{"id":"4c0d0a17-7028-453a-ac2a-da3c275e1233","arxiv_id":"2508.01153","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Feeding ground-truth label embeddings into an STR decoder and progressively masking them based on training loss improves accuracy on several benchmarks while leaving inference unchanged.","lead":"TEACH is a training trick for scene text recognition that feeds the correct text into the model early in training, then gradually hides it so the model learns to read from the image alone. It claims to boost accuracy on standard benchmarks without extra cost at test time, which would make it a cheap upgrade for OCR systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 lists the ViTSTR-S baseline twice with different accuracies and Table 3 reuses Table 2's caption, so the claimed TEACH gains over baselines are not currently verifiable.","rationale":"The reader's weakest assumption was that the benchmark gains are honest out-of-sample estimates, specifically because alpha and beta are tuned on the evaluation benchmarks. That is a valid concern. However, the more immediate problem is that the paper's own tables contradict the reported gains: the ViTSTR-S baseline is shown twice with different numbers, and the ViTSTR+TEACH row is not an improvement over either baseline on several columns. The text in Section 4.3 also claims gains on IC15 and SVTP that the table does not support. This is not about generalization or hyperparameter tuning; it is about the in-sample correctness of the central comparison. I therefore recommend moving the verdict from CONDITIONAL to REJECT for the current version: the empirical support is internally inconsistent and no code is provided to resolve the ambiguity. This is not a judgment that the method is wrong, only that the manuscript as submitted does not provide verifiable evidence for its headline claim.","tokens_in":14000,"tokens_out":10059,"duration_ms":129319,"concrete_test":"Re-run the actual ViTSTR-S and ViTSTR+TEACH experiments on the Real(3.3M) training set with a fixed seed, report the baseline and TEACH rows, and replace both conflicting ViTSTR-S rows in Table 2 with this single measured baseline; then check whether the claimed IC15 (+0.6) and SVTP (+0.8) gains are reproduced. If not, the central empirical claim is falsified as stated.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that TEACH consistently improves the accuracy of ViTSTR and PARSeq. The evidence for this is the delta between baseline rows and TEACH rows in Tables 1-3, but those deltas cannot be trusted as printed. In Table 2, the ViTSTR-S baseline under Real(3.3M) appears twice with different numbers, and the ViTSTR+TEACH row is below both baseline rows on IIIT5K, below the first baseline on IC13/IC15/SVTP, and below the second baseline on IC15/SVTP/CUTE. The text in Section 4.3 nonetheless claims ViTSTR+TEACH 'improves on IC15 and SVTP by 0.6% and 0.8%', which is not consistent with either baseline row. Table 3 also copies Table 2's caption while actually evaluating ArT/COCO-Text/Uber-Text. With no released code, the reader cannot tell which baseline configuration was used. Since the central claim is an empirical improvement claim, an internally inconsistent comparison table is a load-bearing defect: the headline result is not established even in-sample.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TEACH, a training-only curriculum method for scene text recognition that injects ground-truth label embeddings into the decoder input and progressively masks them using a loss-dependent schedule. The authors claim consistent accuracy improvements over ViTSTR and PARSeq baselines without inference overhead, across synthetic and real training data and several benchmarks. The core idea is simple and architecture-agnostic, but the empirical evidence as presented contains internal contradictions (Table 1 SVT drop, duplicate ViTSTR-S rows in Table 2, reused caption in Table 3), an under-specified masking mechanism (Eqs. 4-5), and hyperparameter selection directly on the evaluation benchmarks (Table 6).","tokens_in":14265,"tokens_out":8898,"duration_ms":93991,"significance":"The method has practical appeal: it requires no external language models, no inference-time changes, and could be dropped into existing encoder-decoder STR models. The paper evaluates on a broad set of benchmarks and two base architectures, and the qualitative examples suggest the mechanism can help on distorted or low-contrast text. However, the current manuscript does not establish the central empirical claim. The