acceptodds
Desk rejected at ICLR 2027

Learning from Incomplete and Unaligned Lineages for Oracle Bone Script Decipherment

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Oracle Bone Script (OBS) decipherment seeks the established modern counterpart of an ancient glyph. Related forms preserved across later script periods provide chronologically ordered historical evidence that can help bridge this large visual gap. Yet a cascade of models trained between consecutive stages performs worse than a single model that maps OBS directly to the modern character. We found that the difficulty lies not in the existence of intermediate evidence but in how it is used. For most lineages are incomplete, and the variants of a character at one stage are not paired with those at the next, so each step of the cascade is trained on fragments and restricts cross-period context. We introduce ChronoOBS, which jointly models the six script periods as a chronologically ordered sequence with a shared autoregressive transformer and diffusion head. Missing stages are represented by placeholders, Dynamic Optimal Transport Matching (DOTM) constructs soft cross-period correspondences between unpaired variants, and multi-path decoding aggregates predictions over different subsets of intermediate periods. Under the same underlying architecture, the full ChronoOBS pipeline reaches 24.4% Top-1 accuracy on held-out character categories, compared with 21.8% for direct training under the same architecture and 20.5% for the strongest directly comparable prior image-generation method. On known character categories with period-specific references, the generated intermediate glyphs preserve character identity substantially better than cascaded trajectories. Code and models will be made available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Ancient Script Decipherment, Oracle Bone Script, Computational Paleography

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fTbrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ChronoOBS, a framework for oracle bone script decipherment using incomplete and unaligned historical glyph lineages. Its main contribution is to jointly learn from historical forms across script periods, rather than relying on direct mapping or independent stage-wise translation. The method achieves strong Top-1 decipherment accuracy, compared with direct mapping under the same architecture. Independent verification on known character categories also evaluates the period recognizability and character-identity consistency of generated intermediate trajectories.

Critical strengths
  1. The motivation is interesting: leveraging the context of script evolution to aid decipherment.
  2. The paper proposes an elegant model to realize this idea and achieves strong Top-1 accuracy.
Critical weaknesses
  1. Insufficient validation of intermediate trajectories on unseen character categories.The paper’s primary decipherment task focuses on category-level generalization, yet the intermediate trajectory verification in Section 4.3 and Appendix J is restricted to known character categories. The presented evidence does not establish whether the generated intermediate glyphs preserve character identity and remain structurally plausible on unseen categories. Independent assessment by paleographic experts is also lacking. Consequently, it remains unclear whether these generated trajectories serve as reliable intermediate representations in the core decipherment setting.

  2. Incremental methodological novelty.The overall framework builds on existing autoregressive diffusion approaches. The design of DOTM appears largely incremental, and it yields only a modest improvement over Static OT (path signatures).

Questions for authors

Why were intermediate trajectories not quantitatively evaluated on held-out deciphered categories? Is this due to insufficient period-specific reference glyphs or the chosen evaluation protocol?

9 comments
AuthorsOpenReview ↗

Response to Reviewer fTbr (Part 1)

Dear Reviewer fTbr,

Thank you for your valuable comments and suggestions. We have carefully addressed each comment and provide our point-by-point responses below.

Reply to weakness.

W1. Insufficient validation of intermediate trajectories on unseen character categories. The paper’s primary decipherment task focuses on category-level generalization, yet the intermediate trajectory verification in Section 4.3 and Appendix J is restricted to known character categories. The presented evidence does not establish whether the generated intermediate glyphs preserve character identity and remain structurally plausible on unseen categories. Independent assessment by paleographic experts is also lacking. Consequently, it remains unclear whether these generated trajectories serve as reliable intermediate representations in the core decipherment setting.

Thank you for highlighting these concerns. We have addressed both concerns by extending quantitative trajectory verification to held-out deciphered categories and adding independent expert assessment.

Specifically, we have expanded verifier training to cover all character categories available in each script period, including categories held out from generator training. This enables us to evaluate generated intermediate glyphs on held-out deciphered categories for which authentic period-specific records are available. Evaluation images are excluded from verifier training, and the verifiers are used only for evaluation. The expanded evaluation measures script-period recognizability, character-identity preservation, and retrieval against authentic glyphs, with the complete results shown below.

