{"id":"b72eb3db-7de0-4eeb-b82b-64c5b6b0aa5d","arxiv_id":"2508.03566","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"The full text describes LaTCoder, a block-wise chain-of-thought method for webpage design-to-code, reporting improved layout preservation on two benchmarks, while the title and abstract are for a different paper.","lead":"The abstract and title describe SAM2-UNeXT, but the full text is actually a different paper, LaTCoder, about converting webpage screenshots into HTML code. The LaTCoder text reports large layout-preservation gains, but the document's identity mismatch makes the submission as a whole unverifiable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LaTCoder full-method rows in Table 3 are internally inconsistent with the §3.3 dynamic selector: reported MAE is sometimes lower than both APS/MS candidates, and the selector optimizes the same MAE/CLIP metrics used for evaluation, inflating the headline improvement.","rationale":"The reader's UNVERDICTED verdict is motivated by the title/abstract mismatch (SAM2-UNeXT vs LaTCoder), which is a legitimate reason not to evaluate the claimed method. Taking the full text on its own terms, however, the central claim is that block-wise generation improves layout preservation, and Table 3 is the load-bearing evidence for that claim. I find a concrete internal inconsistency: the dynamic selector defined in §3.3 cannot produce the published full-method MAE values, because the reported full-method MAE is sometimes lower than both of the candidate strategies it is supposed to choose between. This indicates either an undocumented refinement step or non-reproducible numbers. Moreover, the selector optimizes the same MAE and CLIP metrics that are used for evaluation, so the full-method comparisons against monolithic baselines partly reflect selection on the evaluation metric rather than the proposed block-wise pipeline. The reader's weakest assumption about the division algorithm is plausible but secondary, since it affects only a minority of benchmark inputs, whereas the Table 3 inconsistency undermines every reported full-method improvement. A re-derivation of the selector from the reported candidate metrics is a decisive check: if the recomputed values do not match, the headline claim is not supported. Given that the submission is also mismatched in identity, the appropriate disposition is rejection rather than an unverdictable status.","tokens_in":19095,"tokens_out":9520,"duration_ms":98759,"concrete_test":"Recompute the 'LaTCoder' rows of Table 3 from the reported 'LaTCoder (APS)' and 'LaTCoder (MS)' rows using the Verify Score formula in §3.3. If the selected candidate's MAE and CLIP do not match the published 'LaTCoder' values (e.g., GPT-4o/Design2Code-HARD should give MAE=36.21, not 33.93), the full-method results are internally inconsistent and not reproducible from the described procedure. As a corroborating check, apply the same verify-score selection to pairs of monolithic baselines (e.g., select between direct and self-revision) and rerun Table 3; if the block-wise advantage over this best-of-two baseline shrinks or vanishes, the headline improvement is a selection artifact rather than an effect of Layout-as-Thought.","verdict_should_be":"REJECT","load_bearing_attack":"The central empirical claim, that block-wise Layout-as-Thought generation significantly boosts MLLM layout preservation, rests on Table 3. Section 3.3 describes a dynamic selector that computes Verify Score = 0.5*(1-MAE/255) + 0.5*CLIP and chooses between the APS and MS outputs. Given the published per-candidate metrics, this selector must return one candidate's output, so the full method's MAE should equal that candidate's MAE. But the reported 'LaTCoder' rows are arithmetically impossible under this rule: for GPT-4o on Design2Code-HARD, APS has MAE 36.21 and CLIP 0.86, MS has MAE 59.55 and CLIP 0.86, so the selector must pick APS and report MAE 36.21, yet the table reports 33.93. The same pattern appears for Gemini (37.50 vs 40.21/59.72) and for GPT-4o on CC-HARD (36.80 vs 37.53/58.29). Either the full pipeline includes an undocumented refinement or regeneration step after selection, or the numbers are not reproducible from the described procedure. Compounding this, the selector's objective is a convex combination of exactly two evaluation metrics (MAE and CLIP) reported in Table 3, so selecting on it guarantees the selected output looks better on those metrics by construction. The comparison against monolithic baselines (direct, text-augmented, self-revision, DCGen) therefore conflates the benefit of block-wise division/generation with the benefit of picking the better of two candidates on the evaluation metric. The claim that gains are largest for weaker models relies on DeepSeek-VL2, where only APS is used (no selection); there CLIP actually decreases (-4.94% and -1.33%), so content-similarity improvement is not established. The division-algorithm limitation identified by the reader is real but secondary: Table 1 shows it affects only 4-14% of benchmark inputs, whereas the Table 3 inconsistency affects every reported full-method gain.