Paper deep dive
Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning
Rui Song, Lida Shi, Ruihua Qi, Yingji Li, Hao Xu
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/14/2026, 2:32:40 AM
Summary
The paper introduces GEVO, a glyph-driven fine-tuning framework designed to enhance Multimodal Large Language Models (MLLMs) in analyzing the evolution of ancient Chinese scripts. The authors construct a comprehensive benchmark with 11 tasks and over 130,000 instances, evaluating 19 MLLMs. Results show that while existing models struggle with ancient script recognition, fine-tuning significantly improves their performance in evolutionary reasoning and script style identification.
Entities (4)
Relation Signals (2)
GEVO → enhances → MLLMs
confidence 100% · we propose a glyph-driven fine-tuning framework (GEVO) that... enhances their understanding of text evolution.
MLLMs → evaluatedon → Benchmark
confidence 95% · we construct a comprehensive benchmark comprising 11 tasks... to evaluate the capability of MLLMs
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In recent years, rapid advances in Multimodal Large Language Models (MLLMs) have increasingly stimulated research on ancient Chinese scripts. As the evolution of written characters constitutes a fundamental pathway for understanding cultural transformation and historical continuity, how MLLMs can be systematically leveraged to support and advance text evolution analysis remains an open and largely underexplored problem. To bridge this gap, we construct a comprehensive benchmark comprising 11 tasks and over 130,000 instances, specifically designed to evaluate the capability of MLLMs in analyzing the evolution of ancient Chinese scripts. We conduct extensive evaluations across multiple widely used MLLMs and observe that, while existing models demonstrate a limited ability in glyph-level comparison, their performance on core tasks-such as character recognition and evolutionary reasoning-remains substantially constrained. Motivated by these findings, we propose a glyph-driven fine-tuning framework (GEVO) that explicitly encourages models to capture evolutionary consistency in glyph transformations and enhances their understanding of text evolution. Experimental results show that even models at the 2B scale achieve consistent and comprehensive performance improvements across all evaluated tasks. To facilitate future research, we publicly release both the benchmark and the trained models\footnote{this https URL}.
Tags
Links
- Source: https://arxiv.org/abs/2604.11299v1
- Canonical: https://arxiv.org/abs/2604.11299v1
Trouble viewing inline? Open PDF directly →
Full Text
63,495 characters extracted from source content.
Expand or collapse full text
Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning Rui Song 1,2 , Lida Shi 2,3 , Ruihua Qi 2,4 , Yingji Li 1 , Hao Xu 1,2 * 1 College of Computer Science and Technology, Jilin University, China 2 Key Laboratory of Ancient Chinese Script, Culture Relics and Artificial Intelligence, Jilin University, China 3 School of Artificial Intelligence, Jilin University, China 4 School of Archaeology, Jilin University, China songrui,yingjili,xuhao@jlu.edu.cn, shild21,qirh20@mails.jlu.edu.cn Abstract In recent years, rapid advances in Multimodal Large Language Models (MLLMs) have in- creasingly stimulated research on ancient Chi- nese scripts. As the evolution of written char- acters constitutes a fundamental pathway for understanding cultural transformation and his- torical continuity, how MLLMs can be system- atically leveraged to support and advance text evolution analysis remains an open and largely underexplored problem. To bridge this gap, we construct a comprehensive benchmark compris- ing 11 tasks and over 130,000 instances, specif- ically designed to evaluate the capability of MLLMs in analyzing the evolution of ancient Chinese scripts. We conduct extensive eval- uations across multiple widely used MLLMs and observe that, while existing models demon- strate a limited ability in glyph-level compari- son, their performance on core tasks—such as character recognition and evolutionary reason- ing—remains substantially constrained. Moti- vated by these findings, we propose a glyph- driven fine-tuning framework (GEVO) that ex- plicitly encourages models to capture evolution- ary consistency in glyph transformations and enhances their understanding of text evolution. Experimental results show that even models at the 2B scale achieve consistent and compre- hensive performance improvements across all evaluated tasks. To facilitate future research, we publicly release both the benchmark and the trained models 1 . 1 Introduction As carriers of Chinese cultural heritage, a deep un- derstanding of ancient scripts plays a crucial role in preserving and promoting China’s rich cultural legacy (Liu et al., 2025). In recent years, with the rapid advancement of Multimodal Large Language Models (MLLMs), a growing body of research has begun to leverage their capabilities to support * Corresponding author. 1 https://github.com/songruiecho/GEVO the study of ancient scripts, ranging from charac- ter identification to critical interpretation, thereby demonstrating the significant potential of MLLMs in this domain (Yao et al., 2025; Qiao et al., 2025; Cao et al., 2025; Li et al., 2025a). The understanding of ancient scripts is inher- ently tied to the analysis of character evolution, as writing styles from different historical periods exhibit intrinsic structural and semantic connec- tions (Bökset, 2006). Existing studies have ex- plored the evolutionary analysis of ancient scripts, with particular attention to the transformation of character forms from oracle bone inscriptions to regular script (Wang et al., 2022; Guan et al., 2024a; Jiao et al., 2025). However, as MLLMs become increasingly influential in related fields, systematic research on how to evaluate their capabilities in evo- lutionary analysis and how to effectively enhance these abilities remains limited. To overcome the aforementioned gap, this paper comprehensively considers multiple datasets for the analysis of glyph evolution mentioned above, and reconstructs a dataset for the evolutionary anal- ysis at the facsimile level, comprising 7,740 Chi- nese characters and nearly 30,000 corresponding facsimiles. Furthermore, to assess the capability of MLLMs in tasks related to the evolutionary analy- sis of ancient scripts, we construct a detailed eval- uation benchmark based on the aforementioned dataset, with the assistance of paleography experts. This benchmark comprises three main tasks, 11 subtasks, and over 130,000 specific questions. Moreover, we evaluate the benchmark on 19 MLLMs ranging from 1B to 72B in scale and find that existing models exhibit limited capabilities in script style recognition cross-era and ancient script recognition, leading to weak fundamental performance in evolutionary analysis. Nonetheless, we find that with further fine-tuning on a small amount of data, MLLMs can exhibit substantial improvements in temporal attribution capability. arXiv:2604.11299v1 [cs.CL] 13 Apr 2026 Based on this observation, we gain confidence in the ability of MLLMs to handle evolutionary anal- ysis tasks and propose a glyph-driven contrastive fine-tuning method inspired by curriculum learning to encourage models to fully discriminate the subtle differences induced by glyph forms and historical periods. Fine-tuning results on 2B-scale models confirm that the proposed method achieves consis- tent performance improvements across all tasks. In summary, our core contributions are as follows: (i) We contribute a benchmark for evaluating the evolutionary analysis capabilities of MLLMs in Chinese ancient script. (i) We evaluate this benchmark on multiple MLLMs, confirming that the evolutionary analysis capabilities of existing MLLMs are weak, but the recognition ability for script styles can be easily improved through fine- tuning. (i) We propose a glyph-driven fine-tuning framework, which significantly enhances the per- formance of MLLMs. 