Organizations: College of Computer Science and Technology, Jilin University, China · Key Laboratory of Ancient Chinese Script, Culture Relics and Artificial Intelligence, Jilin University, China · School of Artificial Intelligence, Jilin University, China · School of Archaeology, Jilin University, China
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{https://github.com/songruiecho/GEVO}.
Figures & tables
Figure 1: The text evolution process of the character “日” (sun) and the organization of data.
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 .
Figure 3: Task-related instruction construction and optimization process.
Figure 4: Benchmark quantity statistics.
MLLMs
T1.1
T1.2
T1.3
T1.4
T2.1
T2.2
T2.3
T2.4
T3.1
T3.2
T3.3
Average
Closed-source Models
GPT4-all-mini
30.85
51.09
48.30
4.00
0.30
50.10
49.17
1.05
0.00
19.23
24.13
25.29
GPT5-mini
24.65
46.84
46.30
11.50
0.07
49.50
48.70
6.79
0.00
13.60
25.78
24.88
Gemini-3-Flash
21.73
48.71
48.90
24.20
0.03
48.46
52.27
25.17
3.74
12.06
21.52
27.89
Open-source Models
TongGu-VL-2B-Instruct
4.07
50.78
52.90
15.00
8.64
49.79
52.49
1.54
9.27
4.39
26.57
25.04
Table 1: Performance comparison of different MLLMs across all test sets (accuracy %). The evaluation results on the full dataset are provided in Appendix C .
Figure 5: The confusion matrices of script style predictions for different models, with unrecognizable cases removed to better highlight the relationships among glyphs across scripts and periods.
Figure 6: Characters accuracy on different script styles.
Figure 7: Comparison of results for Qwen3-VL-2B-Instruct after SFT ( Qwen3-VL-2B-SFT ) across all tasks.
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, and are considered glyphically 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 maintain a distance between them.
MLLMs
T1.1
T1.2
T1.3
T1.4
T2.1
T2.2
T2.3
T2.4
T3.1
T3.2
T3.3
Average
Qwen3-VL-2B-SFT
68.33
80.14
74.20
68.70
22.57
82.41
82.61
87.76
41.51
86.70
92.66
71.60
GEVO-Stage1
75.00
83.14
78.70
78.10
1.64
62.01
47.84
26.75
4.42
90.13
93.65
58.31
GEVO-Stage2
32.20
57.19
50.40
32.40
28.20
81.62
89.37
92.13
54.64
33.12
95.45
58.79
GEVO *
80.60
88.83
87.00
83.20
39.18
89.32
92.03
96.85
70.19
93.13
98.60
83.54
Table 2: Performance comparison of different GEVO variants across all tasks. * indicates a significant performance improvement under Wilcoxon Signed-Rank Test ( p<0.05 ).
Figure 9: The confusion matrices of script style predictions for Qwen3-VL-2B-SFT and GEVO.
Figure 10: Accuracy of characters across different script styles for Qwen3-VL-2B-SFT and GEVO.
Figure 11: Visualization of similar image representations among “日” (sun) and “口” (mouth) baed on Qwen3-VL-2B-Instruct and GEVO in two-dimensional space. Boxes in different colors are used to distinguish images of script styles corresponding to different modern Chinese characters.
MLLMs
T1.1
T1.3
T1.4
T2.1
T2.3
T2.4
T3.1
T3.2
T3.3
Average
GEVO
69.55
81.95
79.80
56.84
87.84
92.57
90.54
79.75
97.30
81.79
Qwen3-VL-2B-Instruct
24.30
59.05
28.50
41.34
69.59
25.00
70.27
23.09
46.62
43.08
Qwen3-VL-8B-Instruct
45.53
70.55
61.20
47.63
75.34
65.54
71.62
45.62
74.32
61.93
Table 3: Performance comparison on different tasks.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
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.
Figure 13: Detailed instructions for Task 2.
Figure 14: Detailed instructions for Task 3.
MLLMs
T1.1
T1.2
T1.3
T1.4
T2.1
T2.2
T2.3
T2.4
T3.1
T3.2
T3.3
Average
Qwen3-VL-2B-Instruct
17.66
58.87
53.59
25.53
20.82
70.47
73.23
25.82
30.36
17.51
45.88
39.98
Qwen3-VL-8B-Instruct
47.79
76.63
70.22
58.45
29.76
74.71
82.77
69.67
50.96
49.20
71.75
61.99
Qwen3-VL-30B-A3B-Instruct
31.47
68.20
56.31
46.88
29.56
72.34
75.66
65.72
53.60
34.70
66.15
54.60
InternVL3_5-1B-HF
6.66
46.20
44.50
20.42
15.60
46.92
39.14
13.40
21.20
6.18
30.65
26.44
InternVL3_5-8B-HF
17.81
54.02
52.41
26.12
6.47
59.21
39.22
38.78
16.30
20.01
38.57
33.54
MiniCPM-V-2_6
7.87
52.31
52.61
27.71
8.40
77.61
78.02
42.19
8.06
16.61
24.80
36.02
Appendix
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.
Figure 15: During the first stage, the variation of the contrastive learning loss driven by glyphs and the corresponding Locally Weighted Regression fitting curve.
Figure 16: The loss variation of the model in the second stage.
Figure 17: The loss variation of the model in the third stage.
Figure 18: Visualization of similar image representations among “力” (fource) and “刀” (knife) baed on Qwen3-VL-2B-Instruct and GEVO in two-dimensional space.
Figure 19: Visualization of similar image representations among “万” (ten thousand) and “方” (square) baed on Qwen3-VL-2B-Instruct and GEVO in two-dimensional space.