Tracing the Evolution of Oracle Bone Characters Across Three Millennia

Tianhao Fu; Xinxin Xu; Spike Wang; Cunyi Kang; Jian Cao; Xixin Cao

cs.CLSep 28, 2026

Tracing the Evolution of Oracle Bone Characters Across Three Millennia

Authors: Tianhao Fu, Xinxin Xu, Spike Wang, Cunyi Kang, Jian Cao, Xixin Cao

Organizations: Peking University · Independent

Abstract

Of the approximately 4,500 Oracle Bone Inscription (OBI) characters discovered from the Shang dynasty, only about 1,600 have been deciphered. Many computational approaches compare OBI with glyphs from one historical period at a time. However, during the evolution of Chinese characters, significant structural or semantic changes often occur in uncertain dynasties. A single-period reference may be insufficient when relevant forms change substantially between observed eras. Therefore, we propose the \textbf{Manifold-based Script Evolution Framework (MSEF)}, a framework that models the evolution series (OBI, Bronze, Seal, Clerical, Regular) of Chinese characters as the continual evolution of a manifold space. MSEF represents each character as an era-specific manifold point and learns continuous inter-era transition rules via Neural Ordinary Differential Equations. Both manifold space and transition dynamics can be trained end-to-end through character evolution pairs across any two eras.

Figures & tables

Figure 1: Evolution of Chinese script. Taking the character “Ox” as an example, this figure demonstrates the evolution of Chinese characters across five historical periods. This character underwent a significant morphological change in early bronze script, then later reverted to a form similar to its oracle bone appearance. A method restricted to bronze forms could assign low similarity to this particular pair. The example motivates evaluating evidence across multiple periods, though it does not by itself establish an error rate for single-period methods.

Figure 2: Overview of MSEF. Top: Continuous Manifold Evolution. The framework models the 3,000-year evolution of Chinese script as a continuous trajectory (red line) through era-specific manifolds $\mathcal{M}_{t}$ , conditioned on normalized time $t$ (from OBI $t=0.0$ to Regular $t=1.0$ ). The intervals of $t$ define an ordered model-time coordinate; their lengths are not proportional to calendar-year durations. Neural ODEs govern the continuous transitions via forward flows $\phi$ and backward flows $\psi$ . Specifically, each historical period manifold is represented in a shared high-dimensional coordinate space, where each observed glyph corresponds to a point. Inter-period point transitions are governed by the continuous neural ODE. Such a framework enables us to derive candidate character solutions for any dynasty within the character manifold space by leveraging any given dynasty. This facilitates the design of subsequent bidirectional cascading decryption algorithms. Bottom: Cascaded Bidirectional Decipherment. The inference process consists of two stages: (1) Forward Recall : Given an OBI query, the model progressively infers its coordinate in subsequent eras (Bronze $\to$ Seal $\to$ Clerical $\to$ Regular) to retrieve candidate sets. (2) Backward Verification : Candidates from the modern era are traced back through history via inverse flows. At each reverse transition, candidates showing large reconstruction errors (marked with ✗ ) are pruned step-by-step. The final decipherment is determined by the backward compatibility with the original OBI encoding.

Period	Time $t$	Historical Date
OBI	0.00–0.30	c. 1200–1050 BCE
Bronze	0.30–0.70	c. 1050–250 BCE
Seal	0.70–0.85	c. 220 BCE
Clerical	0.85–1.00	c. 200 CE
Regular	1.00	c. 600 CE–present

Table 1: Time encoding for historical periods. Subperiod details are provided in Appendix K .

