{"id":"30ccd86b-14fb-4dd0-a5e9-76a4bed6e68b","arxiv_id":"2508.10113","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A vision-language model trained with progressive radical and pictographic analysis plus dual matching achieves state-of-the-art zero-shot decipherment of Oracle Bone Script.","lead":"This paper uses large vision-language models to decipher Oracle Bone Script by reasoning about the pictures and radicals inside each ancient character. If the method works as claimed, it could give archaeologists a fast, interpretable tool for testing readings of undeciphered Chinese bronze-age writing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot/SOTA claims depend on PD-OBS being disjoint from benchmark test sets; no decontamination is described, so reported accuracy may reflect memorization rather than decipherment.","rationale":"The reader's UNVERDICTED verdict stems from inability to read the full text; I agree no acceptance is warranted. However, the abstract itself provides enough to identify a specific, checkable risk: PD-OBS is the training data, and the claimed zero-shot test must be disjoint from it. The reader's weakest assumption focused on annotation accuracy and comprehensiveness; the more load-bearing failure mode is benchmark contamination or label leakage through identical/near-identical glyphs, which would invalidate both the Top-10 and zero-shot numbers. I therefore recommend making acceptance conditional on a decontamination audit. I am not claiming leakage occurred; I am claiming it is the decisive unverified condition. If the audit shows zero overlap and accuracy is stable, the central empirical claims pass this objection.","tokens_in":7880,"tokens_out":5704,"duration_ms":61018,"concrete_test":"After the dataset is released, perform a decontamination audit: (1) compare the set of modern Chinese character labels in PD-OBS against the label sets of every public benchmark test split; (2) run perceptual hashing (e.g., pHash/dHash) on all OBS glyph images in PD-OBS and each test split to flag exact and near-duplicate images. Re-run the reported Top-10 and zero-shot evaluations restricted to benchmark instances with zero label overlap and no near-duplicate glyph in PD-OBS. If any overlap exists and the restricted accuracy drops materially (e.g., more than 5 points), the zero-shot/SOTA claims are not established by the current experiments.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central empirical claim is state-of-the-art Top-10 accuracy and superior zero-shot decipherment, obtained with a model trained on the authors' 47,157-character PD-OBS dataset. For that claim to hold, the public benchmark test images/characters—especially held-out 'undeciphered' classes—must not appear in PD-OBS, and no test answers may leak through near-duplicate OBS images or modern-character labels. The abstract states no decontamination or split-disjointness procedure, and the garbled full text makes it impossible to verify the construction of PD-OBS or the benchmark splits. Because PD-OBS pairs OBS images with modern Chinese characters, if it was assembled from the same public OBS corpora as the benchmarks, the top-10 retrieval could be produced by image/label memorization. The 'logical analysis' would then be post-hoc rationalization of a retrieval rather than evidence of decipherment. This is a correctness risk in the empirical basis, not a disagreement with the field's consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an interpretable Oracle Bone Script (OBS) decipherment method based on Large Vision-Language Models (LVLMs). It combines radical analysis and pictograph-semantic understanding via a progressive training strategy (radical recognition, radical analysis, pictographic analysis, mutual analysis) and a Radical-Pictographic Dual Matching mechanism for ranking candidate Chinese characters. To support training, the authors introduce the Pictographic Decipherment OBS Dataset (PD-OBS) containing 47,157 Chinese characters annotated with OBS images and pictographic analysis texts. The abstract claims state-of-the-art Top-10 accuracy and superior zero-shot decipherment on public benchmarks, and argues that the model's logical analysis outputs may provide archaeologically valuable references for undeciphered OBS.","tokens_in":8154,"tokens_out":2116,"duration_ms":24392,"significance":"If the claimed results hold, the paper would be a meaningful step toward interpretable and generalizable OBS decipherment: the proposed dataset addresses a real data bottleneck, the progressive radical-to-pictograph training is a sensible pedagogical ordering for LVLMs, and the emphasis on zero-shot evaluation is the right target for archaeological discovery. The manuscript also promises release of dataset and code, which would benefit the community. However, the significance cannot currently be assessed because the full text is corrupted/unreadable and the abstract contains no quantitative results, evaluation protocols, or baselines. The core contribution is an empirical claim without presented evidence.","major_comments":[{"comment":"The body of the manuscript is garbled and unreadable (encoding corruption). No methods, experimental setup, tables, figures, or equations can be verified. This is a load-bearing issue because the paper's central claims—SOTA Top-10 accuracy, superior zero-shot decipherment, and 'logical analysis'—are empirical and currently supported only by the abstract. The authors must provide a readable manuscript with the full method description, evaluation protocol, and numbered results.","section":"Full text (all sections after Abstract)"},{"comment":"The claimed zero-shot and SOTA results depend on the PD-OBS training data being disjoint from the public benchmark test sets, especially any held-out 'undeciphered' characters. The manuscript does not describe any decontamination procedure, image deduplication, or label-overlap analysis between PD-OBS and the benchmarks. Without such details, the reported Top-10 accuracy