{"id":"95e9d44b-1660-42b8-b5e2-33d48d07ceeb","arxiv_id":"2606.18721","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Geometry-Aware Pointer Loss reweights cross-entropy by inverse Manhattan distance to focus gradients on adjacent-cell errors in TSR, yielding SOTA results on PubTabNet and SynthTabNet.","lead":"The paper shows that pointer networks for table structure recognition mostly fail on spatially adjacent cells and introduces a modified loss that weights those errors more heavily using inverse distance. A smart generalist might read it to see how a small change to training objectives can embed geometric knowledge without altering model architecture or inference cost.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Representativeness of the 79.6% adjacent-cell error statistic (Manhattan <=2) across table layouts is untested","rationale":"The reader's weakest_assumption directly identifies the same load-bearing point; the abstract provides no counter-evidence or cross-domain validation that would overturn it.","tokens_in":1706,"tokens_out":279,"duration_ms":13166,"concrete_test":"Apply the identical GAP loss formulation to a held-out TSR benchmark with structurally different tables (e.g., WTW or a custom set of dense scientific tables); recompute both overall F1 and the adjacent-error breakdown; if the adjacent-error reduction disappears or overall performance drops relative to the unweighted baseline, the inductive bias does not generalize.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the observed failure mode (79.6% errors between cells with Manhattan distance <=2 on the evaluated sets) is the dominant general failure mode for pointer-based TSR, so that inverse-distance reweighting of negatives in the pointer cross-entropy will reliably focus gradients without introducing new biases on other layouts. If the statistic is dataset-specific to PubTabNet/SynthTabNet, the reweighting could over-penalize local confusions while under-penalizing long-range pointer errors in tables with different topology or density.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper analyzes failures in pointer-network-based table structure recognition (TSR), reporting that 79.6% of errors occur between cells with Manhattan distance ≤2. It proposes Geometry-Aware Pointer (GAP) Loss, a modification to cross-entropy that applies inverse-distance reweighting to negative candidates so that spatially closer cells receive higher loss weight. The method requires no architecture changes or inference overhead. Experiments on PubTabNet and SynthTabNet show reduced adjacent-cell errors and new state-of-the-art results; code is released.","tokens_in":1817,"tokens_out":539,"duration_ms":25759,"significance":"If the empirical gains hold under the proposed reweighting, the work shows that a lightweight geometric bias injected only at the loss level can measurably improve pointer-based TSR without altering model capacity or runtime. The public code release is a concrete strength that aids reproducibility. The significance is limited by the extent to which the motivating error distribution generalizes beyond the two evaluated datasets.","major_comments":[{"comment":"Error analysis (Introduction and §3): the 79.6% adjacent-cell failure rate is presented as the dominant, general failure mode that justifies inverse-distance reweighting. No additional datasets, table-density variants, or cross-domain splits are used to test whether this statistic is representative; if the distribution is PubTabNet/SynthTabNet-specific, the reweighting may over-penalize local confusions while under-penalizing long-range pointer errors on other topologies.","section":"Introduction / §3"},{"comment":"§4 (Experiments) and loss definition: the manuscript reports consistent gains and SOTA numbers but supplies neither per-run standard deviations, statistical significance tests, nor an ablation isolating the exact inverse-distance formula (including distance metric, zero-distance handling, and normalization). These omissions make it impossible to verify that the reported improvements are robustly attributable to the geometric reweighting rather than other training factors.","section":"§4"}],"minor_comments":[{"comment":"The precise mathematical definition of the GAP loss (reweighting term, handling of the ground-truth cell itself) should be stated as an explicit equation rather than described only in prose.","section":"§3.2"},{"comment":"Figure captions and table footnotes should explicitly state whether the reported metrics are on the official test splits and whether any post-processing differs from the baselines.","section":"Tables 1-2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We respond to each major comment below, indicating planned revisions where appropriate.","responses":[{"response":"The reported 79.6% statistic is derived directly from error cases on PubTabNet and SynthTabNet, the standard benchmarks used throughout the TSR literature and on which GAP Loss is evaluated. The geometric reweighting is motivated by and validated against the failure modes observed in these datasets. While we agree that analysis on further datasets or topologies would help establish broader generality, the current results show consistent error reduction and SOTA performance on the evaluated benchmarks. We will add a short discussion of this scope limitation in the revised manuscript.","revision_made":"partial","referee_comment":"[Introduction / §3] Error analysis (Introduction and §3): the 79.6% adjacent-cell failure rate is presented as the dominant, general failure mode that justifies inverse-distance reweighting. No additional datasets, table-density variants, or cross-domain splits are used to test whether this statistic is representative; if the distribution is PubTabNet/SynthTabNet-specific, the reweighting may over-penalize local confusions while under-penalizing long-range pointer errors on other topologies."},{"response":"We agree that additional statistical reporting and targeted ablations would improve verifiability. In the revised manuscript we will report mean performance and standard deviations over multiple independent runs with different random seeds, include statistical significance tests, and add an ablation isolating the inverse-distance formulation (distance metric, zero-distance handling, and normalization).","revision_made":"yes","referee_comment":"[§4] §4 (Experiments) and loss definition: the manuscript reports consistent gains and SOTA numbers but supplies neither per-run standard deviations, statistical significance tests, nor an ablation isolating the exact inverse-distance formula (including distance metric, zero-distance handling, and normalization). These omissions make it impossible to verify that the reported improvements are robustly attributable to the geometric reweighting rather than other training factors."