SplitMethod / referencePeriod acc.BI Top-5WSC Top-5SS Top-5CS Top-5Avg. Top-5R@1R@10mAPAll-period
ValidationAuthentic glyphs (GT)99.893.895.393.997.095.059.392.170.485.3
ValidationCascade75.456.758.250.548.953.612.518.014.89.7
ValidationChronoOBS w/o DOTM90.361.368.578.983.673.152.967.758.659.1
ValidationChronoOBS92.665.771.181.285.475.954.368.259.262.4
DeciphermentAuthentic glyphs (GT)99.894.294.893.997.295.057.592.370.185.0
DeciphermentCascade70.611.110.67.87.79.34.48.06.22.9
DeciphermentChronoOBS w/o DOTM87.017.719.521.421.720.113.618.215.212.5
DeciphermentChronoOBS88.918.920.522.823.421.414.118.515.713.1

All-period consistency requires the correct character to appear in the verifier's Top-5 at each of the four intermediate periods. On held-out categories, ChronoOBS achieves 88.9% period accuracy, 21.4% mean character Top-5 accuracy, and 13.1% all-period consistency, compared with 70.6%, 9.3%, and 2.9% for Cascade. These results show that the improvement in period recognizability and character-identity preservation extends to categories unseen by the generator. The authentic-glyph rows provide a reference. For these zero-shot categories, identity accuracy at the intermediate periods (18.9–23.4%) is in line with the final decipherment accuracy itself (24.4%).

In the revised manuscript, we have extended Sec. 4.3, “Intermediate Trajectory Verification,” and added the “Decipherment” block to Table 3. Sec. 4.3 and Appendix J now explicitly state that verifier training covers all available character categories in each script period, with evaluation images excluded from training.

AuthorsOpenReview ↗

Response to Reviewer fTbr (Part 2)

To assess structural plausibility beyond automated recognition, we have also added a study with nine experts covering two evaluation tasks: evaluation on held-out deciphered categories and evaluation on real-world undeciphered or disputed characters. Each expert assessed ten randomly sampled OBS glyphs, with five for each task.

For the first task, experts compared trajectories generated by Cascade, ChronoOBS w/o DOTM, and full ChronoOBS. They judged the plausibility of each proposed reading with a yes/no response and rated trajectory consistency and glyph quality on a four-point scale (1 = Poor, 2 = Fair, 3 = Good, 4 = Excellent). Trajectory consistency assesses structural coherence across script periods, while glyph quality assesses visual clarity and well-formedness. We report the percentage of yes responses and the mean rating, with 45 assessments per method and criterion.

CriterionCascadeChronoOBS w/o DOTMChronoOBS
Readings judged plausible (%)6.737.846.7
Trajectory consistency (1–4)1.512.933.11
Glyph quality (1–4)1.423.023.11

ChronoOBS receives the highest scores on all three criteria. In particular, experts rate its trajectory consistency at 3.11/4, compared with 1.51/4 for Cascade and 2.93/4 without DOTM, while its glyph quality is also rated 3.11/4. These judgments complement the automatic identity-preservation results with human assessment of the structural coherence and visual quality of trajectories on held-out categories.

For the second task, we assessed ChronoOBS on real-world undeciphered or disputed characters using the same evaluation setup, with research usefulness as an additional metric:

CriterionChronoOBS
Readings judged plausible (%)31.1
Trajectory consistency (1–4)2.96
Glyph quality (1–4)2.84
Research usefulness (1–4)2.45

On real-world undeciphered or disputed characters, experts rated trajectory consistency at 2.96/4 and glyph quality at 2.84/4, both close to 3 (Good). These assessments indicate that ChronoOBS maintains coherent intermediate trajectories and good generation quality in this setting.

We have added Sec. 4.6, “Expert Evaluation,” and Tables 7–8 to the revised manuscript to report the study protocol and results. Appendix N provides the evaluation-criterion definitions and detailed score distributions.

AuthorsOpenReview ↗

Response to Reviewer fTbr (Part 3)

W2. Incremental methodological novelty.The overall framework builds on existing autoregressive diffusion approaches. The design of DOTM appears largely incremental, and it yields only a modest improvement over Static OT (path signatures).

We would like to clarify where our contributions lie, which go beyond applying existing autoregressive diffusion and optimal transport techniques.

