{"id":"9ad0ebce-f8a1-4f64-9755-9a2722322d2a","arxiv_id":"2506.03798","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CoLa learns compositional latent components of Chinese characters via slot attention and matches them to printed templates, achieving strong zero-shot Chinese character recognition without human-defined decomposition.","lead":"The paper introduces CoLa, a model that learns to break Chinese characters into reusable visual components without human-made radical or stroke labels, then matches those components to recognize characters it never saw during training. On zero-shot character recognition benchmarks it reports large accuracy gains over prior methods, and it can decompose oracle bone characters despite never being trained on them.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains are not shown to come from compositional latent components: no K=1 or teacher-feature control is run, so the central claim is underdetermined.","rationale":"I read the central claim as: CoLa's learned compositional latent components, rather than the surrounding machinery, are what enable the large zero-shot improvements. For that claim to hold, the component factorization must be doing causal work in the recognition pipeline. The paper's own evidence supports internal consistency: the fixed-order experiment in D.3 shows that component alignment can be controlled, and Tables 3-5 report small standard deviations over trials. Those are real supporting observations. However, no experiment isolates the slot-attention bottleneck from the frozen DINOv2 teacher. A simple non-compositional baseline using the same teacher features and template matching would settle whether the decomposition is essential or merely a reconstruction regularizer. The reader's concern about component-order alignment is legitimate, but it is secondary: even perfect alignment would not establish the central claim if a global representation performs equally. The missing ablation is therefore the most load-bearing gap. I set aside the citation placeholders and missing code release; they are addressable but do not touch the mechanism. I would keep the paper's CONDITIONAL status: the method is promising, but the compositional claim needs a decisive control before acceptance.","tokens_in":16364,"tokens_out":6626,"duration_ms":69251,"concrete_test":"On the HWDB 500-class character zero-shot split, train two controls with the same teacher encoder, templates, and loss hyperparameters: (1) CoLa with K=1 (a single global slot, so Eq. 7 becomes ordinary prototype matching on one vector), and (2) prototype matching directly on the teacher feature F (e.g., average-pooled F, classified by the same template-distance rule). If either control reaches within a few points of the reported 68.59%, the slot-attention decomposition is not load-bearing; if both drop substantially, the compositional mechanism is confirmed as the source of the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CoLa's recognition rule (Eq. 7) is a nearest-prototype classifier on the latent representation S, and Eq. 10 trains S by reconstructing teacher features F produced by a frozen DINOv2-plus-CNN encoder. DINOv2 is a strong general-purpose visual encoder; if S simply encodes F with little information loss, the large zero-shot gains in Table 1 could reflect the teacher encoder and template matching rather than the slot-attention factorization. The paper reports no control that removes or collapses the compositional mechanism: no K=1 variant, no direct matching on F or on pooled features, and no ablation of the teacher encoder. The component-order experiment in Appendix D.3 addresses alignment but starts from the assumption that the K×D slot representation is the active ingredient; it does not test whether compositionality is necessary for the accuracy. Thus the central claim that compositional latent components drive the improvement is not yet supported by the evidence as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoLa, a slot-attention-based deep latent variable model that decomposes Chinese character images into K latent components without radical or stroke supervision. The components are decoded to reconstruct teacher features (a frozen DINOv2 encoder plus a trained CNN), and classification is performed by comparing the latent components of the input against per-character template prototypes through a mixture-of-Gaussians rule (Eq. 7). Experiments report large zero-shot gains over prior radical-, stroke-, and CLIP-based methods on handwritten (HWDB), printed-font, and historical-document benchmarks, along with qualitative component visualizations and cross-dataset retrieval experiments on oracle bone characters.","tokens_in":16546,"tokens_out":7507,"duration_ms":71110,"significance":"If the reported gains are robust and genuinely attributable to the compositional latent-component mechanism, CoLa would be a significant step toward practical zero-shot Chinese character recognition that avoids human-defined decomposition schemes. The paper has notable strengths: it reports standard deviations over three trials for its own method (Appendix D.2), makes the template-matching inference rule explicit, and includes a component-order experiment (Appendix D.3) that addresses slot-alignment in a limited way. The central empirical claim is, however, currently underdetermined by the absence of controls that isolate the compositional mechanism, and the printed-font evaluation has a potential template/test overlap. Most of these concerns are addressable with additional experiments, so the paper merits revision rather than