{"id":"8a165a7b-3ead-4d09-ba6d-7e0b26a0e051","arxiv_id":"2501.11800","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TFLOP achieves state-of-the-art table structure recognition by replacing cell-box prediction with a layout pointer that directly associates text regions with HTML tags.","lead":"TFLOP is a table structure recognition system that generates table HTML by pointing each detected text region to the correct table cell tag, instead of predicting cell boxes and matching them afterwards. It reports state-of-the-art accuracy on PubTabNet, FinTabNet, and SynthTabNet, and shows it can handle watermarked and non-English tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FinTabNet/SynthTabNet SOTA may be inflated by using ground-truth cell boxes as test-time input while dual-decoder baselines must predict them; the pointer mechanism is not compared on equal footing.","rationale":"The most load-bearing condition for the paper's central claim is that the layout pointer mechanism, not the input boxes, drives the observed gains. The paper's own Sections 3.1, 4.1, and 4.4 reveal that on FinTabNet and SynthTabNet, TFLOP is given ground-truth cell-level bounding boxes at test time, while the compared dual-decoder methods must predict boxes and align OCR text. This is a benchmark-protocol confound rather than a simple robustness gap: TFLOP never has to solve detection or text-region extraction on those benchmarks, and its structure decoder is prompted with exact cell locations. The small PubTabNet-test advantage over VAST (0.35 TEDS) is consistent with the hypothesis that most of the FinTabNet/SynthTabNet margin comes from oracle boxes rather than the pointer mechanism itself. The reader's stated weakest assumption about OCR box quality is real but secondary; the more immediate issue is that two of the three SOTA claims use GT boxes as input, making the comparison with the cited baselines unfair. The proposed test—switching FinTabNet/SynthTabNet to OCR boxes or giving a dual-decoder oracle boxes—would settle whether the pointer contribution is real. This does not require rejecting the architecture; it requires a controlled evaluation. A code release alone would not fix the benchmark protocol unless it includes the OCR-box variant and an oracle-box baseline.","tokens_in":15328,"tokens_out":5291,"duration_ms":55579,"concrete_test":"Re-run TFLOP on the FinTabNet and SynthTabNet test sets using the same PSENet/Master OCR text-region boxes used for PubTabNet test instead of the provided cell-level annotations, and recompute TEDS/TEDS-Struct. If the reported 99.45/99.40 TEDS values drop materially (toward or below the PubTabNet-test level, or below prior SOTA), the FinTabNet/SynthTabNet results are inflated by oracle box input. As a complementary check, run a dual-decoder baseline such as VAST with oracle GT boxes supplied at the matching stage and see whether the gap to TFLOP closes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the layout pointer mechanism is responsible for the reported SOTA is not yet demonstrated, because two of the three headline benchmarks are confounded by oracle box input. Section 3.1 says TFLOP receives text region bounding boxes from 'cell-level annotations or off-the-shelf OCR engines.' Section 4.1 states that FinTabNet and SynthTabNet provide cell-level annotations, and Section 4.4 explicitly describes FinTabNet as 'free from OCR-related noise.' Thus TFLOP is evaluated on FinTabNet/SynthTabNet with ground-truth cell boxes as test-time input, while the compared dual-decoder baselines (TableFormer, VAST, DRCC) must predict cell boxes from the image and then match OCR-extracted text. The pointer mechanism therefore receives oracle layout information that the baselines do not. This affects even TEDS-Struct, since the structure decoder is prompted with exact cell positions. On PubTabNet test, where OCR boxes are used, the TEDS advantage over VAST is only 0.35 (96.66 vs 96.31), whereas the FinTabNet gap is 1.24 points. No ablation compares TFLOP against a dual-decoder that is also given the same GT boxes, and no experiment reports TFLOP with OCR-derived boxes on FinTabNet/SynthTabNet. The SOTA claim is therefore contingent on a controlled comparison that the paper does not provide.