{"id":"996ffa91-e5ef-4ee5-a414-a04d8b2a085e","arxiv_id":"2607.00747","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"First adaptation of Uncertainty Herding to cascaded TD-TSR pipelines via RankFusion and CAPA extensions yields consistent gains over baselines on four table extraction datasets under annotation budgets of 71-500 documents.","lead":"The paper adapts Uncertainty Herding, a coverage-uncertainty active learning method, to cascaded table detection and structure recognition pipelines with two new extensions. A smart generalist might read it to learn practical ways to cut annotation costs when training document AI models on limited budgets.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Unverified propagation of TD localization errors into CAPA's TSR-stage sampling and calibration","rationale":"The reader's weakest assumption directly identifies the missing isolation of error propagation; the full-text description does not appear to close this gap with a dedicated check, so the central empirical claim remains conditional on that untested premise.","tokens_in":1780,"tokens_out":283,"duration_ms":22810,"concrete_test":"Re-run the four-dataset active-learning protocol with controlled TD noise (inject 15% spurious table boxes at each round while keeping TSR labels fixed); if CAPA's win rate over UHerding drops below 3/4 or its variance increases beyond the original runs, the dependency-exploitation premise does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline claim that CAPA is the most consistent pipeline-aware method (outperforming UHerding on 3/4 datasets) rests on the premise that stage-dependent gating plus dual-manifold coverage safely exploits the TD-to-TSR link. If TD false positives/negatives alter the input distribution to TSR, the per-task uncertainty estimates and gating decisions can become biased; nothing in the reported experiments isolates whether the observed consistency survives such propagation. The abstract states the extensions 'exploit the TD-to-TSR dependency' but supplies no ablation or sensitivity analysis on this point.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper adapts Uncertainty Herding (UHerding) to cascaded table extraction pipelines consisting of Table Detection (TD) followed by Table Structure Recognition (TSR). It introduces two pipeline-aware extensions—RankFusion (dual-manifold coverage) and CAPA (stage-dependent gating plus per-task uncertainty calibration)—that exploit the TD-to-TSR dependency. Experiments on two public and two private datasets across annotation budgets of 71–500 documents show that UHerding generalizes well, while CAPA is the most consistent pipeline-aware variant, outperforming standard UHerding on three of four datasets.","tokens_in":1887,"tokens_out":485,"duration_ms":14654,"significance":"If the empirical results hold under proper statistical controls, the work supplies the first explicit treatment of inter-stage dependencies in active learning for cascaded object-detection pipelines. This is practically relevant for document-analysis tasks where TSR annotation is far more expensive than TD annotation. The hybrid coverage-uncertainty framing and the concrete extensions (RankFusion, CAPA) are reusable beyond tables.","major_comments":[{"comment":"Abstract: the headline claim that 'CAPA emerges as the most consistent strategy, outperforming standard UHerding on three out of four datasets' is presented without error bars, results across random seeds, or any statistical test. Because the text already notes greater variance for RankFusion, the absence of these controls makes the 'most consistent' ranking unverifiable and load-bearing for the central empirical conclusion.","section":"Abstract"},{"comment":"Abstract / Methods (implied): the premise that stage-dependent gating and dual-manifold coverage 'exploit the TD-to-TSR dependency' without introducing new selection bias is not accompanied by any ablation or sensitivity analysis on TD localization errors propagating into TSR-stage sampling and uncertainty estimates. This is the exact mechanism the skeptic note flags and is required to substantiate the pipeline-aware advantage.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract states results 'across two public (PubTables-1M and FinTabNet) and two private table extraction datasets' but does not name the private datasets or give their characteristics; this should be supplied for reproducibility.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their thoughtful comments, which help strengthen the presentation of our empirical results. We address each major comment below.","responses":[{"response":"We acknowledge the need for statistical rigor in supporting the claim. The manuscript will be revised to include results from multiple random seeds with error bars and appropriate statistical tests (such as Wilcoxon signed-rank tests) to verify the consistency of CAPA's performance across datasets.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the headline claim that 'CAPA emerges as the most consistent strategy, outperforming standard UHerding on three out of four datasets' is presented without error bars, results across random seeds, or any statistical test. Because the text already notes greater variance for RankFusion, the absence of these controls makes the 'most consistent' ranking unverifiable and load-bearing for the central empirical conclusion."},{"response":"Our experiments utilize real datasets where TD predictions contain localization errors that propagate to TSR, and the performance improvements demonstrate the effectiveness of the pipeline-aware methods in this realistic setting. To directly address concerns about selection bias, we will add an ablation study in the revised manuscript that simulates controlled TD error rates and analyzes their effect on TSR sampling and overall gains.","revision_made":"yes","referee_comment":"[Abstract] Abstract / Methods (implied): the premise that stage-dependent gating and dual-manifold coverage 'exploit the TD-to-TSR dependency' without introducing new selection bias is not accompanied by any ablation or sensitivity analysis on TD localization errors propagating into TSR-stage sampling and uncertainty estimates. This is the exact mechanism the skeptic note flags and is required to substantiate the pipeline-aware advantage."