2 Related Work AI-Driven Research in Ancient Chinese Charac- ters. With the continuous advancement of algorith- mic capabilities and computing resources, artificial intelligence has gradually permeated in the study of Chinese ancient scripts, which has significantly enhanced the efficiency of character interpretation, form comparison, and evolutionary analysis (Wen, 2011; Li et al., 2026). This not only reduces repeti- tive labor for experts but also expands the bound- aries of traditional research methods, offering new technological paradigms and research pathways for the development of related fields. Numerous bench- marks related to ancient Chinese characters, espe- cially focusing on oracle bone inscriptions (Guan et al., 2024b; Wang et al., 2024b; Zhou et al., 2025), have emerged. Given the pronounced long-tailed distribution and high noise levels characteristic of ancient scripts, research directions such as im- balanced learning (Li et al., 2023a,b, 2025b), de- noising (Shi et al., 2022), few-shot learning (Zhao et al., 2022), and cross-modal learning (Wang et al., 2024a) are also explored. MLLM-Driven Research in Ancient Chinese Characters. With the recent rapid development of MLLMs, related research has expanded from tradi- tional deep learning to multi-task integration and generalization based on MLLMs. Some studies have proposed benchmarks for evaluating MLLMs, encompassing a variety of tasks related to ancient scripts, particularly oracle bone inscriptions. These benchmarks have demonstrated that even the most advanced models still lack the capability to han- dle such tasks effectively (Chen et al., 2025; Liu et al., 2025). Therefore, some approaches enhance the capability for text-related tasks by fine-tuning MLLMs on domain-specific datasets (Cao et al., 2025). There are also methods that leverage the core capabilities of MLLMs to assist experts in interpreting ancient scripts (Qiao et al., 2025). Fur- thermore, some integrate models into agent frame- works, utilizing MLLMs to assist experts in tack- ling more complex tasks (Li et al., 2025a). Un- like previous studies, our primary contribution lies in constructing an evolutionary-related bench- mark to evaluate the capabilities of MLLMs in evolution-related tasks. Additionally, we propose a glyph-based fine-tuning approach that effectively enhances model performance. 3 Benchmark Construction 3.1 Construction of Glyph Evolution Dataset 01 02 03 01 02 0101 01 Chinese Character 日 sun oracle bone script bronze script seal script clerical script regular script Figure 1: The text evolution process of the character “日” (sun) and the organization of data. Construction of Basic Glyph Data. Our dataset is considered based on existing evolutionary anal- ysis datasets. We refer to the study (Wang et al., 2022) to divide the evolutionary process into five stages: Oracle Bone Script, Bronze Inscription, Seal Script, Clerical Script, and Regular Script. Subsequently, we apply the same stages to extract data from Vividict 2 and extract the tracing data of characters in different periods into a unified organizational structure. As shown in Figure 1, this approach results in the same character poten- tially having multiple written forms within a sin- gle period, thereby enhancing the richness of the corpus. Subsequently, we uniformly binarize all font files and manually filter out instances of miss- ing, corrupted, or blank data resulting from char- acter extraction. This process ultimately yields a 2 https://w.vividict.com/ dataset encompassing the evolutionary processes of 7,740 characters. It is important to emphasize that not all characters possess a complete evolu- tionary path, which corresponds to the practical issue of missing evolutionary trajectories in pale- ographic research. Construction of Benchmark. Subsequently, we aim to evaluate the evolution analysis capabilities of MLLMs. Therefore, with the assistance of paleographers, we abstract 11 sub- tasks into three major categories as shown in Fig- ure 2. To adapt to MLLMs, the inputs for all tasks are modeled as a hybrid form of text instructions and images, with the expected output being text. Subsequently, we design task-specific instruction prompts tailored to different MLLMs, framing the tasks in formats such as question answering, bi- nary judgment, and multiple-choice, thereby con- structing the final benchmark. To enable the bench- mark to better align with the fine-tuning–evaluation paradigm, we randomly split each task into training and test sets by 9:1. Specifically, during the instruction construction process, we consulted paleographers to define the general task scope. Subsequently, based on this task scope, we solicit suggestions from ChatGPT to generate a set of candidate instructions, which were then presented again to paleography experts to ensure professionalism. These instructions were subsequently input into multiple locally deployable MLLMs (theQwen3series andInternVLseries evaluated in this paper) to verify their ability to guide the MLLMs in producing standardized an- swers and reasonable outputs. For instructions that proved ineffective, we revise them based on expert feedback and MLLM responses, ultimately ensur- ing the task validity and model compliance of the instructions. This process is illustrated in Figure 3. 3.2 Dataset Statistics Figure 4 illustrates the data distribution of the benchmark across different ancient scripts and their corresponding tasks. Oracle bone inscriptions ac- count for the smallest portion of the dataset, as only slightly more than two thousand oracle bone characters have been deciphered and can be reli- ably mapped to modern Chinese characters (Huang et al., 2019). In contrast, regular script contains the largest number of images, since it is the clos- est in form to modern Chinese characters and thus more abundantly preserved. From a task perspec- tive, T2.1 includes the largest number of instances, as it is designed to comprehensively evaluate the image-level recognition capabilities of MLLMs. Conversely, T3.3 has the smallest data volume be- cause it requires a complete and well-documented evolutionary trajectory for each character, which is unavailable for many characters. Overall, the benchmark is constructed at a moderate and well- balanced scale, enabling a comprehensive and multi-dimensional assessment of MLLMs. 4 Evaluation of MLLMs To validate the capability of existing MLLMs in tasks related to evolutionary analysis, we employ 19 commonly used MLLMs to evaluate all of the aforementioned tasks on test sets as shown in Ta- ble 1. Based on the experimental results, we pro- vide the most important observations as follows: MLLMs have a certain ability to compare script styles and character glyphs, but its char- acter recognition capability is relatively poor. Compared to script style recognition (T1.1), the performance of text recognition tasks (T2.1) is gen- erally lower. This is because character recognition requires strong expert knowledge, and the existing models often lack this specialized expertise. When tasks are concretized to script style (T1.1, T1.2, T1.3) and glyph comparison and selection (T2.1, T2.2, T2.3), the performance often improves com- pared to simple recognition tasks. This indicates that, rather than expert knowledge-driven special- ized recognition tasks, MLLMs are more adept at distinguishing similarities and differences through comparison. Additionally, in some cases, such as theLLaVA-1.5series models, their performance is poor due to difficulties in understanding complex instructions in ancient script research. Additionally, for some closed-source models, they exhibit almost no character recognition capability or frequently refuse to provide valid responses, leading to an av- erage performance that is often inferior to that of local models. Compared with isolated glyph comparison and recognition, performing the same tasks within an explicit evolutionary context leads to improved performance, underscoring the im- portance of modeling the evolutionary process. In most