First, the contribution is not the autoregressive-diffusion backbone itself but how incomplete historical evidence is used. Under the same architecture, cascading through the intermediate periods is worse than ignoring them (9.8% vs. 21.8%), auxiliary supervision changes little (22.1%), and joint modeling is what helps (24.4%). Unlike MAR, where autoregression runs over tokens within one image, ChronoOBS runs it across script periods, with placeholders and period dropping so that incomplete lineages can be used for training and decoded along multiple paths. The surviving historical forms provide valuable evidence for bridging the gap between ancient and modern glyphs. Their organization into discrete, chronologically ordered stages motivates our choice: autoregressive modeling captures cross-period dependencies, while diffusion represents the diversity of glyph shapes within each stage.

Second, we would also like to clarify that the entry labeled “Static OT (path signatures)” refers to the static matching formulation developed in this work: the path-signature cost is the static term of DOTM's cost (Sec. 3.3, Eq. 8). To the best of our knowledge, this work is the first to combine path signatures with optimal transport to establish soft correspondences among ancient Chinese glyph variants across historical script periods. We therefore view this static formulation as one part of our contribution to cross-period glyph matching; its comparison with DOTM examines the additional effect of the adaptive extension within our approach. Against matching baselines that are not ours (Table 4), random matching reaches 22.9% and static OT with standard pixel or VGG costs 23.1–23.2%, whereas our path-signature cost reaches 23.7% and full DOTM 24.4%. The matching design therefore accounts for 1.5 of the 2.6-point gain over Direct (21.8%).

Third, DOTM addresses the difficulty that fixed morphological similarity may not fully capture evolutionary relationships. It updates matching distances using the cross-period structure learned during training and assesses correspondences jointly across whole trajectories. We see its contribution in allowing glyph alignment and evolution modeling to inform one another, so that matching reflects both morphological similarity and compatibility within an evolutionary trajectory.

In summary, our contribution is driven by a specific decipherment challenge: making use of historical evidence with missing stages and unknown variant correspondences. The innovation lies in the problem formulation and the adaptation of these techniques into a joint framework, where alignment and evolution modeling inform one another to make incomplete historical lineages useful for oracle bone script decipherment.

Reply to questions.

Q1. Why were intermediate trajectories not quantitatively evaluated on held-out deciphered categories? Is this due to insufficient period-specific reference glyphs or the chosen evaluation protocol?

Thank you for raising this question. This was due to the original evaluation protocol rather than a lack of reference glyphs: our trajectory verifiers covered only known character categories and therefore could not assess character identity on held-out decipherment categories. Authentic intermediate-period glyphs are available for some held-out categories.

We have now expanded verifier training to cover all character categories available in each script period, with evaluation images excluded from training, and extended trajectory evaluation to held-out deciphered categories with available reference glyphs. As detailed in our reply to W1, we have added these results to Sec. 4.3, Table 3, and clarified the verifier protocol in Appendix J of the revised manuscript.

Reviewer_HohSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Oracle Bone Script (OBS) decipherment using incomplete and unaligned across script periods Chinese characters. The central claim is that cascading models between periods shows lower performance because the variants from the adjacent periods are not paired. For this reason, the paper proposes ChronoOBS which models six script periods jointly using a shared autoregressive transformer and diffusion head. The missing periods are treated as placeholders, while Dynamic Optimal Transport Matching uses the static features and the dynamic updates to construct connections between the unpaired variants. The model additionally uses multi-path decoding that aggregates predictions over different subsets of intermediate periods. The evaluation is extensive. ChronoOBS reaches 24.4% Top-1 accuracy on held-out character categories, compared with 21.8% for direct training under the same architecture and 20.5% for the strongest directly comparable prior generative baseline method. The paper includes controls, ablations and robustness experiments as well as independent validation of generated trajectories on known categories.

Critical strengths

DOTM is an interesting application, because it addresses the fact that many variants appear within a character-category but their correspondences between periods do not appear. The various controls in the experimental analysis increase my confidence that the joint aspect accounts for the improvement.

Critical weaknesses

The paper generalizes decipherment to known deciphered categories, but does not show real-world decipherment and this should be very clear from the beginning. The significant weakness in this paper is already admitted in the Limitations section (Appendix!) and that is the lack of expert validation for the undeciphered cases. Another major weakness is that it is rather obscured whether the reported accuracy is balanced/macro-F1 and not clarified if the reporting numbers are affected by class frequency.

Questions for authors
  1. Is the reported accuracy micro-averaged or macro-averaged?
  2. Are any architectural choices made based on held-out categories? Please clarify in the text.
  3. Have you examined whether the learned correspondences are stable across seeds, initializations, etc? This would help understand whether DOTM learns correspondences that are robust rather than model-specific correspondences.
6 comments
AuthorsOpenReview ↗

Response to Reviewer HohS (Part 1)

Dear Reviewer HohS,

Thank you for your valuable comments and suggestions. We have carefully addressed each comment and provide our point-by-point responses below.