could reflect memorization of near-duplicate OBS images or modern-character labels rather than genuine decipherment. This is a concrete correctness risk that must be addressed with an explicit split-disjointness analysis.","section":"Abstract / PD-OBS dataset construction"},{"comment":"The term 'zero-shot decipherment' is not defined in the abstract or readable text. It is unclear what constitutes the unseen set (e.g., characters absent from PD-OBS, novel glyph variants, or previously undeciphered OBS), what the candidate pool is, and how the Top-10 ranking is constructed. Without a precise protocol and reported numbers (with baselines and variance), the central generalization claim is not testable. Please specify the train/test split, candidate generation, and evaluation metric.","section":"Zero-shot protocol"},{"comment":"The paper claims that the model delivers 'logical analysis processes' with potential archaeological value, but no examples, human evaluation, or quantitative assessment are provided in the readable material. It is important to show that the analysis outputs are grounded in the input glyph and not post-hoc rationalizations of a retrieval ranking. At minimum, include qualitative case studies with expert commentary or a structured evaluation of analysis quality, and clarify how such outputs can be validated by archaeologists.","section":"Logical analysis claims"}],"minor_comments":[{"comment":"The manuscript text is heavily corrupted; even section headings and figure captions are unreadable. The authors should re-upload a correct PDF and verify rendering across viewers.","section":"Full text"},{"comment":"The abstract would benefit from reporting a concrete Top-10 accuracy number and specifying which public benchmarks are used, so that readers can gauge the improvement without reading the full paper.","section":"Abstract"},{"comment":"The provided GitHub repository URL should be checked for availability and documentation; the manuscript should state the license, annotation procedure, and inter-annotator agreement for the pictographic analysis texts.","section":"Dataset / GitHub link"}],"recommendation":"major_revision","confidential_remarks":"The main concern is that the submitted manuscript is effectively unreadable beyond the abstract, so no technical evaluation is possible. The dataset-overlap issue raised by the skeptic is legitimate and must be resolved with a clear decontamination analysis, especially because the PD-OBS dataset is built from the same domain as the benchmarks. I recommend major revision rather than rejection because the proposed direction is plausible and the missing evidence could be supplied in a revised version; however, if the revision does not include readable methods and quantitative experimental details with split-disjointness verification, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — two things to know. This paper introduces a new annotated dataset, PD-OBS, with 47,157 Chinese characters paired with OBS images and pictographic analysis text, and it proposes an LVLM-based method that progressively trains from radical recognition to pictographic reasoning, with a Radical-Pictographic Dual Matching mechanism. That combination is genuinely new in the OBS-decipherment work I've seen, and the zero-shot framing is the right target.\n\nWhat the paper does well on the evidence we have (the abstract; the supplied full text is garbled): the dataset is a concrete, reusable resource if released properly; the training strategy is coherent and interpretable by design; evaluating on public benchmarks and claiming SOTA Top-10 accuracy is the standard bar; and the claim that the model outputs logical analysis sequences is useful for archaeological triage.\n\nSoft spots, in proportion. The abstract reports no numbers, no error bars, and no description of how PD-OBS is split from the public benchmarks. That matters because the central empirical claim is zero-shot decipherment. If the training set overlaps the test images or modern-character labels, top-10 accuracy can be memorization, not decipherment. The stress-test note makes exactly this point, and it lands: no decontamination procedure is described. The 'logical analysis' could then be post-hoc rationalization of a retrieval. These are correctness risks, not mere quibbles, though they may be addressed in the full paper. A smaller gap: the interpretability claim has no evaluation protocol—no human study or archaeologist assessment—so it reads as anecdotal.\n\nWho this is for: computational paleographers and digital-humanities researchers building triage tools for undeciphered scripts. I'd send it to peer review rather than desk-reject: the dataset alone justifies a referee round, and the method is plausible enough to test. The referee should require (1) a statement of disjointness between PD-OBS and benchmark splits, (2) per-split error bars, and (3) release of the split-generation code. If those hold, the SOTA claim deserves attention; if not, this is a dataset paper with an overreaching method.\n\nNet: worth a serious referee, with the decontamination question as the load-bearing review item.","headline":"New annotated OBS dataset and a coherent LVLM-based decipherment method, but SOTA and zero-shot claims hinge on train/test disjointness that the abstract does not establish.","tokens_in":8544,"tokens_out":2625,"would_cite":true,"duration_ms":29785,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Explicit radical and pictographic reasoning lets a vision-language model surpass prior methods at deciphering Oracle Bone Script, including zero-shot settings.","keywords":["Oracle Bone Script","vision-language model","pictographic analysis","radical analysis","zero-shot decipherment","ancient script recognition","interpretability","digital humanities"],"falsifier":"Take every OBS character whose accepted decipherment was published after the PD-OBS dataset was compiled, so the answer cannot be memorized. Ask the model for its top-10 candidate readings