cases, character recognition (T3.1) and script-style identification (T3.2) benefit from be- ing situated within the evolutionary process, as the model can leverage additional contextual informa- tion to support reasoning and evaluation. This ob- servation highlights the necessity of incorporating Task 1. Script Style Recognition Task 2. Character Recognition Task 3. Evolutionary Path Tasks T1.1Givenanimage,determineitsscriptstyle. T1.2Giventwodifferentimagesofthesamecharacter, determinewhethertheybelongtothesamescriptstyle. T1.3Giventwodifferentimagesofdifferentcharacters, determinewhethertheybelongtothesamescriptstyle. T1.4Givenmultipleimages,identifywhichcharacterdoes notbelongtothesamescriptstyleastheothers. A T2.2Giventwocharactersinthesamescriptstyle, determinewhethertheyarethesameChinesecharacter. T2.3Giventwocharactersindifferentscriptstyles, determinewhethertheyarethesamecharacter. T2.4Givencharactersfrommultiplescriptstyles,identify whichcharacterisdifferentfromtheothers. T3.1Giventheevolutionarypathofacharacter,identify theChinesecharacter. T3.2Givenascrambledevolutionarypath,arrangethemin thecorrectchronologicalorder. T3.3Givenamissingsegmentofthepath,fillinthemissing part. B C D C Yes No MLLMs oracle bone T2.1Givenanimageinanyscriptstyle,recognizethe Chinesecharacterswithinit. 日(sun) Yes No D 日(sun) A B C D E CBAED ? A B C D A Figure 2: The 11 basic tasks that constitute the evaluation of MLLMs, along with corresponding examples. Due to limited space, we have provided a detailed expansion of the text instruction formats for different tasks in Appendix A. MLLMs ChatGPT Expert Please provide several tasks most relevant to ancient Chinese character evolution analysis. Based on the relevant task descriptions, provide task- specific instructions for input to the MLLM. Please filter and modify the instructions returned by ChatGPT to ensure they adequately reflect the task information. Observe:Whether the MLLM can generate valid outputs based on the above instructions? Save the instructions. Optimize the instructions based on the output results of the MLLM, ensuring that the MLLM producesmore standardized outputs without altering the semantic meaning of the instructions. Figure 3: Task-related instruction construction and opti- mization process. evolutionary knowledge into ancient script analysis. Nevertheless, owing to the limited domain-specific knowledge of MLLMs in ancient script studies, the observed performance gains remain relatively modest. For T3.3, performance largely depends on the model’s ability to jointly compare glyph forms and semantic meanings; as a result, accuracy in many cases approaches random guessing (25%). This further emphasizes the need to equip MLLMs with more specialized capabilities for ancient script oracle bone bronze seal clerical regular Script 0 2000 4000 6000 8000 Numbers (a) Quantity of different script styles. T1.1 15000 (11.2%) T1.2 9667 (7.2%) T1.3 10000 (7.5%) T1.4 10000 (7.5%) T2.1 29860 (22.3%) T2.2 9740 (7.3%) T2.3 9024 (6.7%) T2.4 5717 (4.3%) T3.1 7006 (5.2%) T3.2 7006 (5.2%) T3.3 5717 (4.3%) T4.1 15000 (11.2%) (b) Distribution of benchmark questions. Figure 4: Benchmark quantity statistics. understanding. 4.1 Analysis of Results Based on the quantitative results above, we aim to further investigate how script style influ- ences predictions. Given that theQwenseries of models demonstrate the most outstanding per- formance on the evaluation benchmark, we fur- ther conduct an in-depth analysis of the predic- tion results from theQwen3-VL-2B-Instructand Qwen3-VL-8B-Instruct in Figure 5 and 6. Figure 5 illustrates the confusion in script styles across adjacent historical periods, particularly among oracle bone, bronze, and seal script, as well as between clerical script and regular script. This observation is well aligned with the prin- MLLMsT1.1T1.2T1.3T1.4T2.1T2.2T2.3T2.4T3.1T3.2T3.3Average Closed-source Models GPT4-all-mini30.85 51.09 48.304.000.3050.10 49.171.050.0019.23 24.1325.29 GPT5-mini24.65 46.84 46.30 11.500.0749.50 48.706.790.0013.60 25.7824.88 Gemini-3-Flash21.73 48.71 48.90 24.200.0348.46 52.27 25.173.7412.06 21.5227.89 Open-source Models TongGu-VL-2B-Instruct4.0750.78 52.90 15.008.6449.79 52.491.549.274.3926.5725.04 Qwen2.5-VL-7B-Instruct31.07 63.91 57.60 38.7023.51 75.46 74.20 42.8340.66 36.40 39.8647.65 Qwen2.5-VL-32B-Instruct24.00 55.33 56.80 29.9023.58 71.25 74.64 50.0039.80 27.72 56.2946.30 Qwen2.5-VL-72B-Instruct18.60 61.49 57.30 41.8024.45 73.31 74.86 54.7249.79 48.88 62.7651.63 Qwen3-VL-2B-Instruct17.81 55.74 51.90 26.7021.23 70.74 71.43 22.7327.39 16.67 44.9338.84 Qwen3-VL-8B-Instruct45.27 76.63 69.70 56.8030.74 74.13 81.62 69.9349.36 50.43 71.6861.48 Qwen3-VL-30B-A3B-Instruct32.81 66.94 58.02 45.3731.12 70.89 77.45 63.8455.91 36.02 64.2854.79 InternVL3_5-1B-HF7.6053.15 51.20 21.9017.35 46.71 43.08 11.7122.117.5733.0428.67 InternVL3_5-8B-HF17.33 58.53 52.00 28.014.6957.49 54.37 36.1915.98 18.98 40.2137.99 InternVL3_5-14B-HF7.3348.50 48.30 33.902.5555.95 54.04 74.4818.69 33.51 74.3041.05 MiniCPM-V-2_6 8B7.6755.02 51.90 28.108.8480.70 76.41 45.987.8516.00 25.8736.76 MiniCPM-V-4_5 8B21.00 62.67 54.10 36.009.3180.49 77.41 64.3413.84 25.83 25.7042.79 GLM-4.1V-9B-Thinking26.00 52.43 51.00 23.2018.05 46.30 43.30 11.3620.687.7334.0930.38 DeepSeekOCR-3B11.12 50.34 48.21 25.879.6451.02 56.03 21.1818.35 15.72 45.0832.05 LLaVA-1.5-7B-HF10.479.208.4026.100.0022.98 24.92 22.900.0012.36 26.2214.87 LLaVA-1.5-13B-HF3.3347.57 50.10 11.500.1759.24 61.465.940.006.5028.9624.98 GEVO*80.60 88.83 87.00 83.2039.18 89.32 92.03 96.8570.19 93.13 98.6083.54 Table 1: Performance comparison of different MLLMs across all test sets (accuracy %). The evaluation results on the full dataset are provided in Appendix C. oracle bone bronze seal clerical regular Predicted Label oracle bone bronze seal clerical regular True Label 1990280 13196350 11523772 04101970 00224110 Qwen3-VL-2B-Instruct oracle bone bronze seal clerical regular Predicted Label oracle bone bronze seal clerical regular True Label 207102709 1683180110 7366151017 45426249 0022284 Qwen3-VL-8B-Instruct Figure 5: The confusion matrices of script style predic- tions for different models, with unrecognizable cases removed to better highlight the relationships among glyphs across scripts and periods. ciples of script evolution, as scripts from adja- cent periods tend to exhibit highly similar stylistic characteristics. As for the sparsity observed in Qwen3-VL-2B-Instruct’s performance on oracle bone, bronze, and seal scripts, it occurs because the model tends to "decline to answer" when it is un- certain. This directly demonstrates the limitations in the capability of a 2B-scale model. Figure 6 reports the prediction accuracy across different script styles. A clear and intuitive trend can be observed: scripts that are closer to the mod- ern era consistently achieve higher recognition ac- curacy. This pattern is consistent with the evo- lutionary trajectory of Chinese characters, which gradually converge toward modern forms, thereby enabling MLLMs to more effectively transfer their intrinsic knowledge of contemporary Chinese char- acters. In contrast, the recognition accuracy for oracle bone bronze seal clerical regular 0 10 20 30 40 50 60 Accuracy (%) 6.67 10.47 15.89 35.36 37.61 Qwen3-VL-2B-Instruct oracle bone bronze seal clerical regular 0 10 20 30 40 50 60 Accuracy (%) 4.86 9.20 20.90 46.51 52.21 Qwen3-VL-8B-Instruct Figure 6: Characters accuracy on different script styles. T1.1 T1.2 T1.3 T1.4 T2.1 T2.2T2.3 T2.4 T3.1 T3.2 T3.3 68.33 80.14 74.20 68.70 22.57 82.41 82.61 87.76 41.51 86.70 92.66 Qwen3-VL-2B-InstructQwen3-VL- 8B-Instruct Qwen3-VL-2B-SFT 0 10 20 30 40 50 60 70 Average Accuracy(%) 38.84 61.48 71.60 Qwen3-VL-2B-Instruct Qwen3-VL-8B-Instruct Qwen3-VL-2B-SFT Figure7:Comparisonofresults forQwen3-VL-2B-InstructafterSFT (Qwen3-VL-2B-SFT) across all tasks. oracle bone script remains below 10%, indicating that current MLLMs possess little to no effective capability for recognizing this script. This observa- tion further underscores the necessity of endowing MLLMs with