Eval.	Metric	Pix2Pix	CycleGAN	BBDM	CDE	OBSD	MSEF
OBS-OCR	Top-1	0.0	0.0	19.5	31.0	41.0	71.5
OBS-OCR	Top-10	0.0	0.0	29.5	47.5	50.5	82.0
OBS-OCR	Top-20	0.0	0.0	34.5	50.0	54.5	84.5
OBS-OCR	Top-50	4.5	8.5	39.0	52.5	58.0	86.5
OBS-OCR	Top-100	13.0	19.0	42.0	56.0	61.0	88.0
OBS-OCR	Top-200	20.0	37.5	46.0	59.5	62.5	89.0

Table 2: Single-round decipherment on HUST-OBS and EVOBC. We follow Guan et al. (2024) ’s protocol. The scores are reported under the cited protocol; adapters and information-matched comparisons are not independently verified.

Model	Params	Normal	Complex	Overall
Random (4-choice)	—	25.00	25.00	25.00
GPT-4o-2024-11-20	—	26.31	25.52	26.23
Gemini 2.5 Pro	—	55.22	39.44	53.66
Claude 4 Sonnet	—	35.93	25.92	34.94
GLM-4.5V-106B	106B	33.19	27.11	32.48
Qwen2.5-VL-72B	72B	25.41	24.98	25.36

Table 3: Results on PictOBI-20k. MSEF subgroup and overall scores are reported; the aggregation is not verified as a common-population average.

Method	R@1	R@5	R@10	R@1%	AP
Diff-Oracle (2023)	52.1 $\pm$ 1.3	68.9 $\pm$ 1.6	77.4 $\pm$ 1.2	85.2 $\pm$ 1.0	0.56
OracleFusion (2025)	58.3 $\pm$ 1.2	74.5 $\pm$ 1.3	81.2 $\pm$ 1.1	88.1 $\pm$ 0.9	0.62
CrossFont (2025)	54.6 $\pm$ 1.4	70.8 $\pm$ 1.4	78.5 $\pm$ 1.2	86.5 $\pm$ 1.0	0.58
OracleSage (2024)	60.1 $\pm$ 1.1	76.2 $\pm$ 1.2	82.8 $\pm$ 1.0	89.5 $\pm$ 0.8	0.64
OracleAgent (2025)	62.8 $\pm$ 1.0	77.9 $\pm$ 1.1	84.6 $\pm$ 0.9	90.8 $\pm$ 0.7	0.67
MSEF (Ours)	72.5 $\pm$ 0.8	86.5 $\pm$ 0.6	91.8 $\pm$ 0.5	95.6 $\pm$ 0.4	0.78

Table 4: Reported FGCCES score summaries. Run-level records and significance tests are unavailable for independent verification.

Data utilization
Configuration	R@1	$\boldsymbol{\Delta}$
Full MSEF	72.5 $\pm$ 0.8	—
Complete chains only	52.8 $\pm$ 1.3	$-$ 19.7 †
w/o fine-grained subperiods	68.3 $\pm$ 0.9	$-$ 4.2 †
Architecture components
Single manifold (no Neural ODE)	62.2 $\pm$ 1.1	$-$ 10.3 †

Table 5: Reported ablation scores on FGCCES. Significance symbols are retained from the source report but not independently verified.

Evaluation Outcome	Pct.
Agrees with expert consensus	72%
Disagrees, later validated	12%
Disagrees, experts maintain	16%

Table 6: Expert-evaluation percentages reported in the manuscript for 100 challenging cases; case-level records are unavailable for independent verification.

Appendix figures & tables46 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Evaluation	Reported value (%)
OBSD	HUST-OBS/EVOBC, OBS-OCR Top-1	41.0
Diff-Oracle	FGCCES, R@1	52.1
CrossFont	FGCCES, R@1	54.6

Appendix

Table 7: Selected baseline results under their stated benchmark protocols. Values across different benchmarks are not directly comparable.

Group	Count	Top-1 (%)	Difference from Low
Low, distance $<0.3$	228	67.5	—
Medium, distance $0.3$ – $0.6$	445	48.3	$-19.2$
High, distance $>0.6$	334	22.7	$-44.8$
Abrupt transition	112	14.2	$-53.3$

Appendix

Table 8: Unvalidated generation-accuracy summary by complexity label. The three distance-group counts sum to 1,007; the membership relationship of the 112 abrupt-transition examples is not established. The four rows are not treated as a disjoint partition.

