acceptodds
Under review as a conference paper at ICLR 2027

When Tables Need Context: Query-Conditioned Table Extraction

Abstract

Most document table extraction systems follow an extract-all paradigm: they reconstruct every table in a document before downstream retrieval or filtering. However, real-world document interaction is often query-driven, where users seek only the tables relevant to a specific information need. We introduce Query-Conditioned Table Extraction (QCTE), a new task that identifies tables satisfying a natural-language condition and reconstructs only the matched tables as structured HTML. We argue that query-conditioned extraction can benefit from document-level context beyond isolated table crops, since captions and surrounding paragraphs provide useful evidence for reliable condition judgment. To this end, we propose PCRC, a framework that combines table-centered context retrieval, progressive evidence incorporation, and reusable KV-cache inference. PCRC progressively introduces contextual evidence only when necessary while sharing visual-textual representations across table selection and reconstruction, substantially reducing redundant multimodal computation. To support this setting, we construct three condition-grounded benchmarks spanning financial, scientific, and visually degraded documents. Across all three datasets, PCRC improves judgment F1 from about 0.73 to 0.90 and composite F1-TEDS by 13-16 points. It requires only 24-29% of the prefill computation of the same inference schedule without cache reuse. These results show that selective document context improves query-conditioned table extraction and can be used efficiently with cache reuse.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Query-conditioned table extraction, Document understanding, Vision-language models, KV-cache reuse, Efficient inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DkKROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces query-conditioned table selection and reconstruction, together with PCRC: paragraph retrieval, progressive evidence incorporation, and KV-cache reuse. Three benchmarks cover financial, scientific, and synthetic degraded documents. Without task-specific finetuning, relevant context improves selection and HTML reconstruction while caching reduces repeated prefill.

Critical strengths
  1. The problem is practical and goes beyond table retrieval. The output includes reconstructed HTML from rendered documents, not merely relevance decisions. Visual reconstruction and document context provide a clear extension to natural-language-conditioned table discovery.

  2. The empirical investigation is substantial. Composite F1 improves from Qwen's 0.597/0.601/0.644 to 0.723/0.756/0.794 (Table 3). Ablations cover random/all/semantic/spatial context, shared/exclusive evidence, stopping, single-judgment inference with identical selected evidence, and another backbone. The strong generic-prefix-cache control in Table 14 is particularly useful.

  3. Annotation and evaluation boundaries are documented carefully. A roughly 5% condition audit covers 217,229 condition–table pairs, with blinded checking, double annotation, and adjudication; gains persist on human-checked labels (§A.1.6, Tables 8 and 12). The paper explicitly reports candidate-conditional quality and identical quality with/without cache reuse.

Critical weaknesses
  1. End-to-end robustness needs stronger evidence. Table 9 excludes unmatched source tables from recall, leaving missed detections unpenalized. Incorrect early rejection also ends evidence collection and prevents extraction (Algorithm 1, lines 3–7). Table 14’s “No reliability gating” control reports only aggregate scores, without measuring erroneous early stops or how often later evidence could correct them. The overall robustness claim in §A.4.9 (lines 1720–1723) therefore needs targeted validation.

  2. The composite metric can reward missing difficult tables. In §A.3 (lines 1251–1258), F1-TEDS = F1 × EQ uses cell-weighted EQ only over true positives. With two equally sized positives, TEDS 1.0 and 0.1, and no false positives, selecting both scores 0.55; missing the second scores 0.667. Higher scores need not mean more complete extraction. Add a metric with a fixed denominator over all positives, zero credit for misses, and a false-positive penalty. This counterexample does not establish incorrect method rankings.

  3. Evidence for gains over a simpler baseline remains limited. With identical evidence, single judgment plus generic prefix caching nearly matches PCRC’s quality (Table 14); generic caching also recovers most prefill savings under the same schedule (§A.4.5). Across 20 documents per dataset, PCRC is only about 3–5% faster, excluding detection, OCR, and NV-Embed indexing (Table 16). Context selection has clearer support; assessing progressive inference and custom caching requires total costs and timing variability. The 3.4–4.1-fold prefill saving over uncached execution does not imply equivalent acceleration over the simpler baseline.