Reply to weakness.

W1. The paper generalizes decipherment to known deciphered categories, but does not show real-world decipherment and this should be very clear from the beginning.

Thank you for pointing this out. Our quantitative evaluation follows the standard protocol of computational OBS decipherment: models are tested on held-out deciphered categories whose established readings serve as ground truth [1-3]. A reading proposed for a genuinely undeciphered character has no ground truth, and its acceptance requires long-term scrutiny by the paleographic community, so it cannot serve as a quantitative benchmark for any method. This scope was stated in Sec. 4.1 of the original submission.

To make this scope clear from the beginning, we have added two sentences to Sec. 1 (Introduction) of the revised manuscript, in the paragraph introducing computational OBS decipherment, specifying that our quantitative benchmark tests recovery of established readings for categories unseen during training, while outputs for undeciphered or disputed glyphs are candidate readings requiring further validation.

We would also like to clarify that the original submission included exploratory demonstrations on real undeciphered or disputed characters in Appendix M, “Candidate Clues for Undeciphered and Disputed OBS.” Figures 9–11 in the original submission (Figures 10–12 in the revised manuscript) present generated trajectories for 36 undeciphered or disputed OBS examples, together with stochastic modern-script hypotheses for genuinely undeciphered categories from HUST-OBC and EVOBC. These examples illustrate the model's application to unresolved material and its potential to provide candidate clues for decipherment.

To complement these qualitative examples, we have added an expert evaluation in Sec. 4.6, “Expert Evaluation,” of the revised manuscript. We describe its protocol and results in detail in our reply to W2 below.

[1] Gechuan Zhang et al. 2021. Deciphering Ancient Chinese Oracle Bone Inscriptions Using Case-Based Reasoning. In Case-Based Reasoning Research and Development: 29th International Conference, ICCBR 2021, pages 309–324, Salamanca, Spain. Springer.

[2] Kaixin Peng et al. 2026. Interpretable Oracle Bone Script Decipherment via Progressive Glyph Analysis with LVLMs. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 32654–32674, Budapest, Hungary. Association for Computational Linguistics.

[3] Haisu Guan et al. 2024. Deciphering Oracle Bone Language with Diffusion Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15554–15567, Bangkok, Thailand. Association for Computational Linguistics.

AuthorsOpenReview ↗

Response to Reviewer HohS (Part 2)

W2. The significant weakness in this paper is already admitted in the Limitations section (Appendix!) and that is the lack of expert validation for the undeciphered cases.

Thank you for highlighting this concern. As explained in our reply to W1, readings for undeciphered characters have no ground truth, so quantitative validation is not possible for any method; expert judgment is the only evidence available. We have therefore added a study with nine experts, each assessing ten randomly sampled OBS glyphs: five from held-out deciphered categories and five from real-world undeciphered or disputed characters.

For held-out deciphered glyphs, experts compared Cascade, ChronoOBS w/o DOTM, and ChronoOBS. They assessed reading plausibility with a yes/no judgment and rated trajectory consistency and glyph quality on a four-point scale (1 = Poor, 2 = Fair, 3 = Good, 4 = Excellent):

CriterionCascadeChronoOBS w/o DOTMChronoOBS
Readings judged plausible (%)6.737.846.7
Trajectory consistency (1–4)1.512.933.11
Glyph quality (1–4)1.423.023.11

For real-world undeciphered or disputed glyphs, experts assessed ChronoOBS's hypotheses and additionally rated their usefulness to research:

CriterionChronoOBS
Readings judged plausible (%)31.1
Trajectory consistency (1–4)2.96
Glyph quality (1–4)2.84
Research usefulness (1–4)2.45

These results provide an initial expert assessment of both the proposed readings and their intermediate trajectories. On real-world undeciphered or disputed characters, experts judged 31.1% of the proposed readings plausible and rated trajectory consistency at 2.96/4, supporting their value as candidate clues for further paleographic study.

We have added Sec. 4.6, “Expert Evaluation,” to the main text of the revised manuscript, including the study protocol and complete results in Tables 7–8. We have also added Appendix N with further evaluation details and score distributions. In Appendix A (Limitations), we have clarified that expert ratings for real-world undeciphered or disputed characters assess the perceived plausibility and research usefulness of candidate readings, while confirming a proposed reading still requires independent historical evidence.