Era pair	Independent R@1	Joint-consistent R@1	Gain
OBI $\to$ Bronze	61.2	74.8	13.6
OBI $\to$ Seal	58.4	71.3	12.9
OBI $\to$ Clerical	56.7	69.5	12.8
OBI $\to$ Regular	54.6	72.5	17.9

Appendix

Table 9: Reported CrossFont retrieval with and without sequence-consistency filtering.

Figure 3: Selected examples of shared correspondence labels and era-specific glyph forms. These examples do not imply universal semantic invariance.

Figure 4: Exploratory workflow: feature construction, cross-era sequence organization, clustering, and baseline error analysis.

Feature	Description
Stroke count	Number of identified strokes
Inflection points	Direction changes in the contour
Endpoint count	Number of open stroke tips
Mean endpoint distance	Mean pairwise endpoint distance
Horizontal/vertical ratio	Relative stroke orientations
Enclosed area ratio	Fraction of enclosed glyph area

Appendix

Table 10: Structural descriptors used in the exploratory analysis.

ID	Description	Count
A	Monotone simplification	312
B	Non-monotone change	224
C	Radical reorganization	158
D	Abrupt transformation	187

Appendix

Table 11: Reported exploratory cluster membership counts.

Figure 5: Illustration of the reported exploratory cluster labels. The diagram is not an independent test of baseline failure mechanisms.

Cluster	Description	Generation error	CrossFont error
A	Monotone simplification	31.2	28.6
B	Non-monotone change	57.8	63.5
C	Radical reorganization	52.3	70.8
D	Abrupt transformation	68.4	41.2

Appendix

Table 12: Reported error percentages conditional on the exploratory cluster assignments.

Family	Configuration	Metric	Value
Retrieval	Baseline	R@1	62.8
Retrieval	+ Era fusion	R@1	65.2
Retrieval	+ Temporal constraint	R@1	67.4
Generation	Baseline	Top-1	41.0
Generation	+ Era conditioning	Top-1	47.3
Generation	+ Consistency loss	Top-1	51.8

Appendix

Table 13: Reported results for particular baseline augmentations. These values are not architectural performance ceilings.

Figure 6: Schematic rule-based matching pipeline. A schematic branching structure is not a proof that finite rule enumeration is mathematically impossible.

Figure 7: Continuous-time latent modeling under the time convention in Appendix K . The diagram does not establish recovery of a unique physical historical trajectory.

Figure 8: Spatial attribution examples across script periods. Highlighted regions describe the selected attribution procedure, not direct measurements of historical evolutionary laws.

Figure 9: Compatibility-score changes under high-saliency and low-saliency occlusions for the displayed examples. Effects vary across examples and should not be described as uniformly catastrophic or negligible.

Figure 10: Visualization of latent representations and selected trajectories. Projected cluster structure is an exploratory diagnostic rather than a proof of topological preservation or historical causality.

Figure 11: Retained, unvalidated attention-pattern visualization. The graphic shows 144 entries, but its correspondence to the reported checkpoint is not independently established here. Head preferences are descriptive, not causal or cognitive evidence.

Figure 12: Reference variant network before the illustrated filtering procedure.

Figure 13: Reference network after the illustrated filtering procedure. Reduced graph size does not establish the historical invalidity of removed variants.

Figure 14: Selected latent vector analogies. Individual examples do not establish a globally character-independent transformation operator or disentangled semantic components.

Record type	Definition
Character identity	Distinct correspondence label
Glyph observation	A recorded image in a script period
Correspondence pair	Two linked observed glyph records
Complete chain	Linked observations across all five periods

Appendix

Table 14: Definitions of counting units used for FGCCES.