}],"tokens_in":1447,"tokens_out":373,"duration_ms":29155,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The new part is the first reported use of Uncertainty Herding on a cascaded object detection task for tables, plus two named extensions: RankFusion for dual-manifold coverage and CAPA for stage-dependent gating plus per-task calibration. The experiments run across four datasets and several annotation budgets, and the abstract states that CAPA beats standard UHerding on three of them while staying more consistent than RankFusion.\n\nThe practical angle is useful. Table extraction pipelines are common in document work, and an active-learning method that tries to respect the two-stage structure is a reasonable direction. Running on both public and private data also helps.\n\nThe main weakness is that the central premise—exploiting the TD-to-TSR dependency without new error propagation—rests on the abstract claim alone. Nothing isolates whether false positives or negatives from the detection stage bias the structure-recognition sampling or the uncertainty estimates. No ablations on that point, no error bars, and the “three out of four” dataset selection cannot be checked without more detail on methods and seeds. Those gaps make the consistency claim hard to evaluate.\n\nThis is for readers already working on active learning inside document AI or cascaded detectors. It is narrow enough that most people outside that niche will not need it, but the empirical setup is concrete enough to merit referee time once the propagation and variance issues are addressed.","headline":"The paper adapts UHerding to cascaded table extraction with two pipeline extensions and reports empirical gains, but leaves the core claim about exploiting TD-to-TSR dependencies untested.","tokens_in":2402,"tokens_out":357,"would_cite":false,"duration_ms":13246,"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":"Pipeline-aware active learning strategies outperform standard methods for cascaded table extraction.","keywords":["active learning","table extraction","cascaded pipelines","table detection","table structure recognition","uncertainty herding","object detection"],"falsifier":"A controlled test on a dataset with high table detection error rates in which CAPA no longer outperforms standard UHerding or produces lower end-to-end table extraction accuracy.","tokens_in":2683,"feed_emoji":"","tokens_out":654,"duration_ms":32048,"temperature":0.7,"pith_summary":"This paper adapts Uncertainty Herding to the cascaded pipeline of table detection followed by table structure recognition. Two extensions are introduced: RankFusion applies dual-manifold coverage over detection and structure spaces, while CAPA adds stage-dependent gating and per-task uncertainty calibration to use the dependency between stages. Experiments on four datasets with annotation budgets from 71 to 500 documents show the adapted methods reduce annotation needs and improve results over baselines. CAPA proves the most consistent, beating standard Uncertainty Herding on three of the four datasets.","feed_headline":"Cascaded sampling cuts labels needed for table extraction","feed_subtitle":"Stage-aware extensions to uncertainty herding use detection-to-structure links to select informative documents across annotation budgets.","key_machinery":"Pipeline-aware extensions to Uncertainty Herding (RankFusion with dual-manifold coverage and CAPA with stage-dependent gating) that exploit the TD-to-TSR dependency for sampling in cascaded table extraction.","core_discovery":"The first adaptation of Uncertainty Herding to cascaded object detection pipelines is presented, with two pipeline-aware extensions that exploit the TD-to-TSR dependency. RankFusion adds dual-manifold coverage over both detection and structure representation spaces, while CAPA further incorporates stage-dependent gating and per-task uncertainty calibration. Experiments across two public and two private datasets show that UHerding generalizes well to table extraction and outperforms each baseline, while among the pipeline-aware variants CAPA is the most consistent and outperforms standard UHerding on three out of four datasets.","pith_inferences":["The stage-dependent approach may transfer to other cascaded vision pipelines where an early localization step informs a later recognition step.","CAPA's consistency could make it suitable for production document systems that prioritize stable performance across varying document collections.","Explicit checks for how detection-stage errors influence structure-stage sampling could identify further refinements to the gating mechanism."],"forward_implications":["UHerding generalizes well to table extraction, outperforming each baseline.","RankFusion achieves higher expected gains but at the cost of greater variance.","CAPA emerges as the most consistent strategy, outperforming standard UHerding on three out of four datasets.","The methods remain effective across annotation budgets from 71 to 500 documents on both public and private datasets."],"fun_headline_variants":["Uncertainty herding adapted to cascaded table extraction","Pipeline extensions balance coverage in table detection","CAPA shows consistent gains across table datasets","Rankfusion adds dual manifold coverage to active learning"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Exploiting the TD-to-TSR dependency via dual-manifold coverage and stage-dependent gating does not introduce new error propagation or selection biases that negate the reported gains.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty herding adapted to cascaded table extraction","Pipeline extensions balance coverage in table detection","CAPA shows consistent gains across table datasets","Rankfusion adds dual manifold coverage to active learning"]},"model":"grok-4.3","cost_usd":0.005447,"raw_usage":{"total_tokens":2568,"prompt_tokens":725,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":54465500,"prompt_tokens_details":{"text_tokens":725,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1788,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":725,"tokens_out":55,"duration_ms":13239,"temperature":1.0,"reasoning_tokens":1788,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-02T14:24:14.189022+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled test on a dataset with high table detection error rates in which CAPA no longer outperforms standard UHerding or produces lower end-to-end table extraction accuracy.","supporting_citations":[],"review_version":1}