W3. Another major weakness is that it is rather obscured whether the reported accuracy is balanced/macro-F1 and not clarified if the reporting numbers are affected by class frequency.

Thank you for raising this issue. The reported validation and held-out decipherment prediction accuracies are macro-averaged across character categories, with five test samples per category. Each category therefore receives equal weight in the reported accuracy; categories with more glyphs in the source corpus do not receive proportionally more evaluation weight. Because every category contributes the same number of test samples, micro- and macro-averaged accuracy coincide, so the reported numbers are not affected by class frequency. This metric is macro-averaged accuracy, rather than macro-F1.

We have added the explicit statement “We report macro-averaged accuracy across categories, with five test samples per category” to Sec. 4.1, “Experimental Setup,” of the revised manuscript.

Reply to questions.

Q1. Is the reported accuracy micro-averaged or macro-averaged?

Thank you for the question. Macro-averaged. As explained in our reply to W3, we report macro-averaged accuracy across character categories, with five test samples per category.

Q2. Are any architectural choices made based on held-out categories? Please clarify in the text.

Thank you for raising this question. No. For the generative model, architecture, hyperparameters, and inference-path choices were determined using only the training data and the known-category validation split. We trained for a fixed 1000 epochs and used the final checkpoint for evaluation, without checkpoint selection. Held-out decipherment categories were reserved for final evaluation and did not guide those decisions.

We have clarified this protocol in Appendix C, “Implementation Details,” of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer HohS (Part 3)

Q3. Have you examined whether the learned correspondences are stable across seeds, initializations, etc? This would help understand whether DOTM learns correspondences that are robust rather than model-specific correspondences.

Thank you for this helpful suggestion. Yes. We have added a stability study with three training seeds for each of two static-cost initializations, pixel-space L2L_2 and path signatures, for six runs in total.

Every five epochs, we compare the learned trajectory probability distributions using Jensen–Shannon divergence (JSD, in bits) and probability overlap. Within an initialization, we compute the metrics for all three seed pairs separately within each character category, average the pairwise scores, and then average across categories. For the cross-initialization comparison, we first average each trajectory's probability over the three seeds within each category and initialization, then compare the resulting two distributions and average across categories. The complete stability results include both the final 200 epochs and the full training course:

ComparisonFinal 200 epochs: mean JSD (bits) ↓Final 200 epochs: mean overlap (%) ↑Full training: peak JSD (bits) ↓Full training: minimum overlap (%) ↑
Across seeds, pixel-space L2L_2 initialization0.007092.960.075976.81
Across seeds, path-signature initialization0.006693.160.081075.76
Across initializations, using seed-averaged distributions0.084271.780.177354.15

The table shows that, over the final 200 epochs, mean cross-seed JSD is low (0.0070 and 0.0066 bits) and mean overlap is high (92.96% and 93.16%) under the two initializations. Although the learned distributions differ across initializations, the seed-averaged distributions still share 71.78% of their probability mass on average over the final 200 epochs. Together with the high cross-seed agreement, this shows that DOTM's learned correspondences are reproducible across seeds and largely preserved when the static cost is changed.

We have added Figure 6 and Table 12 in Appendix E.3 of the revised manuscript to show how cross-seed JSD and probability overlap evolve throughout training and to provide the complete stability results. Under the same initialization, mean cross-seed JSD starts close to zero, rises with fluctuations during early and middle training, and peaks at epoch 470 for pixel-space L2L_2 and epoch 335 for path signatures. It then declines and settles at a low level with small fluctuations over the final 200 epochs. Probability overlap follows the opposite pattern: it starts high, falls to its minimum at the corresponding JSD peak epoch, and then recovers to a high level. Thus, the curves show that cross-seed differences increase temporarily during training and subsequently diminish.

In the revised manuscript, we have summarized the late-training results in the new Sec. 4.5, “Stability of DOTM Correspondences.”

Reviewer_LzK1OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

ChronoOBS predicts modern character forms from oracle-bone glyphs by jointly modeling six script periods. It combines an autoregressive diffusion model, placeholders for missing periods, dynamic optimal-transport matching, and voting across inference paths. Across three category-level splits, it achieves 24.4% held-out Top-1 accuracy, compared with 21.8% for the same-architecture direct model.