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript, despite the metadata header naming a different title, presents LaTCoder, a divide-and-conquer pipeline for converting webpage screenshots into HTML/CSS code with multimodal LLMs. The method first partitions the design into rectangular blocks using a hand-tuned line-detection algorithm, generates code for each block with a Chain-of-Thought style prompt, assembles the block codes through either absolute positioning (APS) or MLLM-based assembly (MS), and finally selects between the two assembled outputs with a verifier that combines MAE and CLIP similarity. The authors evaluate LaTCoder with DeepSeek-VL2, Gemini, and GPT-4o on Design2Code-HARD and on a new CC-HARD benchmark, reporting improvements in TreeBLEU, CLIP, Visual Score, and MAE, plus a pairwise human preference study. The core technical idea is plausible and the new benchmark is a useful contribution, but the headline full-method results are confounded by test-time selection on the evaluation metrics, and several supporting analyses are incomplete.","tokens_in":19505,"tokens_out":7565,"duration_ms":88207,"significance":"If validated, LaTCoder would provide a simple and effective block-wise generation baseline for design-to-code, and CC-HARD would be a useful stress test for layout preservation. The paper also takes reproducibility seriously: it fixes temperature to 0, sets random seeds, and releases code and benchmark links. The main significance is conditional, however, because the full-method rows in Table 3 are produced by a verifier whose score is a convex combination of two of the four reported evaluation metrics; this makes the comparison against single-shot baselines non-equivalent unless the selection effect is quantified and separated. The APS-only results, which are not affected by this confound, still show substantial MAE and TreeBLEU gains, so the underlying approach is worth further scrutiny rather than dismissal.","major_comments":[{"comment":"The dynamic selection verifier is defined as Verify Score = 0.5*(1-MAE/255) + 0.5*CLIP, and MAE and CLIP are also two of the four headline metrics in Table 3. Therefore the 'LaTCoder' row in Table 3 is not the output of a single generation pipeline but the per-sample best of two candidates according to a weighted combination of the evaluation metrics themselves, plus empirically set weights (0.5/0.5). This guarantees an advantage on those metrics relative to either candidate and conflates the contribution of block-wise generation/assembly with the contribution of test-time selection on the target metric. I note that the full-method MAE being lower than both the APS and MS average MAEs is not by itself arithmetically impossible, since per-sample min-selection can lower the average; however, the paper does not report per-sample selection counts or the metrics on the exact selected outputs, so the reported full-method values cannot be independently verified. The authors should either make APS/MS without selection the primary comparison, or use a verifier based on a metric not included in the evaluation table, and should report the selection statistics.","section":"§3.3, Eq. (1), Table 3"},{"comment":"The layout-aware division algorithm assumes rectangular blocks separated by solid-colored horizontal/vertical lines that do not cross text regions. Table 1 explicitly reports that 13.75% of Design2Code-HARD and 4.33% of Design2Code contain irregular layouts, overlapping layouts, or gradient backgrounds. The paper does not state how many samples fail the division step or what fallback is used, even though every downstream step inherits the block decomposition. Because the aggregate scores may be dominated by the 86.25% of 'regular' cases, the authors should provide a per-case failure analysis of the divider and report performance separately on the non-standard subset, or describe a fallback for samples that do not satisfy the rectangular-line assumption.","section":"§3.1, Table 1"},{"comment":"Several of the claimed improvements are small relative to the reported standard deviations. For example, with GPT-4o on Design2Code-HARD the CLIP score changes from 0.84±0.08 for direct prompting to 0.87±0.07 for LaTCoder, and on CC-HARD the CLIP gain is 0.79±0.10 to 0.81±0.09. No paired significance tests are reported for