test-set tuning of alpha and beta means the reported gains may not generalize, and the table inconsistencies prevent verification even in-sample. With corrected experiments and clean reporting, TEACH could be a useful contribution as a lightweight training augmentation; as submitted, its significance is not yet demonstrated.","major_comments":[{"comment":"The definition and interpretation of the masking rate r are internally inconsistent. The text states that 'A high loss results in low masking (more label input), while a low loss increases masking', but Eq. (5) defines r as increasing in Loss. Moreover, the final stage says that when the loss falls below beta the label input is fully removed (r = 0 under Eq. (5)), which is the opposite of what a 'masking rate' of 0 should mean. In addition, Eq. (4) uses a binary mask M in {0,1}^{Lt} but r is a scalar, and the paper never specifies how the scalar r produces M. Because the loss-aware curriculum is the central mechanism, this ambiguity blocks implementation and evaluation of the method.","section":"Section 3.2, Eq. (5)"},{"comment":"The real-data comparison table contains two identical 'ViTSTR-S' baseline rows with different numbers (97.9/96/97.8/89/87.5/91.5/96.2 vs 98.0/95.0/97.2/88.3/87.4/91.8/97.6) and no distinguishing label, and the header includes two 'IC15' columns without indicating which variant (1811 vs 2077) each refers to. Under these rows, ViTSTR+TEACH is below at least one baseline on IIIT5K, IC13, both IC15 columns, and CUTE. The claim in Section 4.3 that 'ViTSTR+TEACH improves on IC15 and SVTP by 0.6% and 0.8%' cannot be reproduced from either baseline row. This makes the central claim of consistent improvement unverifiable for the real-data setting.","section":"Table 2"},{"comment":"The synthetic-data results contradict the headline claim of consistent improvement: ViTSTR+TEACH scores 91.3 on SVT against 93.8 for ViTSTR-S, a drop of 2.5 percentage points, while the caption and Section 4.3 state that TEACH 'improves both ViTSTR and PARSeq baselines across all benchmarks.' This is not a minor fluctuation; it directly refutes the word 'consistent' in the central claim.","section":"Table 1"},{"comment":"The hyperparameters alpha and beta are selected by grid search over accuracy averaged across the six evaluation benchmarks (Table 6), and the selected configuration (alpha=2, beta=0.1) is used for the headline numbers in Tables 1-3. Since the same benchmarks are used for both selection and final reporting, the reported improvements are not out-of-sample estimates. Additionally, Section 4.2 states that beta is set 'slightly below the stabilized loss' of a pretrained baseline, which is a different criterion from the grid search; the manuscript does not reconcile these procedures. The claimed general improvement therefore needs to be re-established with a held-out validation set or with a clearly specified selection protocol that does not use the test benchmarks.","section":"Section 4.4, Table 6"},{"comment":"Table 3's caption is copied from Table 2, mentioning 'six STR benchmarks', 'DiG-ViT, ABINet, and MAERec', and 'five out of six benchmarks', but the table actually reports four columns (ArT, COCO-Text, Uber-Text, Total) and contains none of those models. Since this table is the evidence for the 'challenging large-scale benchmarks' claim, the mismatch prevents the reader from knowing what was evaluated and how the numbers were computed.","section":"Table 3"}],"minor_comments":[{"comment":"The definitions of 'ALL benchmark set' and 'Clean subset' are not used consistently with the tables; please label the duplicate IC13/IC15 columns explicitly (e.g., IC15-1811 and IC15-2077) in every table.","section":"Section 4.1"},{"comment":"Section 3.4 says CRNN is excluded from the main analysis because 'preliminary experiments found it difficult to propagate useful gradients', yet Table 7 reports CRNN+TEACH gains; please clarify why CRNN results are not in the main tables or remove the preliminary-experiment statement.","section":"Section 3.4 and Table 7"},{"comment":"Please provide the exact mapping from the scalar masking rate r in Eq. (5) to the binary mask M in Eq. (4), including how the sequence-length dimension is handled (e.g., Bernoulli sampling, token-level thresholding, or per-position masking).","section":"Section 3.2"},{"comment":"The word 'Predication' appears in the figure text; it should be 'Prediction'.","section":"Figure 3"},{"comment":"In the qualitative comparison, the notation alternates between 'ViTSTR-TEACH' and 'ViTSTR+TEACH' and between 'PARSeq-TEACH' and 'PARSeq+TEACH'; please standardize and clarify which variant (e.g., PARSeqA vs PARSeqN) is used.