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TFLOP, a table structure recognition (TSR) framework that replaces the conventional dual-decoder design (predicting HTML tags and cell bounding boxes, then matching OCR text regions to boxes) with a single decoder that generates OTSL/HTML structure tags while a layout pointer mechanism directly associates each input text-region bounding box with the corresponding data tag. A span-aware contrastive loss is added to improve handling of row/column spans. Experiments report state-of-the-art TEDS/TEDS-Struct on PubTabNet, FinTabNet, and SynthTabNet, with additional demonstrations on watermarked documents and Korean tables.","tokens_in":15635,"tokens_out":6055,"duration_ms":59666,"significance":"The core idea, if validated, is a worthwhile simplification: TFLOP avoids the heuristic box-matching post-processing of dual-decoder systems, and the public release of code supports reproducibility. The span-aware contrastive loss is a reasonable and ablated contribution. However, the headline state-of-the-art claims are not yet convincingly established, because on two of the three main benchmarks TFLOP is evaluated with ground-truth cell boxes, while the dual-decoder baselines must predict them; moreover, the layout pointer mechanism itself is never ablated against a matched baseline that does not use it. The significance of the proposed architecture is therefore contingent on additional controlled experiments.","major_comments":[{"comment":"The state-of-the-art claims on FinTabNet and SynthTabNet are confounded by an input asymmetry. Section 4.1 states that these datasets provide cell-level annotations, and Section 4.4 indicates that on PubTabNet test, where annotations are absent, OCR boxes are used. Thus TFLOP receives ground-truth cell boxes as test-time input on FinTabNet and SynthTabNet, while the dual-decoder baselines (TableFormer, VAST, DRCC) must predict cell boxes and then match OCR text. Because the layout embedding conditions the structure decoder, even the TEDS-Struct scores are not on equal footing. Please provide a controlled comparison: either evaluate TFLOP on FinTabNet/SynthTabNet with OCR-derived boxes, or give a dual-decoder baseline the same ground-truth boxes, or both. Without one of these controls, the 1.24-point FinTabNet TEDS advantage over VAST cannot be attributed to the proposed framework.","section":"§4.1, §4.4, Table 2"},{"comment":"The central mechanism, the layout pointer, is never ablated. TFLOP BASE differs from TFLOP FULL only by the removal of image ROI alignment and span-aware contrastive supervision, so both variants include the pointer. Table 3 therefore measures the contribution of ROI and contrastive losses, not the contribution of the pointer over a dual-decoder head. To support the paper's central claim, an ablation should replace the pointer with a box-predicting head (or otherwise remove the pointer) while keeping the same encoder, layout embeddings, and OTSL tokenization, and should compare end-to-end TEDS after the same (or no) matching post-processing. The discussion in §4.4 describing the TEDS-Struct improvement as a \"side-effect of layout embedding\" is a plausible hypothesis but is not a substitute for this experiment.","section":"§3.5, §4.5, Table 3"}],"minor_comments":[{"comment":"The word \"non-Engligh\" should be \"non-English\".","section":"§5.2"},{"comment":"The word \"constrastive\" should be \"contrastive\".","section":"Figure 3 caption"},{"comment":"The symbol N is used both for the total hidden-state length (B+T) in §3.5 and for the fixed output sequence length in §4.2; please disambiguate (e.g., N_max).","section":"§3.5, §4.2"},{"comment":"The text \"2 x 2 ROIAlign\" should read \"2×2 ROIAlign\".","section":"§3.3"},{"comment":"Please clarify how the temperature τ=0.1 was chosen and confirm whether it is used consistently in Eq. (2) and Eq. (5).","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid engineering contribution with public code, but the novelty of the layout pointer over a dual-decoder baseline is not yet demonstrated. The missing control experiments (oracle-box comparison and a pointer-less ablation) are feasible within the paper's scope, so I do not see grounds for rejection. The paper's reliance on Donut as a shared-author architecture is not problematic, but the incremental novelty over that line of work should be made explicit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the layout pointer is a genuinely different way to do TSR, and the paper is worth reading, but the main SOTA claim is not yet supported. On FinTabNet and SynthTabNet, TFLOP is given ground-truth cell boxes at test time, while the dual-decoder baselines have to predict boxes and match OCR text. That is not an apples-to-apples comparison, and the paper never ablates the pointer itself against a matched dual-decoder head. So the central mechanism's contribution remains unproven.