specialized knowledge to support research on ancient scripts. MLLM Vision Model image features feature space MLLM Vision Model CLIP Stage 1 Stage 3 Stage 2 This is “目” This is “日” ...... Figure 8: The two-stage framework of GDEVA training. In the first stage, characters in different script styles that represent the same character “日" (sun) are treated as a set of positive samples, while other images that are glyphically similar to them are considered negative samples. For example,andare considered glyphi- cally similar, but they are different ways of writing the characters for “日" (sun) and “目" (eye), respectively. Therefore, during the training process, we need to main- tain a distance between them. 4.2 Preliminary Attempt: Few-shot SFT To investigate the learning potential of MLLMs for tasks related to ancient text evolution analysis, we adoptQwen3-VL-2B-Instructand conduct sim- ple supervised fine-tuning (SFT) using 200 ran- domly sampled training examples per task. The SFT results are summarized in Figure 7. Overall, the fine-tuned model,Qwen3-VL-2B-SFT, exhibits substantial performance gains, with an average im- provement exceeding 30% over the original model, and even surpasses 8B-scale models by more than 10%. These results suggest that MLLMs can be effectively adapted to glyph comparison tasks for ancient scripts, as tracings of ancient characters are generally not visually complex. However, we also observe performance degrada- tion ofQwen3-VL-2B-SFTon T2.1 and T3.1. This indicates that a limited number of training samples is insufficient to support robust recognition of an- cient scripts, and may even induce catastrophic for- getting of previously acquired knowledge. Taken together, these findings reveal two key principles for guiding MLLM training: (i) a small number of samples can suffice to enhance the model’s ability to discriminate similar characters; and (i) comprehensive and sufficiently diverse data are essential to endow the model with stable and reliable recognition capabilities for writing sys- tems from temporally distant historical periods. 5 Glyph-driven Curriculum Learning 5.1 Model Framework Inspired by the results of our preliminary experi- ments, we propose a multi-stage fine-tuning frame- work as shown in Figure 8. The framework is built upon curriculum learning, where the model pro- gressively learns to understand evolutionary pro- cesses by tackling tasks of increasing complex- ity, from simple to difficult (Wang et al., 2021). In the first stage, we aim to model the glyph varia- tions of the same character across different script styles, while enhancing the alignability for visual representation. To this end, we independently fine- tune the visual module of MLLM, specifically up- dating the parameters of the visual encoder and the cross-modal projection module, so as to learn multimodal representations that are discriminative with respect to glyph variations while remaining semantically consistent. Specifically, we employ a contrastive learning approach (Chen et al., 2020) to encourage the rep- resentation optimization of the vision model within the MLLM. For each character, its script style- related imagesI n corresponding to different histori- cal periods are regarded as a set of positive samples P = I 1 , ...I n , as they all represent the same character. Additionally, to mitigate interference from visually similar glyphs of different charac- ters, we employ CLIP (Radford et al., 2021) to retrieve the top-kmost visually similar glyph im- ages that do not correspond to the target character, which are then used to construct negative samples N =¬I (1) 1 , ...,¬I (k) 1 , ...,¬I (1) n , ...,¬I (k) n . Here, ¬I (1) 1 represents the first negative sample ofI 1 , with a maximum ofk. Subsequently, the following contrastive learning loss is optimized to encourage the model to learn the similarities and differences between different images: L con =− 1 |P| X I i ∈P log S + i S + i +S − i ,(1) whereS + i = P I j ∈P,I j ̸=I i e s(z i ,z j ) τ andS − i = P I − ∈N i e s(z i ,z − ) τ . Here,s(z i ,z j )is cosine similarity between image representations, andz i ,z − repre- sent the representations ofI i and its corresponding negative sample, respectively. In the second stage, we aim for the model to fur- ther learn the mapping between images and text based on the glyphs it has already learned. Specifi- cally, given an image of a character from any his- torical script period, the model should predict the corresponding modern Chinese character. During this process, the parameters of the language model in the MLLMs are updated and keep the visual model parameters frozen, primarily capturing se- mantic associations. In the third stage, we fine-tune the language model in the MLLMs using task-related instruc- tions. Similarly, SFT is performed on a dataset containing only 200 samples per task to reduce the cost of fine-tuning. We name this glyph-driven evo- lutionary MLLM as GEVO. More details regarding model fine-tuning can be found in Appendix D. 5.2 Result Evaluation Table 2 reports a comparative evaluation of dif- ferent GEVO variants. Specifically,GEVO-Stage1 refers to the setting in which the model, af- ter completing training in Stage 1, proceeds di- rectly to supervised fine-tuning in Stage 3, while GEVO-Stage2denotes the variant in which glyph- based contrastive learning in Stage 1 is omitted and fine-tuning is conducted only in Stages 2 and 3. The experimental results demonstrate that, rela- tive to directly fine-tuningQwen3-VL-2B-SFT, both GEVO variants experience a clear degradation in overall performance. The results forGEVO-Stage1 indicate that glyph-driven contrastive learning ef- fectively enhances performance on glyph-centric tasks, particularly the T1 series and T3.2. This observation further underscores the critical role of glyph information in determining script style and historical period, and suggests that task designs em- phasizing glyph comparison can yield tangible per- formance gains. However,GEVO-Stage1exhibits a pronounced decline on character recognition tasks (the T2 series) as well as T3.1, with accuracy on T2.1 and T3.1 dropping below 10%. This behavior indicates that training focused exclusively on glyph- level signals induces catastrophic forgetting of the model’s already limited character recognition ca- pabilities. In contrast, the results forGEVO-Stage2 show that emphasizing recognition-oriented train- ing improves character identification, leading to no- table gains on T2.1 and T3.1. Nevertheless, these gains come at the cost of a severe degradation in glyph comparison performance, leading to a col- lapse in effectiveness on glyph-related tasks. GEVO effectively balances the model’s glyph comparison and character recognition capabilities, oracle bone bronze seal clerical regular Predicted Label oracle bone bronze seal clerical regular True Label 229351932 1288864250 152825960 13111520765 41041242 Qwen3-VL-2B-SFT oracle bone bronze seal clerical regular Predicted Label oracle bone bronze seal clerical regular True Label 173109213 662111990 33327110 09127724 00011277 GEVO Figure 9: The confusion matrices of script style predic- tions for Qwen3-VL-2B-SFT and GEVO. oracle bone bronze seal clerical regular 0 10 20 30 40 50 60 Accuracy (%) 1.65 6.68 13.53 35.17 38.73 Qwen3-VL-2B-SFT oracle bone bronze seal clerical regular 0 10 20 30 40 50 60 Accuracy (%) 6.60 14.56 32.68 56.97 58.21 GEVO Figure 10: Accuracy of characters across different script styles for Qwen3-VL-2B-SFT and GEVO. yielding substantial improvements across all tasks. Even when compared withQwen3-VL-2B-SFT, GEVO achieves an average performance gain ex- ceeding 10%. Moreover, it outperforms the base- line by more than 10% on both fundamental tasks, T1.1 and T2.1, demonstrating the effectiveness of the training methodology underlying GEVO. By si- multaneously enhancing glyph-level discrimination while preserving character recognition ability, this approach establishes stronger foundational com- petencies, which in turn translate into improved performance across a broader range of downstream tasks. But it is important to note that GEVO’s char- acter recognition performance (39.18%) remains relatively limited, indicating that substantial room for further improvement still exists in enhancing the character recognition capabilities of MLLMs. 