and compare with the published consensus; if the hit rate is no better than a random sample of the candidate vocabulary, the zero-shot claim is refuted.","tokens_in":7847,"feed_emoji":"📜","tokens_out":8029,"duration_ms":94147,"temperature":0.7,"pith_summary":"Oracle Bone Script is the oldest mature Chinese writing system, but most of its glyphs resist automated decipherment because they are rare, abstract, and pictorially varied. This paper argues that the missing ingredient is explicit reasoning about a glyph's component parts (radicals) and its pictorial resemblance to what it means. The authors train a large vision-language model on a new dataset of 47,157 characters, each annotated with an OBS image, the modern Chinese character, and a pictographic analysis text, using a progressive curriculum from radical recognition to full glyph-to-meaning reasoning. A Radical-Pictographic Dual Matching mechanism then aligns the model's own radical and pictographic evidence to rank candidate modern characters, and the paper reports state-of-the-art Top-10 accuracy with superior zero-shot performance. If the results hold, the model's written-out analyses could give archaeologists candidate decipherments for currently undeciphered oracle bone characters.","feed_headline":"Vision-language model deciphers oracle bone glyphs","feed_subtitle":"Radical and pictographic analysis beats prior top-10 accuracy and offers explained leads for undeciphered characters.","key_machinery":"The central mechanism is the Radical-Pictographic Dual Matching, in which the model first identifies the radical components of a glyph, then generates a pictographic semantic analysis, and finally matches both against candidate Chinese characters. The progressive training curriculum—radical recognition, then radical analysis, then pictographic analysis, then mutual analysis—is what leads the model to reason from glyph shape to meaning rather than memorize whole-glyph labels.","core_discovery":"The paper's central claim is that OBS decipherment should be treated not as image classification but as interpretable visual-semantic reasoning. The proposed large vision-language model is trained in stages: radical recognition, radical analysis, pictographic analysis, and mutual analysis that combines both views. A Radical-Pictographic Dual Matching mechanism uses the model's generated analysis to connect visual components with semantic units, which the authors credit for the gains in zero-shot generalization. On public OBS benchmarks, the method reaches state-of-the-art Top-10 accuracy, and it produces a written analysis chain for an unseen glyph. The authors argue this makes the model use","pith_inferences":["An untested consequence of the progressive curriculum is that the same radical-first ordering should help other ancient logographic scripts—such as Egyptian hieroglyphs or Maya writing—that also mix pictorial and phonetic components; building radical annotations for those scripts would test this.","Because dual matching produces both a radical analysis and a pictographic analysis, the agreement between the two can serve as a data-quality signal: disagreement marks glyphs that should be sent for expert review before any decipherment is trusted.","If radical-pictographic reasoning is genuinely more sample-efficient than whole-glyph classification, shrinking the training set should widen the performance gap between this method and image-classification baselines; measuring that gap across training-set sizes would test the mechanism's core premise."],"forward_implications":["On public benchmarks, the method's Top-10 accuracy surpasses previous deep-learning baselines, so OBS recognition no longer has to be treated as pure closed-set classification.","In zero-shot settings, dual matching of radical and pictographic evidence lets the model rank plausible modern Chinese characters for glyphs it has not seen.","The model's written analysis chains give epigraphers a rationale, not just a candidate, so an undeciphered glyph can be examined as a proposed radical-pictograph compound.","The released PD-OBS dataset gives the research community 47,157 annotated OBS-image-to-Chinese-character pairs with pictographic analysis texts, supporting further work on ancient script understanding."],"supporting_citations":[],"fun_headline_variants":["Explainable AI deciphers oracle bone script via radical and pictographic analysis","Zero-shot oracle bone decipherment via radical-pictographic dual matching","LVLM bridges glyphs and meaning for interpretable oracle bone decipherment","Vision-language model decodes undeciphered oracle bones with logical reasoning"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the 47,157-character PD-OBS dataset is accurately labeled and comprehensive enough that a model trained on it generalizes to undeciphered oracle bone glyphs.","fun_headline_variants_meta":{"raw":{"variants":["Explainable AI deciphers oracle bone script via radical and pictographic analysis","Zero-shot oracle bone decipherment via radical-pictographic dual matching","LVLM bridges glyphs and meaning for interpretable oracle bone decipherment","Vision-language model decodes undeciphered oracle bones with logical reasoning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000357,"raw_usage":{"total_tokens":1807,"prompt_tokens":810,"completion_tokens":997,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":919}},"tokens_in":554,"tokens_out":997,"duration_ms":9965,"temperature":1.0,"reasoning_tokens":919,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:37:54.179270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take every OBS character whose accepted decipherment was published after the PD-OBS dataset was compiled, so the answer cannot be memorized. Ask the model for its top-10 candidate readings and compare with the published consensus; if the hit rate is no better than a random sample of the candidate vocabulary, the zero-shot claim is refuted.","supporting_citations":[],"review_version":1}