\n\nWhat is actually new: reformulating text region prediction and matching into direct pointing is a real departure from the dual-decoder line. The span-aware contrastive loss with the overlap-weighted coefficient is new as far as I know, and the ablation shows it helps on complex tables. The architecture is clean — Swin/BART with OTSL tokens — and the paper is honest about PubTabNet test requiring OCR boxes. Source code is promised, which is good.\n\nThe soft spots: the oracle box issue is the big one. The TEDS-Struct gains on FinTabNet (99.56 vs 98.63 for VAST) could simply reflect that TFLOP sees perfect cell positions while VAST sees nothing. On PubTabNet test, where both use OCR boxes, the TEDS advantage over VAST is only 0.35 (96.66 vs 96.31) — that is a more credible result, but still confounded by OTSL tokenization and no pointer ablation. The paper also doesn't retrain the watermark baseline, the Korean evaluation is 30 images, and the single-run scores are presented without variance. Those are minor compared to the comparison confound.\n\nBottom line: this is a solid architecture idea presented with plausible but not clean evidence. The right fix is a controlled experiment: same backbone, same tokenizer, train a dual-decoder head that is also given GT boxes (to factor out the information advantage), or run TFLOP with OCR boxes on FinTabNet. If the pointer still wins, the claim stands. I would send it to a serious referee, because the idea deserves a fair test, but I would push hard for that ablation before acceptance. I would not cite the SOTA numbers as they stand.","headline":"A genuinely different TSR architecture whose headline SOTA claim is undermined by oracle box input on two of three benchmarks and a missing pointer ablation.","tokens_in":16125,"tokens_out":2320,"would_cite":false,"duration_ms":22914,"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":"TFLOP claims that a single decoder can output a table's HTML tags and, at the same time, point each tag at the text region that belongs in that cell, removing the box-matching post-processing used by dual-decoder systems.","keywords":["table structure recognition","layout pointer","HTML tag generation","span-aware contrastive learning","OCR text regions","TEDS","PubTabNet","documents with watermarks"],"falsifier":"Run TFLOP on a fixed set of table images under two OCR conditions: once with cell-level text regions and once with those regions deliberately corrupted by merging two adjacent cells' boxes into one and by deleting one cell's box. If TEDS stays unchanged, the central claim that the pointer removes misalignment would be undercut; if every corrupted-box table loses a measurable amount of TEDS, the framework's advantage over dual decoders would depend on input-box quality rather than on the pointer itself.","tokens_in":15108,"feed_emoji":"📊","tokens_out":9616,"duration_ms":83758,"temperature":0.7,"pith_summary":"The paper claims that table structure recognition can be done without first predicting cell bounding boxes and then matching OCR text regions to them. TFLOP's layout pointer mechanism makes the sequence decoder directly associate each generated data tag with one of the input text-region boxes, so the final HTML is assembled during generation rather than after a heuristic alignment stage. With this mechanism and an added span-aware contrastive loss, TFLOP reports state-of-the-art TEDS on PubTabNet, FinTabNet, and SynthTabNet, with the largest gains on complex rowspan/colspan tables. A sympathetic reader should care because the approach removes a fragile post-processing step and, according to the paper's experiments, carries over to watermarked and non-English industrial tables.","feed_headline":"Layout pointer skips box matching, setting new table-structure records","feed_subtitle":"Direct box-to-tag pointing