5.3 Further Analysis Figure 9 presents the confusion matrix for script style recognition. Although oracle bone script and bronze script still exhibit a certain degree of confusion due to their high glyph-level similarity, the overall prediction quality is substantially im- proved. Compared with Figure 5, the number of correctly and reasonably predicted instances in- creases markedly, and the model no longer fails to respond to specialized queries. In addition, the confusion between seal script and clerical script MLLMsT1.1T1.2T1.3T1.4T2.1T2.2T2.3T2.4T3.1T3.2T3.3Average Qwen3-VL-2B-SFT68.33 80.14 74.20 68.7022.57 82.41 82.61 87.7641.51 86.70 92.6671.60 GEVO-Stage175.00 83.14 78.70 78.101.6462.01 47.84 26.754.4290.13 93.6558.31 GEVO-Stage232.20 57.19 50.40 32.4028.20 81.62 89.37 92.1354.64 33.12 95.4558.79 GEVO*80.60 88.83 87.00 83.2039.18 89.32 92.03 96.8570.19 93.13 98.6083.54 Table 2: Performance comparison of different GEVO variants across all tasks. * indicates a significant performance improvement under Wilcoxon Signed-Rank Test (p < 0.05). is significantly reduced. These results indicate that SFT on related tasks can effectively enhance the MLLM’s understanding and discrimination capability for ancient script–related tasks. Fur- thermore, when comparing the two SFT variants, Qwen3-VL-2B-SFTand GEVO, we find GEVO fur- ther mitigates the confusion between clerical script and regular script observed inQwen3-VL-2B-SFT, while simultaneously enhancing the predictive per- formance on bronze script. This suggests that GEVO effectively benefits from explicit glyph- level comparisons. Figure 10 reports the character recognition accu- racy across different script styles. Compared with the results ofQwen3-VL-2B-Instructshown in Figure 6,Qwen3-VL-2B-SFTexhibits performance gains only on regular script. This suggests that a limited amount of training data is insufficient to substantially improve character recognition, as the pronounced glyph variations across different historical periods of Chinese characters cannot be effectively generalized from a small number of samples. Consequently, ancient character recog- nition is inherently a knowledge-intensive task, which partly explains the weak performance of general-purpose models in this domain. Further- more,Qwen3-VL-2B-SFTshows degraded recogni- tion performance on oracle bone script and bronze script, indicating that MLLMs tend to exhibit rep- resentational bias toward script styles with more stable structures and regular strokes. Such bias weakens the models’ ability to capture the highly heterogeneous and non-standard glyph forms char- acteristic of early scripts. By incorporating charac- ter recognition training tasks, GEVO achieves sub- stantial performance improvements, with accuracy exceeding 50% on both clerical script and regular script. However, the gains on oracle bone script and bronze script remain limited. This observa- tion indicates that, even with fine-tuning, accurate recognition of ancient scripts remains challenging for MLLMs, highlighting the need for more com- prehensive and efficient datasets to facilitate deeper and more robust learning. 5.4 Visualization Analysis Qwen3-VL-2B-Instruct (sun) (mouth) GEVO (sun) (mouth) Figure 11:Visualization of similar image rep- resentations among “日” (sun) and “口” (mouth) baed onQwen3-VL-2B-Instructand GEVO in two- dimensional space. Boxes in different colors are used to distinguish images of script styles corresponding to different modern Chinese characters. We conduct a visualization analysis to explore GEVO’s ability to distinguish glyphs in the repre- sentation space, thereby providing substantial ev- idence for its performance in downstream tasks. For this purpose, Figure 11 presents the represen- tation distribution of the two distinct characters “日” (sun) and “口” (mouth). Since both are pic- tographic characters derived from real-world ob- jects, they are often compressed into a limited num- ber of stable geometric prototypes during the early stages, leading to convergent geometric abstraction at the level of outer contours. For more intuitive visualization, different images are positioned in the corresponding two-dimensional space based on their representations. Among them, GEVO demon- strates superior clustering ability for identical char- acters, positioning different styles of the character “口” (mouth) at similar distances. Additionally, the representations of the glyphs for the same charac- ter “日” () have also become more concentrated. But we still observe that the model struggles to distinguish particularly similar glyphs, such asand. This indicates that GEVO still has room for improvement, particularly in distin- guishing especially similar glyphs. We provide MLLMsT1.1T1.3T1.4T2.1T2.3T2.4T3.1T3.2T3.3Average GEVO69.5581.9579.8056.8487.8492.5790.5479.7597.3081.79 Qwen3-VL-2B-Instruct24.3059.0528.5041.3469.5925.0070.2723.0946.6243.08 Qwen3-VL-8B-Instruct45.5370.5561.2047.6375.3465.5471.6245.6274.3261.93 Table 3: Performance comparison on different tasks. more interesting visualization results with greater semantic differences in Appendix E. 6 Generalization on OOD Datasets Given the potential discrepancies among glyph fac- similes from different sources, we further extract a Out-of-distribution (OOD) subset from OBIsEvolu- tion (Wang et al., 2022) for evaluating the general- ization ability of GEVO. Specifically, we randomly select 150 characters from the dataset and man- ually filtered out questionable samples, resulting in a small-scale dataset containing 148 characters and 717 corresponding facsimiles. Subsequently, following the same procedure adopted in GEVO, we constructed an MLLM evaluation instruction set and re-evaluated GEVO’s inference results on the newly constructed benchmark. It should be emphasized that tasks T1.2 and T2.2 can not be conducted in this OOD benchmark, as each glyph has only a single version and thus lacks multiple variants for evaluation. The experimental results in Table 3 show that GEVO consistently outperforms the baseline MLLMs across all evaluation tasks, demonstrat- ing superior robustness and generalization abil- ity in out-of-distribution settings. Compared with general-purpose MLLMs, GEVO achieves more stable and balanced performance on diverse sub- tasks, indicating its stronger capability in handling challenging ancient character reasoning scenarios. These results validate the effectiveness of the pro- posed framework in enhancing ancient character understanding under distribution shifts. 7 Conclusion This paper introduces a benchmark for evaluat- ing MLLMs on Chinese character evolution tasks. Evaluations of 19 MLLMs reveal persistent weak- nesses in glyph comparison and character recogni- tion, though modest gains can be achieved through simple SFT. Bassed on the findings, we propose a curriculum-inspired fine-tuning approach based on glyph contrastive learning, which improves perfor- mance across tasks. Notably, fine-tuned 2B-scale models surpass all evaluated MLLMs. Limitations Since ink rubbings often contain noise and ex- hibit highly inconsistent glyphs, we explore simpler hand-copied facsimiles to assess the capabilities of MLLMs. In subsequent research, to enhance prac- tical applicability, we will conduct further studies on ink rubbings. Additionally, the character recog- nition performance in this study still falls short of practical application requirements. In the future, we will explore more data augmentation methods to enhance the character recognition capabilities of MLLMs. Furthermore, we encourage researchers to explore more and larger closed-source models to compensate for our limitations, as we are only able to test three closed-source models due to cost constraints. Finally, we also believe that integrating semantics during the process of glyph evolution is crucial, yet the relevant corpus remains scarce. Acknowledgements This work is supported by the National Natural Sci- ence Foundation of China (NSFC): “Research on Understanding Ancient Characters Based on Multi- modal Large Models” (Grant No. 62476111), China Postdoctoral Science Foundation Funded Project (Grant No. 2024M761122), Natural Sci- ence Foundation of Jilin Province (General Pro- gram, Grant No. 20260102295JC), the “Paleog- raphy and Chinese Civiliza tion Inheritance and Development Program” Collaborative Innovation Platform (No. G3829), and the National Social Science Foundation of China (No. 23VRC033). References Roar Bökset. 