any automatic metric, so statements such as 'significantly boosts' in Section 5.1 are not statistically supported. The human evaluation in Section 5.3 and Figure 5 covers only GPT-4o on CC-HARD and omits the total number of paired comparisons, the handling of ties, and inter-annotator agreement. Please add paired tests for the automatic metrics and report the full protocol and statistics for the human study.","section":"Table 3, §5.3, Figure 5"}],"minor_comments":[{"comment":"The submission metadata identifies the paper as arXiv:2508.03566 (cs.CV), 'SAM2-UNeXT', but the body is an ACM KDD '25 paper titled 'LaTCoder' with arXiv:2508.03560 (cs.SE). The editor should verify that the correct manuscript and identifier were submitted.","section":"Metadata"},{"comment":"The text says the line-detection algorithm is applied recursively, but Algorithm 1 as written has no recursion or termination condition for recursion. Please clarify the recursive call and the stopping criterion.","section":"§3.1, Algorithm 1"},{"comment":"The verify-score formula is not numbered and the symbols MAE and CLIP are not formally defined at that point. Consider numbering the equation and noting that MAE is normalized by 255 and CLIP is cosine similarity in [0,1].","section":"§3.3, Eq. (1)"},{"comment":"The table footnote says 'best performance among three variances'; 'variances' should be 'variants', and the text should be explicit that for DeepSeek-VL2 only the APS variant is available.","section":"Table 3"},{"comment":"The abstract and Section 5.3 state that annotators prefer LaTCoder in 'over 60% of cases', but the figure does not report the denominator or the number of samples. Please include the exact counts and the majority-voting details.","section":"§5.3"},{"comment":"The claim that MLLM-based assembly 'often produces better overall aesthetics and smoother transitions' is a subjective assertion without quantitative or visual evidence beyond one case study. Please provide support or temper the claim.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The body of the manuscript is a camera-ready ACM KDD '25 paper with a DOI, while the provided metadata names a different arXiv paper. This is a serious administrative discrepancy that the editor should resolve before further processing. If this is a journal submission of an already-published conference paper, the novelty and self-overlap policies need to be checked. Scientifically, the central design-to-code idea is promising, but the full-method evaluation is confounded by selection on the evaluation metrics, so the revision should focus on separating the generation effect from the selection effect."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The arXiv record is not what it claims: the abstract and title describe SAM2-UNeXT, a segmentation method, but the full text is a KDD paper about LaTCoder, a webpage design-to-code pipeline. I read the LaTCoder content as the actual submission.\n\nWhat LaTCoder contributes is real: an automated line-based division algorithm with OCR text protection, block-wise chain-of-thought prompting, two assembly strategies with a dynamic selector, and a new harder benchmark, CC-HARD. The CoT ablation shows a clear effect, and the human preference study is a reasonable addition. The pipeline is practical and the benchmark should be useful to the design-to-code community.\n\nThe soft spots are not minor. Table 3 reports full-method MAEs that are better than both candidate assembly strategies. For GPT-4o on Design2Code-HARD, APS gives MAE 36.21 and MS gives 59.55. The dynamic selector computes Verify Score = 0.5*(1-MAE/255) + 0.5*CLIP and then picks one candidate, so the full method should inherit that candidate's MAE. The reported 33.93 is arithmetically impossible under the described procedure. The same pattern appears for Gemini and on CC-HARD. Either there is an undocumented refinement step after selection, or the numbers are not reproducible.\n\nSecond, the selector optimizes exactly the two metrics, MAE and CLIP, that are also the paper's evaluation endpoints. Selecting on those metrics means the headline gains against single-shot baselines partly come from test-time metric optimization, not from block-wise generation alone. That confound is load-bearing.\n\nThird, the claim about weaker models relies on DeepSeek-VL2, where CLIP actually decreases, so content-similarity improvement is not established there.\n\nThe division-algorithm thresholds are hand-set, but that is a secondary issue; the Table 3 inconsistency affects every reported full-method gain.