, compare performance to a human baseline, analyse strengths and weaknesses, chart a year of progress in visual capabilities, and publicly release ZeroBench at https://zerobench.github.io.
Jonathan Roberts, Mohammad Reza Taesiri, Ansh Sharma +31
Recent multimodal large language models (MLLMs) achieve strong performance on visual reasoning benchmarks, yet it remains unclear to what extent such performance reflects reasoning directly grounded in visual evidence. We introduce VisReason, a benchmark for vision-centric reasoning in everyday scenarios where perception and inference are tightly coupled. VisReason contains 1,505 questions across 10 categories spanning perceptual, structural, and conceptual reasoning. Our evaluation shows that VisReason poses a qualitatively different challenge from existing benchmarks, exposing substantial gaps between humans and current MLLMs and revealing limited benefits from test-time reasoning strategies. VisReason offers a focused diagnostic for evaluating vision-centric reasoning beyond language.
Longteng Guo, Yifan Wang, Pengkang Huo +4
Institute of Automation, Chinese Academy of Sciences · School of Artificial Intelligence, University of Chinese Academy of Sciences