{"id":"57eb40bc-fc22-4cb3-9fec-e90b44becd18","arxiv_id":"2506.23106","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A neural-network recognizer is used to find stroke-removal patterns that preserve character identity, revealing that many multi-stroke characters tolerate losing about a third of their strokes.","lead":"The paper automatically searches for which strokes can be removed from complex Japanese and Chinese characters before a character-recognition model starts confusing them with other characters. It finds that many characters keep their recognizability even after a third of strokes are gone, which could guide simpler fonts and learning materials.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline claim is about human legibility, but the only evidence is a classifier posterior that the authors themselves concede can overestimate readability (Sec. 5.4); without a human reading test, the central claim is conditional at best.","rationale":"The reader's weakest_assumption — that Pθ(cx | x_-k) is an unvalidated proxy for human legibility — is exactly the load-bearing concern I identify. The paper's own text supports this: Sec. 3.1 introduces the proxy explicitly, Sec. 5.4 concedes that native speakers can find a shape unreadable before the classifier does, and the Conclusion repeats the limitation. The argmax selection over all removal patterns makes the reported high posteriors a best-case estimate, which can mask proxy error. This is not an internal inconsistency; the computational pipeline is coherent and the reported statistics about Pθ are plausible. The issue is external validity: every headline statement about 'legibility' in the abstract and introduction is a human-facing claim, and no human data are provided. The proposed human forced-choice test would directly settle whether the one-third-removal claim transfers to human readers. If it passes, the central claim is empirically supported; if it fails, the paper's contribution should be reframed as a study of classifier behavior under stroke removal, not of human legibility. The reader's CONDITIONAL verdict is therefore appropriate and I do not move it: the paper is transparent about the limitation and the computational results are self-consistent, but the human-facing conclusion remains conditional pending validation.","tokens_in":10257,"tokens_out":3883,"duration_ms":43274,"concrete_test":"Run a human reading experiment on the actual x*_-k outputs. Sample 30 characters per K∈{5,10,15,20}; for each, take x*_-k at k=floor(K/3) (the 'one-third removal' claim) and the largest k for which Pθ remains 1.0 (the 'minimum stays 1.0' result). Render these 240 simplified images plus 30 originals as controls; present to 20 native Japanese readers in a forced-choice task among the true class and 3 confusable same-stroke-count distractors. Compare human accuracy to Pθ at the character level. If mean human accuracy at the one-third point falls significantly below 95% (or is significantly lower than Pθ), the proxy assumption fails and the headline claim must be downgraded to a statement about classifier confidence; if human accuracy tracks Pθ, the concern is empirically retired.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the identification of 'legibility' with DaKanji's posterior Pθ(cx | x*_-k), introduced in Sec. 3.1 ('we use computed legibility as an approximation of human legibility'). The paper's headline results (Sec. 5.1: minimum computed legibility stays 1.0 for k up to 2/5/8 for K=10/15/20) are statements about this posterior, and the chosen x*_-k is the argmax over all C(K,k) removal patterns. Two compounding issues make this proxy insecure for the human-facing claim. First, DaKanji was trained on intact KanjiVG characters; simplified characters are out-of-distribution, and softmax confidence on OOD inputs is not calibrated to human perception. Second, the argmax selection means the reported high posterior is a best-case over many candidates, so even a small systematic overestimation in the proxy is amplified. The authors themselves provide direct evidence of the mismatch in Sec. 5.4: 'native speakers... find the shape unreadable just before it becomes non-legible,' and the Conclusion repeats the limitation. Thus the central claim that one-third stroke removal 'leaves legibility nearly unchanged' is, as currently evidenced, a claim about classifier confidence rather than human legibility; the human-facing conclusion is conditional on validation the paper does not provide.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a computational framework for simplifying multi-stroke CJK characters by deleting strokes. For each character x with K strokes, the method enumerates all C(K,k) removal patterns, renders them as 64x64 images from KanjiVG vector data, and uses the softmax posterior P_theta(c_x | x_-k) of a pretrained DaKanji classifier as a measure of \"computed legibility.\" The pattern maximizing this posterior is selected as the optimal k-stroke removal. Experiments on 1,256 characters with K = 5, 10, 15, and 20 report average and minimum computed-legibility curves, a random-removal baseline, per-character removal tolerance, example optimal removal sequences, and an analysis of whether short or long strokes tend to be removed first. The headline observations are that computed legibility often stays near 1.0 until approximately one-third of the strokes are removed, that removal tolerance varies strongly across characters, and that optimal removals are spatially balanced across radicals. The authors acknowledge in Sections 5.4 and 6 that the classifier is only an approximation of human legibility and can overestimate readability.","tokens_in":10527,"tokens_out":9140,"duration_ms":97231,"significance":"If read as a study of a specific recognizer's robustness to stroke deletion, this is a solid and