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The experimental section appears not to have been checked as a whole: duplicate baseline rows, a reused caption, contradictory text around Eq. (5), and benchmark-level hyperparameter selection together make the headline results unverifiable. If the authors can rerun the experiments with validation-based hyperparameter selection and correct the internal inconsistencies, I would be willing to look at a revised version. I would also ask the editor to verify that the reported baseline numbers (especially the two ViTSTR-S rows in Table 2) match the original sources."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: TEACH is a plausible training-time curriculum for STR—inject ground-truth label embeddings into the decoder, then mask them out as training loss drops. The idea is easy to grasp, model-agnostic, and adds zero inference cost. If it works, it’s a cheap trick that could help several STR architectures. But the evidence in the preprint does not currently establish that it works, and the internal inconsistencies are serious enough that I would not trust the headline numbers without a clean re-run.\n\nWhat’s genuinely new is the specific design: a loss-aware masking rule, r = α·(Loss − β) clamped to [0,1], that schedules the removal of label tokens. That is a fresh twist on scheduled sampling and curriculum learning, but the paper never cites scheduled sampling or compares against it, which is a real omission. The paper deserves credit for testing on both synthetic and real data, for ablating the masking schedule, and for attempting CRNN integration.\n\nThe soft spots are load-bearing. Table 1 shows ViTSTR+TEACH dropping on SVT (91.3 vs 93.8) yet the text claims consistent improvements. Table 2 lists the ViTSTR-S baseline twice with different numbers, and the ViTSTR+TEACH row is below both baselines on IIIT5K, IC13, IC15, SVTP, and CUTE in various columns; the text’s claim of an IC15 improvement is contradicted by every version of the table. Table 3 reuses Table 2’s caption, a minor but telltale copy-paste error. Second, α and β are chosen by grid search over the same six evaluation benchmarks (Table 6), so the reported gains are not truly out-of-sample. Third, the masking mechanism is underspecified: how does the scalar r produce the binary mask M over L_t tokens? That needs spelling out for reproducibility. No code is released.\n\nThe core idea is not circular—the model still has to read visually because the labels are removed—and I would not dismiss it. But the current manuscript does not reliably demonstrate it. A serious referee should see it, because the idea is worth a proper test. My own recommendation: fix or regenerate the tables, hold out the hyperparameter selection, and release code. Until then, treat the method as unverified.","headline":"A plausible training-time curriculum idea for STR whose reported evidence is undercut by inconsistent tables and benchmark-tuned hyperparameters; the method may still work, but the paper needs major revision.","tokens_in":14736,"tokens_out":3937,"would_cite":false,"duration_ms":42360,"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":"This paper claims that injecting ground-truth label embeddings into a scene text recognizer's decoder during training, then masking them with a loss-aware rule, improves accuracy on six benchmarks with no inference-time cost.","keywords":["Scene Text Recognition","Curriculum Learning","Training-time label injection","Loss-aware masking","Encoder-decoder models","Benchmark evaluation","Model-agnostic training","Vision-only inference"],"falsifier":"Fix $\\alpha$ and $\\beta$ on a validation split drawn only from training data, then evaluate on the six benchmarks; if the TEACH models no longer beat their baselines by a margin beyond repeated-run noise, the out-of-sample claim fails. A second check: train with randomly permuted ground-truth labels under the same schedule; if accuracy stays as high as with correct labels, the effect is not semantic guidance.","tokens_in":13777,"feed_emoji":"🔤","tokens_out":10254,"duration_ms":108039,"temperature":0.7,"pith_summary":"TEACH is a training-only scheme for scene text recognition that gives the decoder the correct word during training and then gradually takes it away. The authors claim that starting from full label guidance and masking the labels according to the recent training loss forces the model to learn