{"id":"58b502e9-d513-4476-a530-d9c5e98144da","arxiv_id":"2505.14059","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A two-stage model that analyzes page layout first, then parses text, tables, and formulas in parallel, reports state-of-the-art accuracy and speed on the benchmarks it evaluates.","lead":"Dolphin is a 322-million-parameter document parser that first identifies layout elements in reading order, then crops and parses each element in parallel with task-specific prompts. It reports lower error rates on several page and element benchmarks than much larger models such as GPT-4o and Claude, at roughly twice the throughput of the fastest prior system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Train/test overlap in table training (PubTabNet/PubTab1M) and un-released self-constructed Dolphin-Page/Dolphin-Block benchmarks leave page-level and table SOTA claims unverifiable; a duplicate-image audit would settle it.","rationale":"The paper's contribution is an engineering design—two-stage anchor-based parsing with parallel decoding—and nothing in the architecture itself is contradicted by the text. The FPS advantage (0.1729 vs 0.0971 sequential) and the ablation of type-specific prompts are internally consistent. The mathematical/encoding descriptions are standard. What makes the headline SOTA claim load-bearing is the provenance of the evaluation data. The reader's weakest_assumption correctly identifies that the paper trains on PubTabNet/PubTab1M and evaluates on those same families without ruling out test-split leakage, and that the self-constructed Dolphin-Page/Dolphin-Block benchmarks are neither released nor attested as disjoint from the 30M-sample training corpus. I agree with this assessment; it is the most security-critical assumption. If the overlap exists, the page-level and table results are not evidence of generalization, and the abstract's SOTA claim is unsupported. If the overlap does not exist, the results are plausible and the contribution is a worthwhile, efficient alternative to autoregressive VLMs. The public formula benchmarks provide weak independent support, but they are element-level and the paper does not claim to beat UnimerNet/Mathpix there (Table 3). Therefore the verdict should remain CONDITIONAL, with the condition being a released, overlap-audited benchmark and an explicit statement of train/test split exclusion. I would not move to REJECT because there is no internal inconsistency and the architecture is sensible; I would not ACCEPT until the data hygiene issue is resolved.","tokens_in":17976,"tokens_out":6589,"duration_ms":63629,"concrete_test":"Conduct a duplicate-image audit: compute exact and perceptual hashes (e.g., pHash) for all 7,904 PubTabNet test images, 10,000 PubTab1M test images, 210 Dolphin-Page pages, and 1,856 Dolphin-Block paragraph crops; then scan the released 30M training corpus (or a 1M random subset if full release is impractical) for exact or near-duplicate matches. If any test image matches a training image, the corresponding reported metric is invalid until the model is retrained on a disjoint set; if zero matches are found and the authors additionally document that Dolphin-Page/Dolphin-Block sources are disjoint from the Mixed Documents and synthetic HTML/LaTeX/Markdown corpora, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim rests on benchmark hygiene that the paper never establishes. Section 4.1 says table training uses PubTabNet and PubTab1M; Section 4.2 evaluates on those same benchmark families but never states that the official test splits (7,904 and 10,000 images) were excluded from training. Any overlap would let the model memorize test tables and inflate the reported TEDS 0.9515/0.9625. The new Dolphin-Page (210 pages) and Dolphin-Block (1,856 paragraphs) benchmarks are self-constructed, not released, and no disjointness from the 30M training corpus is attested; the training pipeline includes 0.12M 'Mixed Documents' from educational, publication, and business sources, so overlap with the benchmark's manually annotated pages is plausible. Because the two strongest reported results (page-level edit distances 0.0114/0.0131/0.1028 and table TEDS) depend on these unverified sets, the headline claim is conditional at best. The public formula benchmarks (SPE/SCE/CPE) provide some independent support, but they do not cover the core page-level claim.","agreement_with_reader":"agree"},"referee_report":null,"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-07T15:41:14.703657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}