clearly written contribution: exhaustive enumeration over removal patterns is appropriate, the random-removal control is well designed, and the use of a publicly available pretrained model aids reproducibility. The observations about structural redundancy, radical preservation, and stroke length are plausible and potentially useful for font design and for teaching materials. However, the paper repeatedly frames the results in terms of human legibility, while the evidence is entirely classifier confidence. Because the same posterior is used both to select and to score the simplified characters, and because no human reading experiment or perceptual rating is reported, the human-facing significance is not established. The paper's value would remain substantial if the claims were explicitly scoped to computed legibility, or if a human validation study were added.","major_comments":[{"comment":"The headline claim of the abstract and introduction, namely that simplification by stroke removal preserves legibility and that up to about one-third of strokes can be removed, is worded in terms of human legibility, but the evidence is entirely the posterior P_theta(c_x | x*_-k) of DaKanji. The same posterior is used to select the optimal removal in Eq. (1) and to score it in Eq. (2) and Fig. 6, so the reported \"minimum computed legibility stays 1.0\" result is partly by construction. The manuscript itself concedes in Section 5.4 that native speakers find some shapes unreadable just before the classifier stops recognizing them, and the conclusion lists human-centered validation only as future work. Because no human reading experiment or perceptual rating of the generated stimuli is reported, the human-facing interpretation is unsupported. The computational results would remain a meaningful contribution if the abstract, introduction, and conclusion were respecified as claims about recognizability by DaKanji, or if a human validation study were added, even on a stratified sample around the threshold region.","section":"Section 3.1, Eq. (1), Section 5.4"},{"comment":"The definition T_x = sum_{k=1}^{K-1} P_theta(c_x | x*_-k) is inconsistent with the plotted range in Fig. 10 and with the text in Section 5.3 stating that T_x is around 0.5 even for the least tolerant characters. For K = 20, Section 5.1 reports that the minimum computed legibility is 1.0 for all characters for k = 1,...,8, so Eq. (2) would give T_x at least 8 for those characters, far above the Fig. 10 axis that ends at 1.0. If the figure and the text actually use a normalized value such as T_x/(K-1), the definition must say so; if Eq. (2) is intended as written, the y-axis, the \"T_x ~ 0.5\" statement, and the ranking in Fig. 11 need to be corrected. This issue is load-bearing for the tolerant-versus-intolerant characterization.","section":"Section 3.3, Eq. (2), Fig. 10"},{"comment":"DaKanji is pretrained on intact KanjiVG characters, so the simplified images produced by stroke removal are out-of-distribution for the classifier, and softmax confidence on out-of-distribution inputs is not automatically a proxy for human readability. The paper reports only the model's 99.8% accuracy on intact images and the 98% accuracy on a second font, but no calibration or reliability check on simplified inputs. Section 5.1's observation that removing a stroke can raise computed legibility from 0.90 to 1.00 for the same character illustrates that the metric behaves in a way that is not obviously perceptual. The random-removal baseline is a good control for the relative benefit of selection, but it does not anchor the absolute scale of P_theta for human legibility. The authors should report the distribution of posterior values over all C(K,k) candidates for a subset of characters and, ideally, compare model confidence with human ratings on the same stimuli.","section":"Section 4.2 and Section 5.1"},{"comment":"The conclusions that tolerant characters often contain distinctive hook-like strokes and that intolerant characters tend to have more horizontal strokes are based on visual inspection of only five characters per stroke count, with no quantitative feature analysis or statistical comparison against the full set of 1,256 characters. Since this underlies the paper's finding about tolerance and stroke complexity, the claim should be supported by counts of stroke-shape types, by a comparison with the overall distribution, or by a statistical test.","section":"Section 5.3 and Fig. 11"}],"minor_comments":[{"comment":"The notation K C k is used without a formal definition; write binomial coefficient notation or explicitly define it as \"K choose k.\"","section":"Section 4.1"},{"comment":"The caption uses N in {5, 10, 15, 20} while the text uses K for the number of strokes; be consistent throughout.","section":"Fig. 11"},{"comment":"The observation that removing a stroke can increase computed legibility from 0.90 to 1.00 is surprising and deserves more than a parenthetical remark; it highlights the non-monotonic behavior of the proxy and should be discussed explicitly.","section":"Section 5.1"},{"comment":"State explicitly that the random baseline averages over all C(K,k) patterns while the proposed method selects the maximum; the comparison is valid, but the difference includes a max-versus-mean effect.","section":"Section 5.2 and Fig. 9"},{"comment":"A short reproducibility statement about releasing the generated simplified-character images and the exact KanjiVG subset would strengthen the paper.