lifts table-structure accuracy on all three benchmark sets.","key_machinery":"The load-bearing mechanism is the layout pointer: the decoder's final hidden states are split into box features $\\{\\bar{b}_j\\}$ and tag features $\\{\\bar{t}_k\\}$, and the pointer loss $$L_{ptr} = -\\frac{1}{B}\\sum_{j=1}^{B} \\log \\frac{\\exp(\\bar{b}_j\\cdot \\bar{t}_{k^*}/\\tau)}{\\sum_{k'\\in D}\\exp(\\bar{b}_j\\cdot \\bar{t}_{k'}/\\tau)}$$ forces each text box to point to its corresponding data tag, with a second binary-cross-entropy term using a learned empty-box embedding $\\bar{b}_0$ for tags that have no text. The supporting mechanism is span-aware contrastive supervision, which treats boxes in the same row or column as positives and weights each positive by the span coefficient $c_p(j)=\\mathrm{overlap}(p,j)^2/(\\mathrm{span}(p)\\,\\mathrm{span}(j))$, so cells sharing only part of a row or column are still aligned according to their true coverage. These losses are combined with the tag-classification loss, and the resulting box embeddings form row-wise and column-wise clusters in the ablations, which is what lets the pointer work on tables with complex spans.","core_discovery":"The central claim is that a sequence decoder can serve as both the logical structure predictor and the alignment module. After encoding the table image and a set of text-region bounding boxes, TFLOP auto-regressively generates compact tag sequences in OTSL notation and, from the decoder's last hidden states, computes a dot-product similarity between each box feature and each table-data-tag feature; a pointer loss trains each box to point to its correct tag. Empty cells are handled by a special embedding, and a span-aware contrastive loss organizes box embeddings by row and column overlap so cells participating in rowspans or colspans are pulled together in proportion to their shared span. The paper reports that this design reaches the highest TEDS scores among compared methods on all three benchmarks and interprets the small TEDS-versus-TEDS-Struct gap as evidence that text misalignment errors, which dual-decoder baselines suffer, are largely eliminated.","pith_inferences":["An implication the authors leave implicit is that the alignment burden is shifted, not removed: the pointer is only as good as the input text-region boxes, so the framework relocates the failure mode of dual decoders into the OCR or annotation stage.","The pointer formulation is not tied to HTML; because the decoder emits OTSL tokens with a one-to-one HTML mapping, the same layout-pointer loss could be trained toward LaTeX or a graph representation of the table.","The cross-lingual result is suggestive but rests on only 30 self-annotated Korean tables; a systematic multilingual dataset would be needed to establish that the pointer mechanism, rather than the specific encoder-decoder weights, is what transfers.","The small QA experiment links TSR quality to downstream table QA, which implies that pointer-based alignment could be worth testing in end-to-end document question-answering systems that currently rely on separate structure and text extraction."],"forward_implications":["On the reported benchmarks, the full TFLOP model reaches 96.66 TEDS on PubTabNet test, 99.45 on FinTabNet, and 99.40 on SynthTabNet, all above the compared dual-decoder baselines.","Because the TEDS-versus-TEDS-Struct gap is much smaller for TFLOP than for prior methods on FinTabNet, most remaining errors are structural tag errors rather than text-placement errors.","The ablation shows span-aware contrastive supervision helps complex tables more than simple tables, implying rowspan/colspan recognition is the part of the problem this addition is solving.","The watermark experiment suggests that unwanted text can be handled by filtering input bounding boxes with a small classifier before layout pointing, without changing the decoder or the pointer."],"supporting_citations":[{"why":"Introduces the TableFormer dual-decoder baseline and the SynthTabNet benchmark that TFLOP must beat.","marker":"[Nassar et al., 2022]"},{"why":"The VAST dual-decoder method whose bounding-box matching stage TFLOP removes and whose results are compared on PubTabNet and FinTabNet.","marker":"[Huang et al., 2023]"},{"why":"TableMaster dual-decoder pipeline used as an OCR-based PubTabNet test baseline and as a