2006. Long story of short forms: the evolution of simplified Chinese characters. Ph.D. thesis, Institutionen för orientaliska språk. Jiahuan Cao, Yang Liu, Peirong Zhang, Yongxin Shi, Kai Ding, and Lianwen Jin. 2025. Tonggu-vl: Ad- vancing visual-language understanding in chinese classical studies through parameter sensitivity-guided instruction tuning.In Proceedings of the 33rd ACM International Conference on Multimedia, pages 11111–11120. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In In- ternational conference on machine learning, pages 1597–1607. Zijian Chen, Wenjun Zhang, Guangtao Zhai, and 1 oth- ers. 2025. Obi-bench: Can lmms aid in study of ancient script on oracle bones? In The Thirteenth In- ternational Conference on Learning Representations. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Mar- cel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261. Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chen- hui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, and 1 others. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793. Haisu Guan, Jinpeng Wan, Yuliang Liu, Pengjie Wang, Kaile Zhang, Zhebin Kuang, Xinyu Wang, Xiang Bai, and Lianwen Jin. 2024a. An open dataset for the evolution of oracle bone characters: EVOBC. CoRR, abs/2401.12467. Haisu Guan, Huanxin Yang, Xinyu Wang, Shengwei Han, Yongge Liu, Lianwen Jin, Xiang Bai, and Yu- liang Liu. 2024b. Deciphering oracle bone language with diffusion models. In Proceedings of the 62nd Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 15554–15567. Shuangping Huang, Haobin Wang, Yongge Liu, Xi- aosong Shi, and Lianwen Jin. 2019. Obc306: A large-scale oracle bone character recognition dataset. In 2019 International Conference on Document Anal- ysis and Recognition (ICDAR), pages 681–688. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276. Qingju Jiao, Jingwen Wu, Qi Liu, Han Zhang, Zhan Zhang, Bang Li, Jing Xiong, Guoying Liu, and Yongge Liu. 2025.A graph-based evolutionary dataset for oracle bone characters from inscriptions to modern chinese scripts. npj Heritage Science, 13(1):369. Caoshuo Li, Zengmao Ding, Xiaobin Hu, Bang Li, Donghao Luo, Xu Peng, Taisong Jin, Yongge Liu, Shengwei Han, Jing Yang, and 1 others. 2025a. Oracleagent: A multimodal reasoning agent for oracle bone script research.arXiv preprint arXiv:2510.26114. Jing Li, Xueke Chi, Qiufeng Wang, Kaizhu Huang, Da- Han Wang, Yongge Liu, and Cheng-Lin Liu. 2026. A comprehensive survey of oracle character recogni- tion: Challenges, datasets, methodology, and beyond. Pattern Recognition, 169:111824. Jing Li, Bin Dong, Qiu-Feng Wang, Lei Ding, Rui Zhang, and Kaizhu Huang. 2023a. Decoupled learn- ing for long-tailed oracle character recognition. In International Conference on Document Analysis and Recognition, pages 165–181. Jing Li, Qiu-Feng Wang, Kaizhu Huang, Xi Yang, Rui Zhang, and John Y Goulermas. 2023b. Towards better long-tailed oracle character recognition with adversarial data augmentation. Pattern Recognition, 140:109534. Jinhao Li, Zijian Chen, Runze Jiang, Tingzhu Chen, Changbo Wang, and Guangtao Zhai. 2025b. Mitigat- ing long-tail distribution in oracle bone inscriptions: Dataset, model, and benchmark. In Proceedings of the 33rd ACM International Conference on Multime- dia, pages 7729–7738. Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023. Improved baselines with visual instruc- tion tuning. Preprint, arXiv:2310.03744. Yang Liu, Jiahuan Cao, Hiuyi Cheng, Yongxin Shi, Kai Ding, and Lianwen Jin. 2025. Mcs-bench: A compre- hensive benchmark for evaluating multimodal large language models in chinese classical studies. In Pro- ceedings of the 63rd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 10435–10492. Runqi Qiao, Qiuna Tan, Guanting Dong, MinhuiWu MinhuiWu, Jiapeng Wang, Yifan Zhang, Zhuoma GongQue, Chong Sun, Yida Xu, Yadong Xue, and 1 others. 2025. V-oracle: Making progressive rea- soning in deciphering oracle bones for you and me. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 20124–20150. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. Daqian Shi, Xiaolei Diao, Lida Shi, Hao Tang, Yang Chi, Chuntao Li, and Hao Xu. 2022. Charformer: A glyph fusion based attentive framework for high-precision character image denoising. In Proceedings of the 30th ACM international conference on multimedia, pages 1147–1155. Qwen Team. 2025a. Qwen2.5-vl. Qwen Team. 2025b. Qwen3 technical report. Preprint, arXiv:2505.09388. Mei Wang, Weihong Deng, and Sen Su. 2024a. Oracle character recognition using unsupervised discrimi- native consistency network. Pattern Recognition, 148:110180. Mengru Wang, Yu Cai, Li Gao, Ruichen Feng, Qingju Jiao, Xiaolin Ma, and Yu Jia. 2022. Study on the evolution of chinese characters based on few-shot learning: From oracle bone inscriptions to regular script. Plos one, 17(8):e0272974. Pengjie Wang, Kaile Zhang, Xinyu Wang, Shengwei Han, Yongge Liu, Jinpeng Wan, Haisu Guan, Zhebin Kuang, Lianwen Jin, Xiang Bai, and 1 others. 2024b. An open dataset for oracle bone character recognition and decipherment. Scientific Data, 11(1):976. Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, and 1 others. 2025. In- ternvl3.5: Advancing open-source multimodal mod- els in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265. Xin Wang, Yudong Chen, and Wenwu Zhu. 2021. A survey on curriculum learning. IEEE transac- tions on pattern analysis and machine intelligence, 44(9):4555–4576. Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234. Xing Wen. 2011. Chinese paleography, calligraphy, and pattern recognition: Styles and scripts in excavated ancient chinese documents. In 2011 International Conference on Document Analysis and Recognition, pages 951–956. Xinyu Yao, Mengdi Wang, Bo Chen, and Xiaob- ing Zhao. 2025. Wenyangpt: A large language model for classical chinese tasks. arXiv preprint arXiv:2504.20609. Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, and 1 others. 2024. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800. Xinyi Zhao, Siyuan Liu, Yikai Wang, and Yanwei Fu. 2022. Ffd augmentor: towards few-shot oracle char- acter recognition from scratch. In Proceedings of the Asian Conference on Computer Vision, pages 1622– 1639. Zhihan Zhou, Daqian Shi, Rui Song, Lida Shi, Xi- aolei Diao, and Hao Xu. 2025.Ancientbench: Towards comprehensive evaluation on excavated and transmitted chinese corpora. arXiv preprint arXiv:2512.17756. A More Task Details In practical tasks, to standardize the input for the model, we provide longer and more detailed in- structions compared to those in Figure 12, 13, and 14. A more detailed breakdown of the task com- position is provided in Figure 3. Each task uses accuracy as the evaluation metric, meaning that if the correct answer appears in the generated result, it is considered correct. If the generated result con- tains multiple candidate answers, it is assumed that the model did not understand the instruction and is therefore considered a prediction failure. Addi- tionally, for the path comparison task T3.2, given the difficulty of correctly ordering the entire evo- lutionary path, we adopt a more lenient evaluation strategy: if the prediction for a certain position in the corresponding path is correct, the image is considered correctly predicted. Ultimately, the de- nominator is the total number of images in the path. For example, if the scrambled