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The work is fundamentally sound as a computational study of a recognizer's stroke-deletion robustness, and the authors are transparent about the limitation. The main gap is the mismatch between the human-legibility framing and the classifier-only evidence, which the authors themselves acknowledge in Section 5.4. If the claims are respecified or a human validation study is added, I would support publication. There are no concerns about novelty or attribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kenta, quick take on arXiv:2506.23106. This is the first computational attack on multi-stroke character simplification by stroke removal I know of, and it does it cleanly. They enumerate all C(K,k) removal patterns for each character, score each with DaKanji's posterior for the original class, and keep the argmax. The random-removal baseline is the right control, and the removal-tolerance metric is simple but useful. The empirical findings—roughly one-third of strokes can go before the posterior dips, critical strokes cause a sharp drop, and optimal removals are balanced across radicals—are concrete and reproducible from KanjiVG plus the public DaKanji model. The non-monotonicity of the optimal removal sequence is a nice catch.\n\nThe soft spot is exactly where the stress test puts it: the paper equates posterior probability with legibility. The same classifier picks x*_-k and then scores it, so the high numbers in Fig 6 are partly by construction. DaKanji was trained on intact characters; simplified shapes are out-of-distribution, and softmax confidence on OOD inputs is not calibrated to people. The authors are honest about this—Section 5.4 says native speakers find the shape unreadable just before the recognizer gives up, and the conclusion repeats the limitation. So the central 'legibility' claim is really about recognizer confidence. That is not fatal; it just means the paper should be framed as a study of the recognizer's tolerance, with human validation as future work. The behavioral claims about stroke redundancy are interesting in their own right, but they'd be stronger with even a small human reading experiment.\n\nMinor concerns: only KanjiVG as rendering source, though they do check a second font for the recognizer; and the four stroke counts are a practical necessity but leave the 20+ stroke behavior untested. The related-work section is honest and found no prior computational simplification work, which I believe.\n\nBottom line: this deserves a serious referee. I'd encourage the editor to send it out. The framework is well-executed and the limitations are visible. A careful reviewer could push for reframing or a human validation, but the computational core is solid.","headline":"A clean, reproducible computational study of stroke removal in kanji, but the headline 'legibility' is classifier confidence, not human reading; worth reviewing, with the human-facing claims treated as conditional.","tokens_in":11038,"tokens_out":2128,"would_cite":true,"duration_ms":21863,"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":"Removing up to about one-third of a multi-stroke character's strokes leaves computed legibility nearly unchanged, the paper claims, so stroke removal can yield simplified characters without losing recognizability.","keywords":["stroke removal","character simplification","kanji","computed legibility","character recognition","removal tolerance","multi-stroke characters"],"falsifier":"Conduct a human identification experiment: show native readers the optimal removal sequences for the 1,256 characters (or a sample of the K=10, 15, and 20 classes) and measure how many strokes can be removed before identification accuracy drops below a threshold such as 80%. If human accuracy collapses far before the classifier's score leaves 1.0, the central robustness claim is false; the paper's own observation of unreadability before classifier failure suggests this outcome is plausible.","tokens_in":10064,"feed_emoji":"✂️","tokens_out":5314,"duration_ms":52337,"temperature":0.7,"pith_summary":"The paper tries to establish that multi-stroke characters (Chinese/Japanese kanji) can be simplified by deleting whole strokes without losing legibility, and that the best strokes to delete can be found automatically. It uses a high-accuracy character classifier, DaKanji, as a stand-in for human legibility, and searches every subset of strokes to find the k-stroke removal that keeps the classifier's confidence in the original character highest. Across 1,256 characters with 5, 10, 15, and 20 strokes, the paper reports that up to about one-third of strokes can be removed while the computed legibility stays nearly unchanged, and that this holds for every tested character up to certain removal counts. The payoff would be a principled, repeatable procedure for designing simplified fonts and learning materials, instead of the current manual heuristics. The central caveat, stated by the authors, is that the classifier's score is only an approximation of human legibility and can overestimate real readability.","feed_headline":"Removing one-third of strokes can keep characters recognizable","feed_subtitle":"A classifier-guided search finds which kanji strokes can be deleted safely, enabling principled simplification.","key_machinery":"The central object is the optimal k-stroke removal x*_-k, defined by x*_-k = argmax_{x_-k} Pθ(cx | x_-k), where θ is the pretrained DaKanji classifier (an EfficientNet Lite-0 network) and x_-k ranges over all C(K,k) ways to delete k strokes from a K-stroke character. This argmax turns an intractable human-legibility judgment into a reproducible score, and the sequence x*_-1, ..., x*_{-(K-1)} (the optimal removal sequence) is what the paper analyzes. A second quantity, removal tolerance T_x = sum_k Pθ(cx | x*_-k), ranks characters by how fast their computed legibility collapses as strokes are removed.","core_discovery":"The central claim is that stroke removal can be optimized by choosing, for each k, the subset of k strokes whose removal leaves Pθ(cx | x_-k), the classifier's posterior probability for the original class, as large as possible. With that