Critical strengths
  1. The controlled comparisons address the central question of how intermediate evidence should be used. The same-architecture direct and auxiliary-supervision baselines, together with fixed-path and matched-generation-budget controls, show that the gains are not explained simply by auxiliary targets or additional stochastic samples. This provides meaningful evidence for joint conditioning and path diversity.

  2. The evaluation separates category generalization from recognition of new samples within known categories. Held-out categories are excluded across script periods from training and matching. Three complete split-specific runs, an alternative OCR evaluator, and independently trained trajectory verifiers provide useful checks beyond the main accuracy table.

Critical weaknesses
  1. The intermediate trajectories are not independently validated on held-out categories. The trajectory verifier covers known categories, whereas held-out performance is measured through recognition of the final modern glyph. The authors appropriately describe generated trajectories as hypotheses, but the current evidence does not establish that these trajectories remain meaningful on the categories requiring generalization. This limits their value as interpretable evidence for decipherment.

  2. The alternative OCR evaluation does not include the strongest same-architecture controls. Table 8 retains an advantage over prior generators, but omits Direct and Direct + auxiliary. Since the central controlled gain is 2.6 percentage points over Direct and the decoder uses OCR-based voting, these comparisons are important for establishing whether the benefit of joint lineage modeling persists under an independent recognizer.

Questions for authors

For held-out deciphered categories with available intermediate records, is there any comparison between the generated intermediate forms and those authentic records, independent of the known-category verifier?

5 comments
AuthorsOpenReview ↗

Response to Reviewer LzK1 (Part 1)

Dear Reviewer LzK1,

Thank you for your valuable comments and suggestions. We have carefully addressed each comment and provide our point-by-point responses below.

Reply to weakness.

W1. The intermediate trajectories are not independently validated on held-out categories. The trajectory verifier covers known categories, whereas held-out performance is measured through recognition of the final modern glyph. The authors appropriately describe generated trajectories as hypotheses, but the current evidence does not establish that these trajectories remain meaningful on the categories requiring generalization. This limits their value as interpretable evidence for decipherment.

Thank you for highlighting this concern. We have addressed this by expanding the verifiers' category coverage and extending trajectory evaluation to held-out deciphered categories for which authentic intermediate-period glyphs are available. Specifically, the verifiers are now trained on real glyphs from all character categories available in each script period, including categories held out from generator training. Evaluation images are excluded from verifier training, and the verifiers are used only for evaluation. This expanded coverage enables us to evaluate the character identity of generated intermediate glyphs on held-out categories as well as known categories.

The complete trajectory-verification results are shown below, including both validation and held-out decipherment categories, all four period-specific character-recognition scores, and the character-retrieval metrics.

SplitMethod / referencePeriod acc.BI Top-5WSC Top-5SS Top-5CS Top-5Avg. Top-5R@1R@10mAPAll-period
ValidationAuthentic glyphs (GT)99.893.895.393.997.095.059.392.170.485.3
ValidationCascade75.456.758.250.548.953.612.518.014.89.7
ValidationChronoOBS w/o DOTM90.361.368.578.983.673.152.967.758.659.1
ValidationChronoOBS92.665.771.181.285.475.954.368.259.262.4
DeciphermentAuthentic glyphs (GT)99.894.294.893.997.295.057.592.370.185.0
DeciphermentCascade70.611.110.67.87.79.34.48.06.22.9
DeciphermentChronoOBS w/o DOTM87.017.719.521.421.720.113.618.215.212.5
DeciphermentChronoOBS88.918.920.522.823.421.414.118.515.713.1

All-period consistency requires the correct character to appear in the verifier's Top-5 at each of the four intermediate periods. On held-out categories, ChronoOBS improves mean character Top-5 accuracy over Cascade by 12.1 percentage points and all-period consistency by 10.2 points. The authentic-glyph rows provide a reference for interpreting these recognition and retrieval scores. Together, these results show improved character-identity preservation on categories unseen by the generator. For these zero-shot categories, identity accuracy at the intermediate periods (18.9–23.4%) is in line with the final decipherment accuracy itself (24.4%), while period accuracy remains high (88.9%).