rejection.","major_comments":[{"comment":"The central claim that the reported gains are due to compositional latent components is underdetermined. CoLa's recognition rule (Eq. 7) is a nearest-prototype classifier in the slot-latent space, and the only learning signal for S is reconstruction of the teacher features F (Eq. 10), where F comes from a frozen DINOv2 plus a CNN trained on the training classes. The paper provides no K=1 ablation, no control that matches F (or averaged F) directly to templates without slot decomposition, and no ablation that replaces the DINOv2 teacher with a randomly initialized or weaker encoder. Without these controls, the large improvements in Table 1 are equally consistent with a strong-teacher-plus-template-matching explanation in which the slot-attention factorization is not the active ingredient. Please add at least a K=1 variant and a direct teacher-feature matching baseline.","section":"Sec. 4.2, Table 1; Eq. (10)"},{"comment":"The Printed benchmark may be contaminated by template-font overlap. The Printed dataset is generated from 105 font files, and the template set is generated from 'commonly used printed fonts' (Appendix B). The paper does not state whether the fonts used to render the test images are disjoint from the fonts used to generate templates. If they are not disjoint, the Printed rows in Table 1 (e.g., 78.10 at 500 classes and 92.70 at 2755) largely reflect near-duplicate template retrieval rather than zero-shot generalization. Please report the font split, or rerun the Printed experiments with template fonts excluded from the set of fonts used to generate test images.","section":"Appendix B; Sec. 4 dataset construction"},{"comment":"The class-prediction term in Eq. 7 compares component k of the input with component k of each template, which requires that the fixed global order epsilon aligns slots semantically across inputs, templates, styles, and the printed-to-handwritten domain gap. The only evidence provided is a small set of visualizations in Appendix D.3. Please add a quantitative assessment of slot correspondence: for example, measure whether randomly permuting slot indices across styles degrades accuracy, or evaluate assignment consistency of the learned components across the N templates of the same character. Without such a check, the nearest-component comparison in Eq. 7 may be based on mismatched components.","section":"Sec. 3.2; Appendix D.3"},{"comment":"The paper acknowledges in Eq. 19 that the two KL regularizers are identically zero because the variational distribution and the generative process share the same encoder. This means the ELBO reduces to an expected reconstruction plus an expected classification term, with no posterior regularization; the 'deep latent variable model' framing is therefore not a standard variational autoencoder. The authors should either provide a principled reason for keeping the sampling step and the Gaussian likelihood, or reframe the method as a deterministic encoder-decoder with a teacher-feature reconstruction objective and a classification loss. This issue does not by itself invalidate the empirical results, but it affects the theoretical novelty claim.","section":"Appendix A.4, Eq. (19)"}],"minor_comments":[{"comment":"The phrase 'outperforms previous methods in both character the radical zero-shot CCR' is ungrammatical; it should be 'in both character and radical zero-shot CCR.'","section":"Abstract"},{"comment":"K is described as the 'maximum number of components in the images,' but in practice it is a fixed number (K=3 in the experiments); the wording should be 'the number of latent components.'","section":"Sec. 3.1"},{"comment":"The numerator in Eqs. (20) and (21) uses \\tilde{T}_{c,n}, which appears to be a typo for \\tilde{T}_{y,n} as in Eq. (18); please correct it.","section":"Appendix A.4, Eqs. (20)-(21)"},{"comment":"Tables 1 and 2 report only a single run for all baseline methods; adding standard deviations or stating the number of seeds for the key comparisons would make the comparison fairer, especially given the very large margins claimed.","section":"Tables 1 and 2"},{"comment":"The historical-document dataset is not released and its size and class distribution are not given; please provide dataset statistics and a data-release plan to support reproducibility.","section":"Appendix B; Sec. 4.2"},{"comment":"The caption contains a typo: 'Fix initialization' should be 'Fixed initialization.'","section":"Fig. 8 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong fit for the journal's scope, but the central attribution claim — that compositional latent components, rather than the strong teacher encoder or template-font duplication, drive the reported zero-shot gains — is not yet established. The required controls (K=1 ablation, direct teacher-feature matching, font-overlap analysis, and a quantitative slot-alignment check) are straightforward and within the manuscript's scope, so I recommend major revision rather than rejection. I would also gently encourage the authors to make the historical-document data and code available if possible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper reports very large zero-shot Chinese character recognition gains using slot attention to decompose characters into unsupervised latent components, then matching templates in that latent space. If the numbers hold, it's a real step forward. But the paper doesn't yet show that the compositional factorization is what's doing the work.