}],"tokens_in":1410,"tokens_out":443,"duration_ms":23418,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core idea is straightforward. The authors observe that 79.6% of pointer errors fall between cells with Manhattan distance at most 2, then modify the loss so negatives closer to the ground-truth cell receive higher weight via inverse distance. This keeps the model and inference unchanged while shifting gradient focus to the reported failure mode.\n\nThe work does what it sets out to do on the given benchmarks. It supplies a clear error breakdown that motivates the change, shows consistent improvement on PubTabNet and SynthTabNet, and stays within the existing pointer-network setup. The modification itself is a direct, non-circular extension of cross-entropy, which makes the claim easy to inspect.\n\nThe soft spot is the representativeness of that 79.6% figure. If the adjacent-error pattern is particular to the two evaluated datasets rather than a general property of pointer-based TSR, the reweighting could over-penalize local confusions on tables with different density or topology. The abstract gives no ablations on alternative layouts or significance numbers, so the full paper needs to demonstrate that the gains hold without introducing new long-range errors.\n\nCitation patterns look typical for the subfield. The math and data handling appear solid on the surface, with no free parameters or self-referential fitting.\n\nThis paper is for people already working on document table parsing or pointer networks for structured output. A reader outside that niche will find little to take away. It deserves a serious referee because the idea is concrete, the experiments use accepted benchmarks, and the central assumption can be tested directly from the reported results.","headline":"The paper adds an inverse-distance reweighting to the pointer cross-entropy loss that targets adjacent-cell mistakes in table structure recognition and reports gains on two standard datasets.","tokens_in":2276,"tokens_out":400,"would_cite":false,"duration_ms":19989,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Reweighting cross-entropy loss by inverse cell distance reduces adjacent errors in table structure recognition.","keywords":["table structure recognition","pointer network","geometry-aware loss","inverse distance weighting","spatial locality","adjacent cell errors","tsr"],"falsifier":"Training a pointer network on a table dataset where most errors occur between cells farther apart and observing that the geometry-aware loss performs worse than standard cross-entropy.","tokens_in":2604,"feed_emoji":"📐","tokens_out":579,"duration_ms":24925,"temperature":0.7,"pith_summary":"Pointer networks for table structure recognition fail most often by mixing up cells that sit next to each other. Standard cross-entropy loss treats every wrong prediction the same regardless of location. The Geometry-Aware Pointer loss changes the objective so that mistakes between nearby cells receive larger gradient updates through inverse distance weighting. This adjustment requires only a change in the loss calculation and leaves the network and inference unchanged. Experiments on two standard table datasets show fewer local errors and higher overall accuracy.","feed_headline":"Distance-weighted loss cuts table recognition errors","feed_subtitle":"Pointer networks make fewer mistakes on nearby cells when training emphasizes close errors over all errors equally.","key_machinery":"Geometry-Aware Pointer Loss applies inverse distance weighting to negative candidates in the cross-entropy objective so that closer cells exert stronger influence on the gradient.","core_discovery":"The paper establishes that reweighting the pointer network's cross-entropy loss according to the Manhattan distance between predicted and ground-truth cells directs stronger gradients toward spatially adjacent mistakes, which account for 79.6 percent of errors, and that this yields improved recognition accuracy on PubTabNet and SynthTabNet.","pith_inferences":["This reweighting could be adapted to other pointer-based prediction tasks that involve spatial layouts.","Future work might examine whether the same weighting helps when tables have irregular layouts not seen in the training data.","Combining this loss with other geometric features in the model architecture might produce further gains.","Testing on tables from different domains like financial reports or scientific papers would check if the error pattern persists."],"forward_implications":["Adjacent-cell errors decrease because gradients emphasize local mistakes over distant ones.","State-of-the-art results are reached on the evaluated table datasets.","The same model architecture works with only a change in the loss calculation.","No additional cost occurs during inference.","Geometric information incorporated at the loss level aids robustness in table structure tasks."],"fun_headline_variants":["Geometry aware loss cuts adjacent table errors","Distance reweighting cuts pointer network errors","GAP loss targets spatial mistakes in TSR","Proximity weighted loss reduces table structure errors","Inverse distance loss directs gradients to neighbors"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The pattern that most errors occur between cells within a Manhattan distance of two is typical for table structure recognition in general.","fun_headline_variants_meta":{"raw":{"variants":["Geometry aware loss cuts adjacent table errors","Distance reweighting cuts pointer network errors","GAP loss targets spatial mistakes in TSR","Proximity weighted loss reduces table structure errors","Inverse distance loss directs gradients to neighbors"]},"model":"grok-4.3","cost_usd":0.008472,"raw_usage":{"total_tokens":3810,"prompt_tokens":627,"num_sources_used":0,"completion_tokens":53,"cost_in_usd_ticks":84724500,"prompt_tokens_details":{"text_tokens":627,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3130,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":627,"tokens_out":53,"duration_ms":20730,"temperature":1.0,"reasoning_tokens":3130,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T21:24:36.157161+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Training a pointer network on a table dataset where most errors occur between cells farther apart and observing that the geometry-aware loss performs worse than standard cross-entropy.","supporting_citations":[],"review_version":1}