evolutionary path of a character is Oracle Bone Script→Regular Script →Bronze Inscription, and the predicted result is "Bronze Inscription→Regular Script→Oracle Bone Script," then the prediction accuracy is 1/3. Furthermore, the 11 tasks correspond to different capabilities of the model, and we further explain why these instructions are included in the bench- mark. •T1.1.It demonstrates the model’s basic glyph recognition capability by determining the script period corresponding to a given char- acter image, which is one of the most funda- mental tasks in evolutionary analysis. There- fore, we emphasize placing it at the forefront. •T1.2. It serves as an effective extension of T1, aimed at comparing script styles when pre- sented with different glyphs of the same char- acter. Based on experimental results, provid- ing the character information in advance im- proves the model’s comparative performance. •T1.3. It is also an extension of the basic glyph comparison task, testing the model’s discrimi- native ability by comparing glyphs under the prior condition of different characters. In most cases, it performs better than T1.1 but slightly worse than T1.2. This indicates that the model is more adept at comparing different glyphs of the same character. •T1.4. It is a higher-order extension of T1.1, used for comparison among multiple different glyphs. •T2.1. It represents another fundamental task in ancient script studies: character recognition. Due to the significant changes in glyph forms during the evolution of ancient scripts, the character recognition capability of the model is generally poor. •T2.2. It is an extension of the basic charac- ter recognition task, determining whether two glyphs represent the same character within the same script style. •T2.3. Similarly, it is an extension of the basic capability of T2.1, determining whether two glyphs represent the same character under the premise of different script styles. • T2.4. It is a high-order extension of T2.1, designed to evaluate a model’s capability in comparative recognition and selection among multiple characters. •T3.1. This is a character recognition task within an evolutionary path, which encour- ages the model to utilize more evolutionary context to determine the exact character. Com- pared to T2.1, this task is simpler because it provides more background knowledge for the model to reference. •T3.2. This is a task of ordering evolutionary sequences, designed to encourage the model to reconstruct the chronological order of glyph evolution within a shuffled evolutionary path. It requires the model to possess basic capabil- ities in calligraphy style recognition and the ability to reorganize the sequences according to temporal progression. Therefore, it is a more complex form of T1.1. •T3.3. This is also an important task in evo- lutionary analysis, involving the completion of a path when a specific segment is missing. In this process, the model needs to compre- hensively consider factors such as the glyph structure and semantics of the characters, and frame the completion task as a retrieval prob- lem. However, due to the vast search space of glyph retrieval, which is not well-suited to MLLMs, we have reformulated it as a multiple-choice question. Task 1. Script Style Recognition T1.1 请根据上图判断此文字所属的书体时代类别。可选类别包括：甲骨文、金文、篆文、隶书、楷书。输出仅且必须为五个类别 之一的单个汉字词，绝对禁止任何形式的解释、理由、分析、过程描述、列表、示例、序号、标点或其它文字。若不能确定， 请只输出：无法判断。 <image>Pleasedeterminethescriptstylecategorytowhichthischaracterbelongsbasedontheaboveimage.Theoptionalcategories include:oraclebonescript,bronzeinscrip,sealscript,clericalscript,andregularscript.Theoutputmustbeandcanonlybeasingle Chinesewordfromoneofthefivecategories.Absolutelynoexplanations,reasons,analyses,descriptionsoftheprocess,lists, examples,serialnumbers,punctuation,orothertextofanyformareallowed.Ifuncertain,pleaseoutputonly:Cannotdetermine. 请根据下列两张图判断它们是否属于同一书体时代。可选答案仅包括：是、否。绝对禁止任何形式的解释、理由、分析、过 程描述、列表、示例、序号、标点或其它文字。 T1.2 Pleasedeterminewhetherthefollowingtwoimagesbelongtothesamescriptstylebasedonthegivenoptions.Theonlypossible answersare:yes,no.Absolutelynoexplanations,reasons,analyses,processdescriptions,lists,examples,numbering,punctuation,orany othertextisallowed.<image1><image2> T1.3 Pleasedeterminewhichofthefollowingmultipleimagesdoesnotbelongtothesamescriptstyleastheothers.ThefouroptionsareA <image1>;B<image2>;C<image3>;D<image4>,whereonlyoneisthecorrectanswer.Theoutputmustbeandcanonlybeasingleletter fromA,B,C,orD.Absolutelynoexplanations,reasons,analyses,descriptionsoftheprocess,lists,examples,serialnumbers, punctuation,orothertextofanyformareallowed.Ifunabletodetermine,pleaseoutputonly:Cannotdetermine. 请根据下列多张图判断哪一张不属于其他图片的书体时代。四个选项分别为A <image1>；B <image2>；C <image3>；D <image4>，其中只有一个是正确答案。输出仅且必须为A、B、C 或D 中的单个字母，绝对禁止任何形式的解释、理由、分 析、过程描述、列表、示例、序号、标点或其它文字。若无法判断，请只输出：无法判断。 T1.4 Figure 12: Detailed instructions for Task 1. We use the special character ‘< image >’ to represent an image. T1.2 and T1.3 share the same instruction, but in constructing the context for T1.2, it is necessary to ensure that the two images correspond to the same modern Chinese character. In contrast, T1.3 requires ensuring that the images correspond to different modern Chinese characters. Task 2. Character Recognition 下面给出一张古文字图像，请根据图像内容识别文字，并直接输出对应的现代汉字。禁止任何形式的解释、理由、分 析、过程描述、列表、示例、序号、标点或其它文字。若不能确定，请只输出：无法判断。<image> T2.1 Belowisanimageofanancientscript.Pleaserecognizethetextbasedontheimagecontentanddirectlyoutputthecorrespon dingmodernChinesecharacter.Absolutelynoexplanations,reasons,analyses,descriptionsoftheprocess,lists,examples,seri alnumbers,punctuation,orothertextofanyformareallowed.Ifuncertain,pleaseoutputonly:Cannotdetermine.<image> T2.2 请根据下列两张图判断它们是否属于同一个字。可选答案仅包括：是、否。绝对禁止任何形式的解释、理由、分析、 过程描述、列表、示例、序号、标点或其它文字。<image1><image2> Pleasedeterminewhetherthetwoimagesbelowbelongtothesamecharacter.Theonlypossibleanswersare:Yes,No. Absolutelynoexplanations,reasons,analyses,descriptionsoftheprocess,lists,examples,serialnumbers,punctuation,or othertextofanyformareallowed.<image1><image2> T2.3 请根据下列多张图判断哪一张对应的字与其他字不同。四个选项分别为A <image1>；B <image2>；C <image3>；D <image4>，其中只有一个是正确答案。输出仅且必须为A、B、C 或D 中的单个字母，绝对禁止任何形式的解释、理 由、分析、过程描述、列表、示例、序号、标点或其它文字。若无法判断，请只输出：无法判断。 Pleasedeterminewhichofthefollowingmultipleimagescorrespondstoacharacterthatisdifferentfromtheothers.Thefour optionsareA<image1>;B<image2>;C<image3>;D<image4>,whereonlyoneisthecorrectanswer.Theoutputmustbeandcanonly beasingleletterfromA,B,C,orD.Absolutelynoexplanations,reasons,analyses,descriptionsoftheprocess,lists,examples, serialnumbers,punctuation,orothertextofanyformareallowed.Ifunabletodetermine,pleaseoutputonly:Cannotdetermine. T2.4 Figure 13: Detailed instructions for Task 2. B More MLLMs Details We evaluate several common MLLMs includ- ing:TongGu-VL-2B-Instruct(Cao et al., 2025) (An expert model trained on a cultural her- itage dataset, which has been reported to ex- hibit stronger comprehension of ancient Chinese scripts compared to other models.),Qwen2.5-VL series (Team, 2025a) andQwen3-VLseries (Team, 2025b),InternVL-3_5series (Wang et al., 2025),MiniCPM-V-2_6andMiniCPM-V-4_5(Yao et al., 2024),GLM-4.1V-9B-Thinking(GLM et al., 2024),DeepSeekOCR(Wei et al., 2025), LLaVA-1.5series (Liu et al., 2023). Addition- ally, we compared three closed-source models: GPT-4o-mini,GPT-5-mini(Hurst et al., 2024), andGemini-3-Flash(Comanici et al., 2025). All API calls were made through third-party inter- faces 3 . We do not keep detailed statistics on the ex- penses, but including the cost of model debugging, evaluating the benchmarks on the three models ex- ceeded 500$. Therefore, we do not explore more closed-source models due to cost constraints. C More Evaluation Results Table 4 presents the performance of different MLLMs on reasoning tasks across the entire dataset (including both the training and test sets). Con- sistent with the results in Table 1, theQwen3se- ries models achieve the best performance, with Qwen3-VL-8B-Instructobtaining the highest av- erage performance. Additionally, MLLMs show certain potential for glyph discrimination tasks, though text recognition remains a significant chal- lenge. The consistent trends indicate that the test 3 https://api.xi-ai.cn/ Task 3. Evolutionary Path Tasks T3.1 下方依次给出同一汉字在不同历史时期的字形。