To complement this evaluation, we obtained trajectory-consistency judgments from nine experts. In the held-out portion of the study, each expert assessed five randomly sampled OBS glyphs from held-out deciphered categories, comparing the trajectories generated by Cascade, ChronoOBS without DOTM, and full ChronoOBS. Cross-period trajectory consistency was rated on a four-point scale (1 = Poor, 2 = Fair, 3 = Good, 4 = Excellent), and we report the mean expert rating. ChronoOBS scored 3.11/4, compared with 1.51/4 for Cascade and 2.93/4 without DOTM. These expert judgments complement the automatic evaluation of identity preservation on held-out categories.

In the revised manuscript, we have expanded the trajectory evaluation in Sec. 4.3, “Intermediate Trajectory Verification,” and added a separate “Decipherment” block to Table 3, reporting recognition and retrieval results alongside the authentic-glyph references. Sec. 4.3 and Appendix J now explicitly state that verifier training covers all available character categories in each script period, with evaluation images excluded from training. The expert-study protocol and the held-out trajectory-consistency comparison are reported in Sec. 4.6, “Expert Evaluation,” Table 7.

AuthorsOpenReview ↗

Response to Reviewer LzK1 (Part 2)

W2. The alternative OCR evaluation does not include the strongest same-architecture controls. Table 8 retains an advantage over prior generators, but omits Direct and Direct + auxiliary. Since the central controlled gain is 2.6 percentage points over Direct and the decoder uses OCR-based voting, these comparisons are important for establishing whether the benefit of joint lineage modeling persists under an independent recognizer.

Thank you for the constructive comment. We have added both Direct and Direct + auxiliary to the PP-OCRv5 evaluation.

This comparison retains the generated images, category splits, budget of 16 final RS outputs, and majority-first voting rule. Only the post-hoc OCR recognizer changes; voting uses the labels and confidences returned by that recognizer. Direct and Direct + auxiliary both use 16-seed voting.

MethodValidation: RS-OCRValidation: PP-OCRv5Decipherment: RS-OCRDecipherment: PP-OCRv5
Pix2Pix5.34.30.00.0
Palette7.96.60.00.0
DRIT++11.69.90.10.1
CycleGAN18.415.80.30.2
Sundial-GAN44.343.96.64.8
BBDM55.848.88.26.1
CDE61.954.517.413.4
OBSD69.461.020.516.2
ChronoOBS (Direct)84.672.821.817.1
ChronoOBS (Direct + auxiliary)84.873.122.117.2
ChronoOBS (Joint)86.776.624.419.4

Under PP-OCRv5, Joint improves held-out decipherment over Direct by 2.3 percentage points and over Direct + auxiliary by 2.2 points. On validation, the gains are 3.8 and 3.5 points, respectively. The ordering therefore persists under the independent recognizer, although absolute scores differ. This supports the conclusion that the controlled advantage is not confined to RS-OCR.

We have added the two controls in Appendix D, “PaddleOCR Evaluation,” Table 11 of the revised manuscript (Table 8 in the original submission).

Reply to questions.

Q1. For held-out deciphered categories with available intermediate records, is there any comparison between the generated intermediate forms and those authentic records, independent of the known-category verifier?

Thank you for raising this question. Yes. As detailed in our reply to W1, we have expanded verifier training to cover all character categories available in each script period, including categories held out from generator training. This expanded coverage enables us to evaluate generated intermediate glyphs on held-out deciphered categories with available authentic records. In particular, the character-retrieval metrics in Table 3 (R@1, R@10, mAP) use each generated intermediate glyph as a query against authentic glyphs, giving a direct comparison with the authentic records of held-out categories (R@1: 14.1% for ChronoOBS versus 4.4% for Cascade).

We sincerely appreciate the reviewer’s time, effort, and constructive feedback. We hope that our responses have adequately addressed the reviewer’s concerns, and we would be happy to provide any further clarification if needed.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We would like to express our sincere gratitude to all the reviewers and ACs for their time and effort in providing detailed feedback on our paper. We appreciate the thoughtful and constructive comments, as well as the recognition of the strengths of our work. We are particularly encouraged by the reviewers' acknowledgment of the motivation, the model design, and the controlled experimental evidence. For example:

  • The controlled comparisons address the central question of how intermediate evidence should be used... This provides meaningful evidence for joint conditioning and path diversity. (Reviewer LzK1)
  • The various controls in the experimental analysis increase my confidence that the joint aspect accounts for the improvement. (Reviewer HohS)
  • DOTM is an interesting application, because it addresses the fact that many variants appear within a character-category but their correspondences between periods do not appear. (Reviewer HohS)
  • The motivation is interesting: leveraging the context of script evolution to aid decipherment. The paper proposes an elegant model to realize this idea and achieves strong Top-1 accuracy. (Reviewer fTbr)