\n\nRecommendation: this arXiv submission should be sent back to the authors to fix the identity mismatch and to explain or correct the full-method numbers. Once those are resolved, LaTCoder deserves a serious review. As it stands, the central empirical claim does not hold up.","headline":"The arXiv record is not what it claims: abstract/title describe SAM2-UNeXT, body is LaTCoder, and the LaTCoder results have internal number inconsistencies that undermine the main empirical claim.","tokens_in":20109,"tokens_out":3411,"would_cite":false,"duration_ms":38137,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Block-by-block webpage code generation, anchored to the design's detected layout regions, preserves layout far better than whole-page generation, especially for smaller multimodal models.","keywords":["design-to-code","layout preservation","multimodal large language models","chain-of-thought","block-wise generation","layout-aware division","CC-HARD benchmark","TreeBLEU"],"falsifier":"Take a set of webpage designs with gradient backgrounds, overlapping sections, and irregular grids where the dividing algorithm is shown to produce wrong or misaligned blocks; if the full LaTCoder pipeline performs no better than monolithic generation on those cases, the claim that block-wise layout anchoring is what drives the improvement is refuted. A sharper version: replace the algorithm's detected bounding boxes with ground-truth boxes; if results improve substantially, then the divider, not anchored generation, is the true bottleneck.","tokens_in":18941,"feed_emoji":"🧩","tokens_out":8595,"duration_ms":91280,"temperature":0.7,"pith_summary":"Dozens of multimodal language models (MLLMs) now promise to turn a webpage screenshot into HTML and CSS, but they routinely scramble the layout because they try to generate the entire page in one pass. LaTCoder instead treats the design as a set of layout 'thoughts': a lightweight algorithm divides the image into rectangular blocks along solid-colored lines (using OCR so lines never cut through text), each block's code is generated separately with a chain-of-thought prompt, and the blocks are reassembled at their original coordinates. The paper reports that, across GPT-4o, Gemini, and DeepSeek-VL2 on two benchmarks, this block-wise pipeline improves DOM-tree structure similarity and pixel error relative to direct prompting and prior divide-and-conquer baselines, with the largest relative gains for the smallest model. It also contributes CC-HARD, a harder benchmark with deeper DOM trees and more complex layouts, and a human preference study that favors LaTCoder in most pairwise comparisons.","feed_headline":"Cutting webpages into blocks sharply improves AI-generated layout","feed_subtitle":"LaTCoder splits a design into anchored blocks, codes each block, then reassembles — weak models gain the most.","key_machinery":"The load-bearing object is the Layout-as-Thought (LaT) pipeline, a block-wise code-generation process in which each image block is treated as an independent 'thought' for a multimodal model to convert into code. The dividing algorithm is the critical mechanism: it scans rows and columns at a 5-pixel grid interval for solid-colored lines at least 50 pixels apart, uses OCR to ensure no dividing line crosses text regions, ignores edge pixels, merges blocks smaller than 300×300 pixels, and recurses to depth 3, producing a set of bounding boxes (BBoxes) that anchor every block. Each cropped block is then fed to an MLLM with a chain-of-thought prompt that instructs it to analyze the section, generate code, and check text, color, and style consistency. Finally, two assembly strategies — absolute positioning (APS) and MLLM-based assembly (MS) — are combined via a reference-free verifier that computes a score as the equal-weight sum of normalized MAE and CLIP similarity, dynamically selecting the better output.","core_discovery":"The central claim is that Layout-as-Thought (LaT) — decomposing a webpage design into rectangular image blocks via a lightweight, OCR-aware line-detection algorithm, generating each block's HTML/CSS code independently with a chain-of-thought prompt, and reassembling the blocks using their recorded bounding boxes — alleviates the layout-information loss inherent in monolithic design-to-code generation. LaTCoder reports consistent improvements over direct prompting and prior divide-and-conquer baselines across three backbone MLLMs (GPT-4o, Gemini, DeepSeek-VL2) and two benchmarks: on Design2Code-HARD it raises TreeBLEU from 0.16 to 0.20 for GPT-4o, and on the new CC-HARD benchmark it raises TreeBLEU from 0.09 to 0.16, reducing MAE by roughly 43%. The paper further claims that the relative gains are largest for the weakest model (DeepSeek-VL2), and that human annotators prefer LaTCoder's output in at least 60% of pairwise comparisons against each baseline.","pith_inferences":["The same