请根据其演化关系，直接给出该字对应的现代规范汉字。禁止解释、 分析或输出除单个现代汉字以外的内容。甲骨文: <image1>, 金文: <image2>, 篆文: <image3>, 隶书: <image4>, 楷书: <image5>, 现代汉字： BelowaretheglyphsofthesameChinesecharacterindifferenthistoricalperiods,listedinorder.Basedonitsevolutionary relationship,directlyprovidethecorrespondingmodernstandardizedChinesecharacterforthisword.Explanations, analyses,oranycontentotherthanasinglemodernChinesecharacterareprohibited. OracleBoneScript:<image1>,BronzeInscript:<image2>,SealScript:<image3>,ClericalScript:<image4>,RegularScript: <image5>,ModernChineseCharacter: T3.2 严格按照图像在输入中的顺序逐一判断它们分别属于哪个书体时代。输出格式必须是：图像编号-书体名称，例如： 1-甲骨文；2-篆文；3-隶书；4-金文；5-楷书。禁止模仿上述示例的顺序输出，严格保持与图像输入顺序一致，禁 止输出解释。<image1><image2><image3><image4><image5>,对应时期： Strictlydeterminewhichscriptperiodeachimagebelongsto,followingtheorderoftheimagesintheinput.Theoutput formatmustbe:imagenumber–scriptname,forexample:1–OracleBoneScript;2–SealScript;3–ClericalScript;4–Bronze Inscription;5–RegularScript.Donotimitatetheorderoftheexampleabove;strictlymaintainconsistencywiththeorderof theimageinput.Explanatoryoutputsareprohibited.<image1><image2><image3><image4><image5>,CorrespondingPeriods: T3.3 下方展示的是同一汉字在不同时期的部分书体写法：甲骨文:<image1>;金文:<image2>;篆文:<image3>;楷书: <image4> 其中[隶书]缺失。请根据汉字形体的历史演化规律，判断缺失的图像最有可能对应选项A、B、C、D中 的哪一个。禁止输出解释，只需给出一个选项字母。 选项：A:<image5>;B:<image6>;C:<image7>;D:<image8> 答案： BelowarepartialscriptformsofthesameChinesecharacterfromdifferentperiods:OracleBoneScript:<image1>;Bronze Inscript:<image2>;SealScript:<image3>;RegularScript:<image4>. The[ClericalScript]ismissing.BasedonthehistoricalevolutionpatternsofChinesecharacterforms,determinewhich optionA,B,C,orDthemissingimagemostlikelycorrespondsto.Explanationsareprohibited;onlyprovidetheletterofthe chosenoption. Options:A:<image5>;B:<image6>;C:<image7>;D:<image8> Answer: Figure 14: Detailed instructions for Task 3. MLLMsT1.1T1.2T1.3T1.4T2.1T2.2T2.3T2.4T3.1T3.2T3.3Average Qwen3-VL-2B-Instruct17.66 58.87 53.59 25.5320.82 70.47 73.23 25.8230.36 17.51 45.8839.98 Qwen3-VL-8B-Instruct 47.79 76.63 70.22 58.4529.76 74.71 82.77 69.6750.96 49.20 71.7561.99 Qwen3-VL-30B-A3B-Instruct 31.47 68.20 56.31 46.8829.56 72.34 75.66 65.7253.60 34.70 66.1554.60 InternVL3_5-1B-HF 6.6646.20 44.50 20.4215.60 46.92 39.14 13.4021.206.1830.6526.44 InternVL3_5-8B-HF 17.81 54.02 52.41 26.126.4759.21 39.22 38.7816.30 20.01 38.5733.54 MiniCPM-V-2_6 7.8752.31 52.61 27.718.4077.61 78.02 42.198.0616.61 24.8036.02 MiniCPM-V-4_521.26 63.80 54.94 35.489.0480.71 78.61 62.0415.14 25.68 24.9624.96 GLM-4.1V-9B-Thinking25.16 51.37 51.33 30.2019.16 73.39 71.02 46.3439.75 30.86 50.5344.46 DeepSeekOCR-3B10.43 51.55 49.06 24.548.8252.65 55.21 20.0619.47 14.96 46.1932.09 LLaVA-1.5-7B-HF1.588.319.2324.940.0727.49 24.89 26.760.1112.18 27.5114.82 LLaVA-1.5-13B-HF 2.8445.21 46.846.930.1648.78 54.655.820.066.2613.1020.97 Table 4: Performance comparison of different MLLMs across all tasks (accuracy %). Due to cost constraints, we do not evaluate the full dataset on additional closed-source models. 0 2 4 6 Loss Original Training Loss 01000200030004000500060007000 Steps 1.0 0.9 0.8 0.7 Loss Smoothed Training Loss (LOWESS) Figure 15: During the first stage, the variation of the contrastive learning loss driven by glyphs and the corre- sponding Locally Weighted Regression fitting curve. set distribution aligns with the overall dataset dis- tribution, allowing it to serve as a representative proxy for evaluating model performance on the entire dataset—at only 1/10 of the inference cost. 01020304050 Steps 1 2 3 4 Loss Figure 16: The loss variation of the model in the second stage. 0510152025 Steps 0.5 1.0 1.5 2.0 2.5 Loss Figure 17: The loss variation of the model in the third stage. D Model Fine-tuning Details We use LlamaFactory 4 for fine-tuning. During the first stage of fine-tuning, we package each image into a standard context template, which includes two key inputs:“type” : “text”, “text” : char and“type” : “image”, “image” : img. Here ‘char’ represents the standard modern Chinese char- acter text, and ‘img’ represents the path to the corre- sponding image. Subsequently, the language model parameters of MLLMs (taking Qwen3-VL as an ex- ample) are frozen, and corresponding image repre- sentations are obtained for computing the lossL con . In the second stage, we encapsulate the instructions and answers from T2.1 (drawn exclusively from the training set to prevent information leakage) into a standard training template, and fine-tune the lan- guage model component of the MLLMs to obtain the second fine-tuned version. Finally, in the third stage, we fine-tune the language model component across all tasks in the benchmark to obtain the final evolutionary understanding model. Figure 15 illustrates the variation of the over- all loss function during training in the first stage. Although the learning loss exhibits significant fluc- tuations across different images, the overall fitting curve indicates that the loss function is effectively decreasing. Figure 16 shows the variation in loss values dur- ing the second stage of model training. Compared to the loss from the glyph comparison learning in the first stage, the loss in the second stage is smoother. This is because developing the model’s character recognition capability is relatively sim- pler than comparing glyphs, especially considering the strong reasoning abilities already present in the existing MLLMs. Similarly, Figure 17 illustrates the loss during the third stage of training. Given the relatively small training dataset and the rapid convergence observed, we argue that SFT data is not required in large quantities for evolutionary analysis in MLLMs. In the task-specific SFT stage of the second/third phase, we employ the same learning rate of 1e-5 for 3 epochs, with a warmup ratio of 0.1. It should be noted thatQwen3-VL-2B-SFTalso follows the same training strategy on identical samples. This pro- cess do not employ the LoRA strategy and involve fine-tuning on the full set of parameters. All ex- periments are conducted on 4*A100 80GB GPUs. We save the model parameters after training and 4 https://w.llamafactory.cn/ evaluate them across various tasks. E More Visualization Results Qwen3-VL-2B-Instruct (fource) (knife) GEVO (fource) (knife) Figure 18:Visualization of similar image repre- sentations among “力” (fource) and “刀” (knife) baed onQwen3-VL-2B-Instructand GEVO in two- dimensional space. Qwen3-VL-2B-Instruct (ten thousand) (square) GEVO (ten thousand) (square) Figure 19: Visualization of similar image representa- tions among “万” (ten thousand) and “方” (square) baed onQwen3-VL-2B-Instructand GEVO in two- dimensional space. Figure 18 and Figure 19 also confirm the influence of glyph similarity on the model.For “力” (fource) and “刀” (knife), Qwen3-VL-2B-Instructexhibits a deficiency in distinguishing characters that share similar glyphs but are actually different characters ( and). Correspondingly, GEVO tends to assign larger relative distances to the two. As previously discussed, GEVO still lacks the corresponding capability to distinguish between extremely similar glyphs (and), which is also one of the future research directions. The same phenomenon is also observed between “万” (ten thousand) and “方” (square), where GEVO can distinguish the two characters in the bottom right corner and maintain a greater distance, whereas in the results ofQwen3-VL-2B-Instruct, the two characters partially overlap. Another interesting finding is that although these characters share similar writing styles during cer- tain evolutionary stages, they possess fundamen- tally distinct semantics. For example, “万” (ten thousand) and “方” (square) differ only slightly in their written composition, but the meanings they convey are vastly different. Based on this, we also hope that our research can inspire professional pale- ographers to study and explain the aforementioned phenomena from the perspective of MLLMs’ un- derstanding of glyphs.