CAS-SIAT-XinHai/FontHalu

Font-based Hallucinations in Multimodal Large Lange Models

Benchmark Construction

The benchmark construction process, illustrated in Figure 2, consists of four key stages: font selection, image template creation, content generation, and benchmark assembly.

Font Selection. In this stage, we collect a diverse set of fonts, including both commonly used and rare artistic or commercial fonts, to ensure comprehensive coverage of font variations. Ultimately, we gather 19,274 distinct fonts, each applied to render identical content on 250 × 1500 pixel blank images. The content consists of two lines of Lorem Ipsum text and one line of numerical data, generating a total of 19,274 images. These images are then processed using a pre-trained ViT model (ViT-Base-Patch16-224) [8 ] to extract feature vectors from the final hidden layer. To reduce dimensionality, we apply UMAP [27], followed by K-means clustering [24] to group the feature vectors into 200 clusters. From each cluster, five representative fonts are selected, yielding a total of 1,000 candidate fonts for the benchmark. For a detailed explanation of the font selections algorithm, see appendix A.

Template Creation. In this phase, 200 images are selected from publicly available datasets, including DocVQA [26], FUNSD [ 13] and SROIE [ 12]. We then annotate each image with 2–3 bounding boxes. The annotation criteria are as follows:

• The bounding box (bbox) should ensure that the corresponding keyword is clearly visible, facilitating subsequent queries about the content within the bbox based on the keyword.
• The bounding box should align as closely as possible with the original content’s location, avoiding discrepancies with the actual key content.
• When annotating, the size of the bounding box should match the actual content’s dimensions, avoiding excessive cropping or expansion.
• The bounding box should not obscure any crucial information, ensuring that all key content is fully presented.

To facilitate controlled content placement, we binarize these images by removing the content within the annotated regions while preserving the location information. This process results in 200 annotated template images.

Content Generation. In this phase, we design various data types based on the annotated bounding boxes, including Lorem Ipsum text, numbers, dates, times, names, addresses, email addresses, etc. These content types are selected to reflect diverse real-world text data, ensuring broad scenario coverage. We utilized a data generation library to auto-generate diverse text contents. For each task, we follow specific content generation strategies:

• VarRec Task: The selected fonts are used to write Lorem Ipsum content on blank images.
• SemAna Task: Questions are first generated based on the key information associated with the annotated bounding boxes in the image templates. Then, specific content is written in these bounding boxes using various fonts while applying one of four randomly selected formatting modes: strikethrough, highlighting, overlay, or normal. These modified texts serve as our ground-truth answers.
• SentAna Task: The test set of the Sentiment1401 dataset is rendered in different fonts.

Finally, we manually verify the dataset, removing low-quality samples to ensure data integrity. The filtering criteria are as follows: (1) Images with truncated content are filtered out. (2) Images with garbled content, where the corresponding font fails to render ASCII characters, are excluded. Images where the content is unclear or fuzzy (i.e., human-readable keywords are difficult to discern) are removed. (3) During the verification process, three reviewers assess the image, and it is only discarded if all three reviewers agree that the content is unreadable.

Task	Images	Questions
VarRec	250	250
SemAna	780	1877
-Normal	589	928
-Strikethrough	124	132
-Overlay	317	393
-Highlight	343	424
SentAna	200	200

Overall performance of models

Model	Creator	Access	VarRec	SemAna	SentAna	ACC
Qwen2-VL-7B	Alibaba	Weights	0.6280	0.6819	0.5690	0.6649
Qwen2.5-VL-7B	Alibaba	Weights	0.4560	0.6622	0.6120	0.6352
Gemini-Pro-1.5	Google	API	0.2680	0.6718	0.5315	0.6148
Gemini-Pro-Vision	Google	API	0.0480	0.6548	0.5696	0.5816
GPT-4o-2024-08-06	OpenAI	API	0.4280	0.5717	0.6702	0.5661
MiniCPM-V-2.6	OpenBMB	Weights	0.3640	0.5876	0.5510	0.5601
Glm-4v-9B	Zhipu AI	Weights	0.3440	0.5850	0.5918	0.5599
Phi-3-Vision-128k-Instruct	Microsoft	Weights	0.1080	0.6244	0.4577	0.5527
InternVL2.5-1B	Shanghai AI Lab	Weights	0.2160	0.6095	0.4504	0.5517
InternVL2.5-2B	Shanghai AI Lab	Weights	0.2120	0.6137	0.4136	0.5509
Qwen2-VL-72B-Instruct	Alibaba	Weights	0.3760	0.5637	0.6259	0.5499
InternVL2-4B	Shanghai AI Lab	Weights	0.1880	0.5855	0.4597	0.5306
InternVL2.5-8B	Shanghai AI Lab	Weights	0.2080	0.5690	0.5227	0.5259
InternVL2.5-4B	Shanghai AI Lab	Weights	0.2240	0.5738	0.4599	0.5251
Minimax-01	Minimax	API	0.2533	0.5431	0.3305	0.4910
Pixtral-12B	Mistral AI	API	0.2960	0.4997	0.5916	0.4872
InternVL2-8B	Shanghai AI Lab	Weights	0.1480	0.5328	0.4726	0.4857
InternVL2-1B	Shanghai AI Lab	Weights	0.1320	0.5605	0.2608	0.4849
InternVL2-2B	Shanghai AI Lab	Weights	0.1160	0.5429	0.3669	0.4798
CogVLM2-LLaMA3-19B-Chat	Zhipu AI	Weights	0.0840	0.5003	0.5574	0.4616
LLaMA3.2-90B-Vision-Instruct	Meta	API	0.0840	0.4758	0.4947	0.4359
Phi3.5-Vision-Instruct	Microsoft	Weights	0.1200	0.4513	0.5091	0.4218
Ovis1.5-LLaMA3-8B	Alibaba	Weights	0.1200	0.4156	0.4541	0.3879
LLaVA-V1.6-Vicuna-7B-hf	UW–Madison	Weights	0.0120	0.4182	0.4744	0.3805
LLaMA3-LLaVA-Next-8B-hf	UW–Madison	Weights	0.0240	0.3953	0.4856	0.3647
LLaVA-V1.6-Mistral-7B	UW–Madison	Weights	0.0160	0.3937	0.4567	0.3597
LLaMA3.2-11B-Vision-Instruct	Meta	API	0.1440	0.3031	0.3077	0.2866
Deepseek-VL-7B-Chat	DeepSeek	Weights	0.0760	0.2339	0.4026	0.2339
InstructBLIP-Vicuna-7B	Salesforce	Weights	0.0000	0.0027	0.3285	0.0349
BLIP-VQA-Base	Salesforce	Weights	0.0000	0.0000	0.2844	0.0284

Acknowledgement

1. zi2zi
2. IAM Handwritten Forms Dataset
3. SynthText
4. LAM The Ludovico Antonio Muratori (LAM) dataset is the largest line-level HTR dataset to date and contains 25,823 lines from Italian ancient manuscripts edited by a single author over 60 years.
5. MJSynth
6. SROIE2019
7. CASIA-HWDB
8. Laia: A deep learning toolkit for HTR