Reported quantity	Value
Estimated velocity-field Lipschitz ratio	2.34
Estimated flow Lipschitz ratio	8.7
Maximum sampled velocity norm	2.51
Mean adjacent-era distance	0.82
Reported trajectory smoothness statistic	0.15

Appendix

Table 15: Reported finite-sample sensitivity diagnostics. Sampled maxima are not certified global bounds.

Metric	Value
OBI $\to$ Bronze $\to$ OBI error	0.023
OBI $\to$ Seal $\to$ OBI error	0.041
OBI $\to$ Regular $\to$ OBI error	0.068
Composition error	0.015

Appendix

Table 16: Reported numerical consistency measurements. They are distinct from candidate-identification accuracy.

Metric	OBI	Bronze	Seal	Clerical	Regular
Intra-radical distance	1.35	1.28	1.18	1.12	1.08
Inter-radical distance	3.52	3.68	3.85	3.92	4.05
Distance ratio	2.61	2.88	3.26	3.50	3.75
Silhouette score	0.45	0.48	0.52	0.55	0.58

Appendix

Table 17: Reported radical-conditioned geometry. These are descriptive clustering measurements, not proofs of semantic preservation.

Pair type	Velocity cosine	Trajectory angle
Same radical	0.78	$11.2^{\circ}$
Same semantic category	0.71	$15.8^{\circ}$
High co-occurrence	0.65	$19.3^{\circ}$
Different radical	0.35	$42.5^{\circ}$
Random pairs	0.31	$45.7^{\circ}$

Appendix

Table 18: Unvalidated velocity-cosine and trajectory-angle summaries. The two columns are distinct reported quantities; the angle estimator is unspecified and is not inferred by applying arccos to a mean cosine.

Reported quantity	Value
Mean absolute divergence	0.08
Maximum sampled absolute divergence	0.35
Estimated OBI-to-Regular volume ratio	1.12
Mean Jacobian determinant	1.08

Appendix

Table 19: Reported volume-related diagnostics. A sampled maximum divergence is not a certified supremum over the state space.

Transition	Mean norm	Standard deviation
OBI $\to$ Bronze	1.85	0.42
Bronze $\to$ Seal	1.42	0.31
Seal $\to$ Clerical	1.12	0.25
Clerical $\to$ Regular	0.45	0.12

Appendix

Table 20: Reported latent velocity magnitudes under the experimental coordinate system.

Era	Direction statistic	Clusters for 80%
OBI	0.42	5
Bronze	0.31	4
Seal	0.25	3
Clerical	0.18	2
Regular	0.12	2

Appendix

Table 21: Reported direction-dispersion diagnostics.

Type	OBI–Br	Br–Se	Se–Cl	Cl–Re
Pictographic	2.15	1.58	1.25	0.52
Indicative	1.92	1.45	1.18	0.48
Ideographic	1.78	1.42	1.12	0.45
Phono-semantic	1.65	1.35	1.05	0.42

Appendix

Table 22: Reported velocity norms by character-type annotation.

Figure 15: Schematic local neighborhood around an OBI “Ox” anchor. Nearby points illustrate similar variants or semantic associations; the distant “Wood” example illustrates a dissimilar reference. The surface is an illustration, not a measured radical-group scatter plot.

Field	Query side	Training/reference side
Glyph and provenance	Observed image and artifact metadata	Observed reference records
Answer identity	Excluded from model input	Training identities only; candidate labels are reference outputs
Dictionary definition	Query answer definition excluded; unavailable block zero-filled	Definitions of permitted reference identities
Structural annotation	Image-observable components only; no answer-dependent labels	Permitted reference annotations
Artifact context	Available neighbors; no hidden query answer or answer-equivalent variant	Separate from proof of artifact-disjoint splitting
Correspondence graph	No held-out query-to-answer edges	Only permitted reference correspondences

Appendix

Table 23: Intended information-access rules; execution is not certified by this table.