We also thank the reviewers for their constructive feedback on potential improvements of this work. We have addressed all the questions and concerns in the individual responses. Additionally, we have uploaded a revised version of the manuscript, where all modifications are highlighted in BLUE. A summary of these modifications is as follows:

  • Trajectory Verification on Held-out Categories. We have extended the independent trajectory verification from known categories to held-out deciphered categories. The verifiers now cover all character categories available in each script period and are used only for evaluation. Table 3 includes a new “Decipherment” block reporting period accuracy, character identity, retrieval against authentic glyphs, and all-period consistency, which shows that the advantage of joint modeling over cascading extends to categories unseen by the generator (Sec. 4.3, Appendix J). (reviewers LzK1, fTbr)
  • Expert Evaluation. We have conducted a study with nine experts on both held-out deciphered characters and real-world undeciphered or disputed characters, covering reading plausibility, trajectory consistency, glyph quality, and research usefulness. This complements the automatic evaluation with expert judgment, the only evidence available for characters without ground truth (Sec. 4.6, Tables 7–8, Appendix N). (reviewers HohS, fTbr)
  • Same-architecture Controls under an Independent OCR Evaluator. We have added Direct and Direct + auxiliary to the PP-OCRv5 evaluation. The controlled advantage of joint modeling persists under the independent recognizer: 19.4% versus 17.1% and 17.2% on held-out categories (Appendix D, Table 11). (reviewer LzK1)
  • Stability of DOTM Correspondences. We have added a stability analysis with three training seeds for each of two static-cost initializations. The learned correspondences are highly consistent across seeds, with about 93% probability overlap over the final 200 epochs (Sec. 4.5, Appendix E.3). (reviewer HohS)
  • Evaluation Scope and Protocol. We have clarified in the Introduction that our quantitative benchmark tests recovery of established readings on held-out deciphered categories. This is the standard protocol in computational OBS decipherment, since readings of genuinely undeciphered characters have no ground truth; outputs for such characters are candidate readings. We have also specified that accuracy is macro-averaged across categories with five test samples per category (Sec. 4.1), and that no architectural or hyperparameter choice used held-out categories (Appendix C). (reviewer HohS)

We hope these clarifications address the reviewers' concerns. We are open to any further comments or questions and look forward to hearing from the reviewers during the discussion period. Thank you once again for your valuable feedback.

AuthorsOpenReview ↗

Final General Response

We sincerely thank all three reviewers and the Area Chair for their time and constructive feedback. We appreciate the reviewers’ recognition of the motivation, model design, and controlled comparisons. Below we summarize the main revisions and follow-up analyses:

  • Trajectory Verification on Held-out Categories. We extended independent verification of intermediate trajectories to deciphered character categories held out from generator training. ChronoOBS outperformed Cascade in script-period recognition, character identity preservation, retrieval against authentic glyphs, and consistency across all four intermediate periods (Sec. 4.3, Table 3). Our follow-up analysis also found that stronger identity preservation across periods was associated with higher final decipherment accuracy.
  • Expert Evaluation. Nine specialists assessed held-out deciphered glyphs and real-world undeciphered or disputed glyphs. In the follow-up, we documented their qualifications and reported the full rating distributions for unresolved cases (Sec. 4.6, Appendix N).
  • Independent OCR Controls. We added the same-architecture Direct and Direct + auxiliary baselines to the PP-OCRv5 evaluation. The joint model retained its advantage on held-out categories (19.4% vs. 17.1% and 17.2%; Table 11).
  • Stability of DOTM Correspondences. Using three seeds per initialization, we observed about 93% cross-seed probability overlap within each initialization and 72% overlap between the two seed-averaged distributions over the final 200 epochs (Sec. 4.5).
  • Evaluation Scope and Protocol. We clarified that the benchmark measures recovery of established readings using macro-averaged accuracy with five test samples per category. Generator design and tuning used only training and known-category validation data. Proposed readings for unresolved glyphs require further paleographic validation.
  • Methodological Contributions. We clarified how joint modeling uses incomplete lineages and explained the respective roles of static path-signature matching and DOTM’s adaptive updates in aligning unpaired glyph variants.

We are encouraged that all three reviewers gave positive final ratings (3: Weak acceptance). We will carefully polish the camera-ready manuscript in light of the reviewers’ suggestions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.