comparison on Korean tables.","marker":"[Ye et al., 2021]"},{"why":"DRCC semi-autoregressive baseline for FinTabNet comparisons and error-accumulation context.","marker":"[Shen et al., 2023]"},{"why":"Supplies the supervised contrastive loss that span-aware contrastive supervision extends.","marker":"[Khosla et al., 2020]"},{"why":"Defines OTSL tokenization, the compact tag sequence with 1-to-1 HTML mapping used by the decoder.","marker":"[Lysak et al., 2023]"},{"why":"Donut architecture whose Swin encoder and BART decoder configuration TFLOP adopts.","marker":"[Kim et al., 2022]"},{"why":"Provides PubTabNet and the TEDS evaluation metric used in all benchmarks.","marker":"[Zhong et al., 2020]"},{"why":"PSENet text detection used to obtain PubTabNet test text-region boxes.","marker":"[Wang et al., 2019a]"},{"why":"Master OCR recognition used with PSENet for PubTabNet test text extraction.","marker":"[Lu et al., 2021]"}],"fun_headline_variants":["Direct pointer skips box alignment, tops three table benchmarks","TFLOP pointer mechanism removes text-region matching, sets SOTA","Span-aware contrastive pointer lifts table structure parsing to top","Pointer loss directly links text boxes to tags, no matching needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the input text-region bounding boxes, whether from cell annotations or an off-the-shelf OCR engine, are complete and accurate enough to serve as pointing targets; if OCR misses, splits, or merges text regions (or leaves watermark boxes), the pointer can only associate the tags it is given with those imperfect boxes.","fun_headline_variants_meta":{"raw":{"variants":["Direct pointer skips box alignment, tops three table benchmarks","TFLOP pointer mechanism removes text-region matching, sets SOTA","Span-aware contrastive pointer lifts table structure parsing to top","Pointer loss directly links text boxes to tags, no matching needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2853,"prompt_tokens":964,"completion_tokens":1889,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1818}},"tokens_in":580,"tokens_out":1889,"duration_ms":13318,"temperature":1.0,"reasoning_tokens":1818,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:50:59.419523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TFLOP on a fixed set of table images under two OCR conditions: once with cell-level text regions and once with those regions deliberately corrupted by merging two adjacent cells' boxes into one and by deleting one cell's box. If TEDS stays unchanged, the central claim that the pointer removes misalignment would be undercut; if every corrupted-box table loses a measurable amount of TEDS, the framework's advantage over dual decoders would depend on input-box quality rather than on the pointer itself.","supporting_citations":[{"cited_title":"Image-based table recognition: data, model, and evaluation","cited_arxiv_id":null,"evidence_quote":"Provides PubTabNet and the TEDS evaluation metric used in all benchmarks."},{"cited_title":"Tableformer: Table struc- ture understanding with transformers","cited_arxiv_id":null,"evidence_quote":"Introduces the TableFormer dual-decoder baseline and the SynthTabNet benchmark that TFLOP must beat."},{"cited_title":"Improving table structure recog- nition with visual-alignment sequential coordinate mod- eling","cited_arxiv_id":null,"evidence_quote":"The VAST dual-decoder method whose bounding-box matching stage TFLOP removes and whose results are compared on PubTabNet and FinTabNet."},{"cited_title":"Divide rows and conquer cells: Towards structure recognition for large tables","cited_arxiv_id":null,"evidence_quote":"DRCC semi-autoregressive baseline for FinTabNet comparisons and error-accumulation context."},{"cited_title":"Supervised con- trastive learning","cited_arxiv_id":null,"evidence_quote":"Supplies the supervised contrastive loss that span-aware contrastive supervision extends."},{"cited_title":"Optimized Table Tokenization for Table Structure Recognition","cited_arxiv_id":"2305.03393","evidence_quote":"Defines OTSL tokenization, the compact tag sequence with 1-to-1 HTML mapping used by the decoder."},{"cited_title":"Ocr-free document understanding trans- former","cited_arxiv_id":null,"evidence_quote":"Donut architecture whose Swin encoder and BART decoder configuration TFLOP adopts."}],"review_version":1}