Setting	Value
Input dimension	352
Latent dimension	256
Encoder layers	12 Transformer layers
Velocity network	3-layer MLP
Velocity activation	Tanh
Velocity normalization	Spectral normalization

Appendix

Table 24: Architecture and optimization configuration.

Resource	Role in this comparison
Oracle-241	Character recognition
OBC306	Character recognition
OBI-IJDH	OBI image resource
OracleSage	Multimodal OBI resource
EVOBC	Multi-era character evolution resource
FGCCES	Cross-era correspondence records

Appendix

Table 25: Roles of selected existing resources and FGCCES. This table compares task focus, not dataset scale.

Method	1	2	3	5	8	10
OBS-OCR
OBSD	41.0	56.0	67.5	76.5	79.5	80.0
MSEF	71.5	80.5	85.0	88.5	90.5	91.0
PaddleOCR
OBSD	30.0	40.0	46.0	53.0	57.5	58.5
MSEF	58.5	67.0	72.5	77.0	79.5	80.5

Appendix

Table 26: Unvalidated multi-round score summaries (percent). Attempt generation, feedback, and duplicate handling are not identified by the supplied records; these values are not established cumulative hit rates or evidence of independent candidate diversity.

Evaluation	Comparator	MSEF	Difference
OBS-OCR Top-1	OBSD: 41.0	71.5	30.5
PictOBI accuracy	Gemini: 53.66	72.18	18.52
FGCCES R@1	OracleAgent: 62.8	72.5	9.7

Appendix

Table 27: Main-text score summaries under different prediction interfaces. Differences are arithmetic only: task adapters and information matching are not established, and the PictOBI aggregate has an unresolved subgroup-weighting convention.

Task	R@1	R@5
OBI $\to$ Bronze	$80.2\pm 0.9$	$90.5\pm 0.7$
Bronze $\to$ Seal	$85.5\pm 0.7$	$93.8\pm 0.5$
Seal $\to$ Clerical	$87.8\pm 0.6$	$95.2\pm 0.4$
Clerical $\to$ Regular	$89.2\pm 0.5$	$96.0\pm 0.3$
OBI $\to$ Seal	$74.5\pm 1.0$	$86.2\pm 0.8$
OBI $\to$ Clerical	$73.2\pm 1.1$	$85.0\pm 0.9$

Appendix

Table 28: Reported retrieval results by era pair.

Eras	R@1	R@5
OBI, Regular	$61.2\pm 1.2$	$76.5\pm 1.0$
+ Bronze	$65.7\pm 1.0$	$80.8\pm 0.8$
+ Seal	$69.5\pm 0.9$	$84.2\pm 0.7$
+ Clerical	$72.5\pm 0.8$	$86.5\pm 0.6$

Appendix

Table 29: Unvalidated era-augmentation summary A. This identifier denotes a report series, not a verified run configuration. Unreconciled pair counts are omitted.

Reported configuration	R@1
OracleSage	60.1
OracleAgent	62.8
MSEF, OBI–Modern only	$63.5\pm 1.1$
MSEF, complete chains only	$52.8\pm 1.3$
MSEF, + Bronze	$66.8\pm 1.0$
MSEF, + Seal	$70.2\pm 0.9$

Appendix

Table 30: Unvalidated restricted-supervision summary B. Configuration equivalence with summary A is not established; no matched-information advantage is inferred. Unreconciled pair/chain counts are omitted.

Assigned category	Share (%)
Glyph damage	45
Character splitting or correspondence ambiguity	28
Scribal variants	16
Other/model-related diagnostics	11

Appendix

Table 31: Reported coding of 200 failure cases.

Group	R@1	R@5
Bin	$75.8\pm 0.9$	$88.5\pm 0.7$
Li	$73.5\pm 1.0$	$86.8\pm 0.8$
Chu	$72.2\pm 1.1$	$85.5\pm 0.9$
Zi	$70.5\pm 1.2$	$83.8\pm 1.0$
Huang	$68.2\pm 1.3$	$81.5\pm 1.1$

Appendix

Table 32: Unvalidated score summaries for five named scribal groups. The count column is omitted because its counting unit is unresolved.