optimal choice, the paper reports that computed legibility stays at 1.0 for all characters with K=10 up to two removals, K=15 up to five removals, and K=20 up to eight removals, and that on average legibility remains high until roughly one-third of strokes are gone. Some characters even become more legible to the classifier after one or two strokes are removed. The paper further claims that the optimal removal sequence is not greedy (the best way to remove k+1 strokes is not always an extension of the best way to remove k), that random removal is far more damaging than optimal removal, and that strokes are removed in a balanced way across the character so that no radical disappears entirely.","pith_inferences":["The paper's robustness result resembles error-correcting-code redundancy; a natural extension is to quantify the minimum number of strokes needed for unique identification of each character class as an information-theoretic capacity.","Because the classifier's score can increase after removing a stroke, some original strokes are actually noise for recognition; a testable extension is to check whether native readers also prefer the cleaned forms.","The paper's own admission that native speakers find shapes unreadable just before classifier failure suggests a human-legibility experiment would likely shift the one-third threshold downward, perhaps much more for low-K characters.","The framework could be inverted to help learners: instead of simplifying to the classifier's optimum, use removal tolerance as a measure of which strokes are essential, and prioritize teaching those strokes first."],"forward_implications":["If the claim holds, font designers can use the same search to produce simplified typefaces (like highway or low-resolution fonts) whose stroke deletions are guaranteed to keep a classifier's recognition intact.","The one-third removal threshold gives a concrete target for textbook and learning-material simplification: keep at least two-thirds of strokes, distributed across radicals, to preserve character identity.","The non-monotonic removal sequence implies simplification is not a greedy process, so software tools should search combinations rather than remove strokes one at a time.","The balanced-removal rule (never erase an entire radical) can be stated as a design guideline from the optimal sequences.","The framework extends to any script where strokes are available as vector data, so it can serve as a benchmark for future simplification algorithms."],"supporting_citations":[{"why":"Supplies the prior cognitive-psychology result that removing about 30% of strokes leaves text readability intact, which the paper's one-third plateau corroborates and extends to per-character search.","marker":"[15]"},{"why":"Documents the official 1950s Chinese simplification scheme, the real-world manual baseline that the proposed systematic search is contrasted with.","marker":"[10]"},{"why":"Provides historical evidence on how character simplification has evolved, supporting the paper's claim that no computational simplification approach existed before.","marker":"[4]"}],"fun_headline_variants":["Algorithm finds which strokes to delete yet keep characters legible","Deleting a third of strokes still leaves characters recognizable","Optimized stroke removal preserves legibility in complex scripts","Non-greedy stroke removal yields more legible simplified characters","Simpler characters: globally optimal stroke deletion keeps them clear"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire conclusion depends on treating DaKanji's posterior probability as equivalent to human legibility, and the paper provides no human reading test to confirm that equivalence.","fun_headline_variants_meta":{"raw":{"variants":["Algorithm finds which strokes to delete yet keep characters legible","Deleting a third of strokes still leaves characters recognizable","Optimized stroke removal preserves legibility in complex scripts","Non-greedy stroke removal yields more legible simplified characters","Simpler characters: globally optimal stroke deletion keeps them clear"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000388,"raw_usage":{"total_tokens":2014,"prompt_tokens":879,"completion_tokens":1135,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":1054}},"tokens_in":495,"tokens_out":1135,"duration_ms":12038,"temperature":1.0,"reasoning_tokens":1054,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:48:59.175104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Conduct a human identification experiment: show native readers the optimal removal sequences for the 1,256 characters (or a sample of the K=10, 15, and 20 classes) and measure how many strokes can be removed before identification accuracy drops below a threshold such as 80%. If human accuracy collapses far before the classifier's score leaves 1.0, the central robustness claim is false; the paper's own observation of unreadability before classifier failure suggests this outcome is plausible.","supporting_citations":[{"cited_title":"Reading and Writing25, 951–979 (2011)","cited_arxiv_id":null,"evidence_quote":"Supplies the prior cognitive-psychology result that removing about 30% of strokes leaves text readability intact, which the paper's one-third plateau corroborates and extends to per-character search."},{"cited_title":"Xinhua Publishing House, Beijing, China (1956)","cited_arxiv_id":null,"evidence_quote":"Documents the official 1950s Chinese simplification scheme, the real-world manual baseline that the proposed systematic search is contrasted with."},{"cited_title":"Open Mind6, 264–279 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides historical evidence on how character simplification has evolved, supporting the paper's claim that no computational simplification approach existed before."}],"review_version":1}