stronger visual features, so that at test time, when no label is available, it recognizes text more accurately. Across six standard benchmarks, models equipped with TEACH improve on their base counterparts under both synthetic and real training data, with the largest gains on distorted, occluded, and low-quality text. The method is positioned as a lightweight alternative to language-model-based recognizers because it needs no external pretraining and adds no inference-time computation.","feed_headline":"Label hints during training boost scene text accuracy","feed_subtitle":"A training-only curriculum improves recognition on six benchmarks while leaving inference unchanged.","key_machinery":"The central object is the loss-aware masking rule $r = \\max(0, \\min(1, \\alpha\\cdot(\\mathrm{Loss}-\\beta)))$, which sets the fraction of label tokens replaced by a padding token in each training step. High loss means little masking, so label guidance stays on; low loss means heavy masking, and once loss falls below $\\beta$ the label stream is fully padded out. This rule is what turns a fixed label-injection trick into a curriculum, letting the decoder shift from reconstructing the label embedding to reading visual features, while keeping the ability to relax masking if loss spikes on hard samples. The companion mechanism is the concatenation of text embeddings with visual tokens in the decoder input, so the model must decide which stream to trust at each step.","core_discovery":"On its own terms, the central claim is that a curriculum of text hints can bootstrap visual recognition: during training the ground-truth label sequence is embedded and concatenated with the visual tokens in the decoder input, and this label stream is masked according to $r = \\max(0, \\min(1, \\alpha\\cdot(\\mathrm{Loss}-\\beta)))$, where Loss is the previous batch's loss. High loss keeps the labels visible; low loss hides them; once loss falls below $\\beta$ the labels vanish entirely. The authors report that this schedule consistently improves word accuracy over the ViTSTR and PARSeq baselines on IIIT5K, SVT, IC13, IC15, SVTP, and CUTE, with the largest improvements on blurry, distorted, and occluded samples, and that PARSeq+TEACH reaches state-of-the-art results on five of six synthetic-trained benchmarks and on IC15, SVTP, and CUTE with real-data training. They further report gains on ArT, COCO-Text, and Uber-Text and on the recurrent CRNN architecture, supporting the claim that the strategy is model-agnostic; at inference the model takes only the image, so no extra computation or parameters are added.","pith_inferences":["If the evaluation honestly measures out-of-sample behavior, the method implies that other structured prediction tasks, such as handwritten text recognition, image captioning, or speech recognition, could use a similar loss-aware ground-truth hint curriculum that vanishes at test time.","The reported hyperparameter search is a caution flag: $\\alpha$ and $\\beta$ were chosen over the same six benchmark sets, with $\\beta$ tied to a baseline's stabilized loss, so the true out-of-sample gain is likely smaller than the headline tables suggest until the schedule is fixed on a held-out split.","A testable extension of the paper's own logic is to train with permuted labels under the same masking rule; if accuracy stays as high as with correct labels, the benefit comes from regularization or noise tolerance rather than semantic alignment.","Inspecting decoder attention on padding positions during late training could show whether the model truly reads vision-only or merely learns a constant bias from the padded label slots."],"forward_implications":["TEACH can be plugged into existing encoder-decoder STR models without changing their networks, adding parameters, or altering inference, so any future backbone can inherit the recipe directly.","Models trained with synthetic data get a consistent lift: the paper reports ViTSTR+TEACH gaining up to 1.7 points on IC13 and 1.6 points on CUTE, and PARSeq+TEACH reaching 97.5% on IIIT5K.","The same pattern holds under real-data training, where PARSeq+TEACH reaches 99.2% on IIIT5K and 97.6% on CUTE and surpasses strong baselines such as ABINet and MAERec on several benchmarks.","Because loss-aware masking beats linear decay in the ablation, the adaptive schedule, not mere label injection, is the load-bearing component."],"supporting_citations":[{"why":"CRNN baseline; used to test TEACH on a recurrent decoder and support the model-agnostic claim.","marker":"[1]"},{"why":"ViTSTR baseline; one