Method	Seconds/query	Parameters	GPU memory
OBSD	0.85	89M	8GB
OracleFusion	0.95	89M	8GB
OracleSage	1.8	7B	24GB
OracleAgent	2.3	7B	28GB
MSEF	0.18	140M	6GB

Appendix

Table 33: Unvalidated computational summaries. Common hardware, batching, precision, timing boundaries, and parameter-counting scope are not established; no cross-method speed or memory advantage is inferred.

Model	R@1	R@5
Diffusion PF-ODE	61.4	76.8
MSEF Neural ODE	72.5	86.5

Appendix

Table 34: Unvalidated retrieval summaries for two dynamics parameterizations. The previous cycle-error column is omitted because its integration paths are not matched.

Candidate type	Mean backward error	Removal	Retention
Valid targets	0.025	1.2%	98.8%
Distractors	0.142	82.4%	17.6%

Appendix

Table 35: Unvalidated candidate-filtering diagnostics. The analyzed candidate population and weighting are not established; the percentages are not independently verified estimates of the FPDR and TPR definitions in the text.

Figure 16: Top-5 Bronze retrieval candidates for 40 selected OBI queries. The green box indicates the annotated target. These examples illustrate retrieval behavior and are not a substitute for evaluation on the complete test population.

Figure 17: Local neighborhoods in the OBI and Bronze representations. Each row contains an anchor and its retrieved neighbors under the declared distance. The visualization characterizes selected neighborhoods rather than proving preservation of global topology.

Figure 18: Selected glyph observations with fine-grained metadata. OBI variants are annotated by scribal-group labels where available. Bronze variants use the displayed subperiod abbreviations L.S, E.Z, M.Z, L.Z, Sp, and Wa. The symbol “/” denotes an unavailable observation in the displayed collection; it does not establish historical extinction.

Field	Description
LBQJQW / jgwlbq
Use	Source for OBI glyph images and scribal-group information used in constructing FGCCES.
Source	https://www.jgwlbq.org.cn/
Terms	Publicly accessible academic/cultural-resource website. We found no standardized machine-readable open license in the cited source.
Release treatment	We cite the source and do not claim ownership of the original glyph images. Raw assets are redistributed only when permitted by the source terms; otherwise, the No jgwlbq assets or FGCCES-derived records are included in the current public package. Any future redistribution requires checking source terms and documenting item-level permissions.
CCAMC

Appendix

Table 36: External and newly introduced assets used in this work. Source terms and release treatment are separate from ownership.

Explore similar work

Jul 20, 2026cs.HC

AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning

Approximately 3,000 of the 4,500 oracle bone script (OBS) characters remain undeciphered due to fragmentary inscriptions and sparse evidence. Current AI approaches fail to replicate expert workflows that integrate form analysis, contextual semantics, and philological reasoning. We introduce AlphaOracle, a human-workflow-inspired framework that systematizes OBS decipherment using the largest digitized corpus to date. Its multi-stage pipeline comprises: (i) rubbing parsing; (ii) radical-based morphological analysis with diachronic modeling; (iii) contextual retrieval with semantic alignment; and (iv) philological validation against classical sources. Each stage generates explicit, confidence-weighted evidence chains, culminating in interpretable reports for scholarly verification. Across multiple test characters, AlphaOracle's readings strongly agreed with expert interpretations. In a study of 86 domain specialists, it reduced analysis time by 64% and 79% of participants rated it highly useful. Notably, AlphaOracle resolves the character "Lao" as a toponymic or clan designation, offering concrete revisions to Shang administrative and social interpretations. These results suggest that computational methods aligned with philological practice can facilitate OBS research and provide a conceptual reference for studies of other undeciphered scripts.