divide-and-conquer principle could extend to other structured visual-to-code tasks, such as mobile UI screens, dashboards, or document layouts, although the hand-set division thresholds would likely need retuning.","The line-based divider is the fragile part of the pipeline; using a learned layout parser or structural knowledge from the source would be a natural next step for irregular, gradient, or overlapping designs.","The finding that decomposition gives the largest boost to the weakest model suggests that block-wise generation acts as an inference-time substitute for model capacity, which may generalize to other long-context generation problems."],"forward_implications":["Adopting block-wise, layout-anchored generation should improve layout fidelity and pixel accuracy for any MLLM doing design-to-code, with no fine-tuning required.","Smaller open-source MLLMs such as DeepSeek-VL2 see the largest relative gains, which narrows the gap to commercial models on this task.","The CC-HARD benchmark provides a more discriminating evaluation of layout and structural complexity than Design2Code-HARD, as shown by larger performance gaps across methods.","Absolute-positioning assembly is responsible for the largest MAE reductions, while MLLM-based assembly yields better aesthetic continuity; combining them via the verify score improves overall output.","The CoT-based block prompt is a major contributor: replacing it with a simplified prompt drops TreeBLEU from 0.16 to 0.13 and visual score from 0.80 to 0.71 on CC-HARD."],"supporting_citations":[{"why":"Supplies the Design2Code-HARD benchmark and the direct, text-augmented, and self-revision baselines that LaTCoder must outperform.","marker":"[42]"},{"why":"Introduces the WebCode2M dataset-driven TreeBLEU metric used to measure DOM-tree structural similarity.","marker":"[16]"},{"why":"DCGen is the prior divide-and-conquer baseline that also generates code from screenshots with subregion descriptions; LaTCoder compares against it.","marker":"[47]"},{"why":"DeepSeek-VL2 serves as the open-source backbone MLLM on which LaTCoder reports the largest relative gains.","marker":"[30]"},{"why":"Gemini is a commercial backbone MLLM used to evaluate the method's generalizability.","marker":"[3]"},{"why":"GPT-4o is the strongest backbone MLLM used in the main experiments and in the human preference study.","marker":"[32]"},{"why":"Provides the CLIP similarity used both as a content-similarity metric and as a component of the verify score.","marker":"[36]"},{"why":"Web2Code motivates the MLLM-as-a-judge verifier idea that the paper tests and ultimately replaces with the hybrid MAE+CLIP score.","marker":"[54]"},{"why":"The screenshot-to-code project provides the motivating example of layout failure in monolithic generation.","marker":"[2]"}],"fun_headline_variants":["SAM2-UNeXT adds DINOv2 and dual-resolution for sharper segmentation","High-res SAM2-UNeXT improves segmentation on four benchmarks","SAM2-UNeXT: A simple baseline that extends SAM2 with DINOv2","SAM2-UNeXT: Dual encoder and dense glue lift segmentation accuracy","Upgraded SAM2 with DINOv2: SAM2-UNeXT for downstream tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a webpage design can be split into rectangular blocks separated by solid-colored lines that never cross text; if a design has gradients, overlaps, or irregular elements, the division step can mangle it and the rest of the pipeline inherits the error.","fun_headline_variants_meta":{"raw":{"variants":["SAM2-UNeXT adds DINOv2 and dual-resolution for sharper segmentation","High-res SAM2-UNeXT improves segmentation on four benchmarks","SAM2-UNeXT: A simple baseline that extends SAM2 with DINOv2","SAM2-UNeXT: Dual encoder and dense glue lift segmentation accuracy","Upgraded SAM2 with DINOv2: SAM2-UNeXT for downstream tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1462,"prompt_tokens":927,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":425}},"tokens_in":543,"tokens_out":535,"duration_ms":6180,"temperature":1.0,"reasoning_tokens":425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:20:57.527762+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of webpage designs with gradient backgrounds, overlapping sections, and irregular grids where the dividing algorithm is shown to produce wrong or misaligned blocks; if the full LaTCoder pipeline performs no better than monolithic generation on those cases, the claim that block-wise layout anchoring is what drives the improvement is refuted. A sharper version: replace the algorithm's detected bounding boxes with ground-truth boxes; if results improve substantially, then the divider, not anchored generation, is the true bottleneck.","supporting_citations":[],"review_version":1}