of the two base architectures that TEACH is added to in the main experiments.","marker":"[4]"},{"why":"PARSeq baseline; the other base architecture and the one that produces the headline state-of-the-art numbers.","marker":"[5]"},{"why":"ABINet; a vision-language baseline that TEACH is compared against to show gains without external language pretraining.","marker":"[7]"},{"why":"DiG-ViT-B; a strong transformer baseline that PARSeq+TEACH surpasses on five synthetic-data benchmarks.","marker":"[24]"},{"why":"MAERec-B; a strong recent baseline that PARSeq+TEACH surpasses under real-data training.","marker":"[26]"},{"why":"MJSynth; one of the two synthetic training sets used in the main experiments.","marker":"[27]"},{"why":"SynthText; the other synthetic training set used alongside MJSynth.","marker":"[28]"}],"fun_headline_variants":["Training-time label hints sharpen scene text recognition","Curriculum of text hints lifts OCR accuracy","Hidden label curriculum trains better OCR models","Masked label hints guide scene text training","Label masking teaches models to read without hints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the benchmark numbers honestly measure what the method would do on new data, since the two masking hyperparameters were tuned with the evaluation benchmarks in view (the grid-search table and a baseline's stabilized loss), and if that tuning leaked into the headline results, the reported generalization gains could shrink on unseen test sets.","fun_headline_variants_meta":{"raw":{"variants":["Training-time label hints sharpen scene text recognition","Curriculum of text hints lifts OCR accuracy","Hidden label curriculum trains better OCR models","Masked label hints guide scene text training","Label masking teaches models to read without hints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000145,"raw_usage":{"total_tokens":1157,"prompt_tokens":903,"completion_tokens":254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":190}},"tokens_in":519,"tokens_out":254,"duration_ms":3908,"temperature":1.0,"reasoning_tokens":190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:48:56.682892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix $\\alpha$ and $\\beta$ on a validation split drawn only from training data, then evaluate on the six benchmarks; if the TEACH models no longer beat their baselines by a margin beyond repeated-run noise, the out-of-sample claim fails. A second check: train with randomly permuted ground-truth labels under the same schedule; if accuracy stays as high as with correct labels, the effect is not semantic guidance.","supporting_citations":[{"cited_title":"TEACH: Text Encoding as Curriculum Hints for Scene Text Recognition","cited_arxiv_id":"2508.01153","evidence_quote":"CRNN baseline; used to test TEACH on a recurrent decoder and support the model-agnostic claim."},{"cited_title":"Experiments Setup Following standard practice in scene text recognition (STR) [19, 5], we adopt both synthetic and real-world datasets for train- ing and evaluation","cited_arxiv_id":null,"evidence_quote":"ViTSTR baseline; one of the two base architectures that TEACH is added to in the main experiments."},{"cited_title":"TEACH injects ground-truth labels into the input sequence during training and gradually masks them based on model performance","cited_arxiv_id":null,"evidence_quote":"PARSeq baseline; the other base architecture and the one that produces the headline state-of-the-art numbers."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"ABINet; a vision-language baseline that TEACH is compared against to show gains without external language pretraining."},{"cited_title":"What if we only use real datasets for scene text recogni- tion? toward scene text recognition with fewer labels,","cited_arxiv_id":null,"evidence_quote":"DiG-ViT-B; a strong transformer baseline that PARSeq+TEACH surpasses on five synthetic-data benchmarks."},{"cited_title":"Levenshtein ocr,","cited_arxiv_id":null,"evidence_quote":"MAERec-B; a strong recent baseline that PARSeq+TEACH surpasses under real-data training."},{"cited_title":"Multi- modal text recognition networks: Interactive enhance- ments between visual and semantic features,","cited_arxiv_id":null,"evidence_quote":"MJSynth; one of the two synthetic training sets used in the main experiments."},{"cited_title":"Petr: Rethinking the capability of transformer-based language model in scene text recognition,","cited_arxiv_id":null,"evidence_quote":"SynthText; the other synthetic training set used alongside MJSynth."}],"review_version":1}