Yuliang Liu, Haisu Guan, Pengjie Wang +11

School of Software Engineering, Huazhong University of Science and Technology, Wuhan 430074, China · Key Laboratory of Oracle Bone Inscriptions Information Processing, Anyang Normal University, Anyang 455002, China · School of Electronic Information Engineering, South China University of Technology, Guangzhou 510641, China

Historical Manuscripts Independent-Pool Single-Draw Oracle Oracles Deep Learning

Ask PDF

Sep 15, 2026cs.CV

FROD: Feature Matching Residual Denoising Oracle Bone Decipher

Oracle bone script (OBS), one of the earliest Chinese writing systems, plays an important role in the study of Chinese etymology. Traditional decipherment relies heavily on domain experts who analyze characters through semantic context and structural evolution. To assist this labor-intensive process, we formulate OBS decipherment assistance as a cross-era image translation task and propose FROD (Feature Matching Residual Denoising Oracle Bone Decipher). Although many OBS characters differ substantially from their modern counterparts, they often preserve local topological invariants at the radical level. During training, FROD leverages fast feature matching to provide gated segmentation supervision: paired samples with sufficient matches are processed patch-wise to align fine-grained radicals, whereas low-similarity pairs are trained holistically to avoid mismatched artifacts. In addition, a Residual Denoising Diffusion Model (RDDM) jointly estimates noise and residual signals, thereby reducing the positional drift and stroke disorder commonly observed in standard diffusion models. Finally, a multi-stage font stylization refinement network refines the generated images by eliminating edge noise and stabilizing stroke structures. On our augmented character-disjoint dataset, FROD achieves higher Top-1 recognition accuracy than the evaluated baselines, with a 3.8% absolute gain over OBSD.

Yanbin Hou, Biao Xiong, Guojun Xu +4

School of Artificial Intelligence, Wuhan University of Technology, Wuhan 430070, China

Unpaired Image-To-Image Translation Iterative Denoising Methods Oracles

Ask PDF

Jun 30, 2026cs.CV

Beyond Single Character: Evaluating MLLMs for Sentence-Level Oracle Bone Inscription Understanding

Existing AI-assisted oracle bone inscription (OBI) visual recognition and understanding studies mainly focus on character-level, ignoring the long-form textual coherence and contextual dependencies embedded in complete divination charges. Recently, the powerful visual perception capabilities of multimodal large language models (MLLMs) have opened new possibilities for OBI information processing. In this work, we introduce S-OBI, a novel benchmark for evaluating MLLMs in Sentence-level OBI understanding. Instead of using noisy and incomplete rubbings as the visual input, S-OBI synthesizes clear and standardized sentence-level OBI instances through glyph substitution and composition. According to 95 original rubbings with translations that have been identified, corrected, and verified by experts, we replace characters in the original rubbings with corresponding clean glyph samples sourced from existing OBI datasets while preserving the overall inscriptional structure and semantic organization. This mitigates the influence of low-level distortions and enables a more focused evaluation of sentence-level OBI understanding. Based on this, we design semantic matching, semantic slot extraction, and contextual reasoning tasks and obtain 695 question-answer pairs. Experiments reveal the inferiority of contemporary MLLMs on sentence-level OBI understanding. In particular, visual perception errors in unmasked regions propagate through the reasoning chain, leading to erroneous predictions for masked characters, which indicates that sentence-level OBI understanding in current models remains strongly dependent on character-level recognition. Overall, S-OBI provides a diagnostic benchmark for evaluating whether MLLMs can move beyond isolated character recognition toward structured inscription-level understanding.

Ziqi Li, Zijian Chen, Tingzhu Chen +1

Nanjing University of Science and Technology, Nanjing 210094, China · Shanghai Jiao Tong University, Shanghai 200240, China · Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China

Historical Manuscripts Multimodal Large Language Models

Ask PDF