\n\nWhat's genuinely new: applying object-centric slot attention to Chinese characters, with no radical or stroke supervision, and using the resulting components for zero-shot recognition via template matching. The reported gains are consistent across HWDB, printed fonts, and a historical document set, with e.g. 68.59% vs 21.79% for CCR-CLIP at 500 HWDB classes. They also provide std devs over three trials for CoLa's own results, which is more than many papers do. The visualizations of learned components are suggestive, though not quantified. The OBC retrieval demo is a nice qualitative touch.\n\nThe soft spots are real, though. The stress-test note is on point: the model's recognition rule is a nearest-prototype classifier on the latent S, and S is trained to reconstruct features from a frozen DINOv2 teacher. DINOv2 is a strong general visual encoder. Without a K=1 control, or direct matching on the teacher features themselves, we don't know whether slot attention is buying the accuracy or whether the teacher plus template matching alone would do nearly as well. This is the main missing experiment, and it's a straightforward one. The reader's other concerns also hold: no error bars on baseline tables, no released code/data, the historical dataset is under-specified, and there are unresolved citation placeholders in the appendix. The KL terms vanish by construction because q and p share parameters, so the variational framing is somewhat decorative.\n\nThe core empirical claim is probably right in the narrow sense that CoLa with this architecture does well on these benchmarks. But the broader claim that unsupervised compositional decomposition is the cause is not yet supported. The paper deserves a serious referee, but it needs a major revision: run K=1 and teacher-feature matching baselines, add error bars to all tables, release code and data, and fix the appendix.\n\nWho's this for? People working on zero-shot recognition, compositional representation learning, and document analysis. I'd bring it to a reading group and I'd cite it if I were working on zero-shot CCR, but I would not rely on the mechanistic claim until the ablations are in.\n\nMy recommendation: send it to peer review, but make the missing controls a condition for acceptance. It's a solid contribution with an addressable flaw.","headline":"Large zero-shot gains from slot-attention decomposition, but the key ablation (is it the compositionality or just the teacher features?) is missing.","tokens_in":17050,"tokens_out":2331,"would_cite":true,"duration_ms":21365,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CoLa learns Compositional Latent components of Chinese characters without human-defined decomposition and uses them for zero-shot Chinese character recognition.","keywords":["Chinese character recognition","zero-shot recognition","compositional latent components","slot attention","radical zero-shot","template matching","variational inference","oracle bone characters"],"falsifier":"Run the 500-class HWDB character zero-shot experiment with a random decomposition order sampled independently for each image at test time; if accuracy stays near the reported 68.59%, the fixed order is not load-bearing, while a large drop would confirm that slot alignment is the mechanism. A complementary check is to permute the three template components at test time and see whether recognition and retrieval accuracy fall.","tokens_in":16140,"feed_emoji":"🀄","tokens_out":7708,"duration_ms":75803,"temperature":0.7,"pith_summary":"This paper introduces CoLa, a deep latent variable model that learns to split Chinese character images into a small number of latent components without using radical or stroke labels. The central claim is that these learned components act as a shared alphabet: for a previously unseen character, CoLa encodes both the input image and printed-font templates into the same latent-component space, and the class is chosen by matching the input's components against each class's template components. The authors report that this beats earlier zero-shot Chinese character recognition methods, with 68.59% versus 21.79% for CCR-CLIP on the 500-class HWDB handwritten setting and an average accuracy gain of about 60% in radical zero-shot settings. If correct, the model shows that decomposition-and-recombination for recognizing new characters can be learned directly from data and can transfer across writing styles, historical documents, and even oracle bone scripts.","feed_headline":"Zero-shot Chinese OCR jumps to 68.6% with learned components","feed_subtitle":"CoLa learns compositional parts without radical labels and transfers these parts to oracle bone scripts.","key_machinery":"The machinery is slot attention with a shared component order, plus template matching. Slot attention is an iterative attention mechanism that groups image features into $K$ slots; here $K=3$, and each slot becomes a compositional latent component. A frozen teacher encoder's visual features are the reconstruction target, so the slots must encode structure rather than low-level pixels. Templates are encoded with the same slot-attention encoder, and class prediction treats each character class as a mixture of Gaussians whose mean is the average of its template components, which turns zero-shot recognition into a latent-space comparison. The fixed observed variable $\\epsilon$, which initializes the slots in the same order for all images, is what makes component $k$ comparable across styles and across the printed-to-handwritten domain gap.","core_discovery":"CoLa's central discovery is that slot attention, trained by reconstructing frozen teacher features and by predicting classes through latent template matching, discovers compositional components of Chinese characters that are consistent enough across images to support zero-shot recognition. The components are not human-defined radicals, yet they are repeatable and structurally meaningful: with three slots, each component attends to a distinct region of the character. A fixed global component order $\\epsilon$ aligns slot $k$ across every input and template image, so recognition can compare component $k$ of an unseen handwritten character with component $k$ of printed templates. The class decision is a nearest-class rule in latent space, where each class is a Gaussian mixture component centered at the mean of its $N$ template component vectors. On the paper's experiments, this unsupervised decomposition outperforms prior radical- and stroke-based methods in both character and radical zero-shot settings, and the same model trained on historical documents decomposes oracle bone characters without retraining.","pith_inferences":["Beyond the paper, a stronger test would be cross-writer or cross-era style shift, where the printed-to-handwritten gap is replaced by an unseen handwriting style gap; the fixed-order alignment has not been tested there.","Because the decoder turns each component into a mask and feature map, the learned components could be reused for generation or style transfer, not just recognition.","If the fixed-order assumption is what matters, then slot permutation at test time should destroy accuracy; that experiment would separate semantic components from positional slots as the actual carrier of generalization."],"forward_implications":["Human-defined radical or stroke annotations are not required for zero-shot Chinese character recognition; the components are discovered from data.","Recognition is a fast template-matching operation in latent space, avoiding the slow autoregressive decoding of radical or stroke sequences (the paper reports 9 ms versus 1666 ms per batch for DenseRAN).","The same decomposition transfers to unseen scripts: a model trained only on historical documents parses oracle bone, Japanese, and Korean characters and retrieves visually similar samples.","Rare or unseen radicals do not break recognition because the model's components are tied to visual structure, not to a predefined radical inventory."],"supporting_citations":[{"why":"Supplies the stochastic gradient variational Bayes estimator used to derive the ELBO training objective.","marker":"[37, 38]"},{"why":"Supplies the slot attention module that CoLa uses to decompose character images into latent components.","marker":"[32]"},{"why":"CCR-CLIP is the matching-based zero-shot CCR baseline that the paper compares against and reports outperforming.","marker":"[17]"},{"why":"The stroke-level zero-shot CCR method whose dataset construction and zero-shot splits CoLa follows.","marker":"[15]"},{"why":"The frozen DINOv2 teacher encoder whose features CoLa reconstructs to learn structural components.","marker":"[39]"},{"why":"The Spatial Broadcast Decoder converts each latent component into a mask and feature map for reconstruction.","marker":"[40]"}],"fun_headline_variants":["CoLa learns Chinese character parts without radical labels","CoLa's latent components enable zero-shot Chinese OCR","Unsupervised decomposition beats radical-based zero-shot Chinese OCR","CoLa learns to decompose characters, generalizes to oracle bones","CoLa discovers compositional Chinese character components unsupervised"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fixed global component order keeps slot $k$ aligned across every input and template image, so comparing component $k$ between a handwritten input and a printed template is semantically meaningful; if that alignment fails, the class prediction compares mismatched components and accuracy collapses.","fun_headline_variants_meta":{"raw":{"variants":["CoLa learns Chinese character parts without radical labels","CoLa's latent components enable zero-shot Chinese OCR","Unsupervised decomposition beats radical-based zero-shot Chinese OCR","CoLa learns to decompose characters, generalizes to oracle bones","CoLa discovers compositional Chinese character components unsupervised"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001136,"raw_usage":{"total_tokens":4726,"prompt_tokens":964,"completion_tokens":3762,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":3686}},"tokens_in":580,"tokens_out":3762,"duration_ms":27532,"temperature":1.0,"reasoning_tokens":3686,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:54:52.732364+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the 500-class HWDB character zero-shot experiment with a random decomposition order sampled independently for each image at test time; if accuracy stays near the reported 68.59%, the fixed order is not load-bearing, while a large drop would confirm that slot alignment is the mechanism. A complementary check is to permute the three template components at test time and see whether recognition and retrieval accuracy fall.","supporting_citations":[{"cited_title":"Weissenborn, T","cited_arxiv_id":null,"evidence_quote":"Supplies the slot attention module that CoLa uses to decompose character images into latent components."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CCR-CLIP is the matching-based zero-shot CCR baseline that the paper compares against and reports outperforming."},{"cited_title":"Zero-Shot Chinese Character Recognition with Stroke-Level Decomposition","cited_arxiv_id":"2106.11613","evidence_quote":"The stroke-level zero-shot CCR method whose dataset construction and zero-shot splits CoLa follows."}],"review_version":1}