Questions for authors
  1. Does PCRC’s robustness advantage survive counting upstream misses? Report erroneous early-stop rates and the fraction correctable by later evidence (Table 9, Algorithm 1, Table 14).

  2. Do the rankings and gains in Table 3 persist with a fixed denominator over all positives, zero credit for misses, and a false-positive penalty?

  3. After including or amortizing detection, OCR, and NV-Embed indexing, how much speedup remains over single judgment with identical evidence and generic prefix caching? Report paired time differences and their variability (Tables 14 and 16).

9 comments

We thank the reviewer for recognizing the importance of our research. We address the concerns below.

1. Evidence of end-to-end robustness

End-to-end robustness needs stronger evidence. Table 9 excludes unmatched source tables from recall, leaving missed detections unpenalized.

To address this concern, we conducted condition judgment and HTML reconstruction using table regions cropped directly from the datasets' annotated bounding boxes, rather than detector-generated proposals. This eliminates the influence of upstream table detection errors, allowing us to assess whether PCRC's quality and token-saving advantages persist when ground-truth table regions are provided.

  • Input construction: all component regions and text are obtained using the datasets' source annotations, PyMuPDF, PP-Structure, and the original OCR procedure for PUB. DEG retains the original masking of characters occluded by watermarks, stains, and stamps.
  • Controlled comparison: all compared methods receive the same annotated table regions, paragraph blocks, and text inputs. Reference HTML remains the reconstruction target for evaluation.

Table R3a. Composite quality with source-annotated boxes (main-paper Table 3 layout). TEDS is the extraction metric. FIN, PUB, and DEG denote FINCONTAB, PUBCONTAB, and DEGCONTAB.

MethodFIN PcP_{c}FIN RcR_{c}FIN F1F_{1}PUB PcP_{c}PUB RcR_{c}PUB F1F_{1}DEG PcP_{c}DEG RcR_{c}DEG F1F_{1}
Qwen+TATR0.2850.3610.3190.5900.7600.6640.3690.4050.386
Qwen+MinerU0.4650.5890.5200.4720.6080.5310.5330.5850.558
TATR+TAPAS0.0210.3290.0390.5790.7490.6530.0240.3790.045
TATR+Qwen0.2920.2860.2890.6400.7910.7080.3280.3310.329
MinerU+Qwen0.4130.5250.4620.4430.6000.5100.4690.5500.506
MinerU+TAPAS0.0430.3630.0770.3480.5140.4150.0330.4870.062
Qwen0.5400.6840.6040.5310.6840.5980.6150.6750.644
PCRC (ours)0.7060.7200.7130.7570.7440.7500.8220.8000.811
PCRC w/o cache0.7060.7200.7130.7570.7440.7500.8220.8000.811
  • Quality remains close to the original results. With annotated boxes, PCRC achieves composite F1 of 0.713 / 0.750 / 0.811 on FIN / PUB / DEG, compared with 0.723 / 0.756 / 0.794 in the original Table 3. The changes are only -1.0 / -0.6 / +1.7 percentage points.
  • PCRC retains its quality lead on all three datasets. It achieves the highest composite F1, tied with its no-cache variant. The strongest non-PCRC baselines score 0.604 / 0.708 / 0.644, leaving PCRC ahead by 10.9 / 4.2 / 16.7 percentage points.

Table R3b. Per-condition VLM token consumption with source-annotated boxes (main-paper Table 4 layout). Pre., Dec., and Tot. are prefill, decode, and total tokens in thousands; R.R. is the recomputation ratio in percent. FIN, PUB, and DEG denote FINCONTAB, PUBCONTAB, and DEGCONTAB.

MethodFIN Pre.FIN Dec.FIN Tot.FIN R.R.PUB Pre.PUB Dec.PUB Tot.PUB R.R.DEG Pre.DEG Dec.DEG Tot.DEG R.R.
Qwen30.561.2331.79—10.061.6511.71—15.030.7415.77—
PCRC35.173.7038.8727.1513.733.1516.8827.9020.263.8324.0923.14
PCRC w/o cache129.563.70133.26100.049.223.1552.37100.087.563.8391.39100.0
PCRC schedule + generic prefix cache36.063.7039.7627.8314.123.1517.2728.6920.503.8324.3323.41
  • The token-saving advantage also remains. PCRC uses 35.17 / 13.73 / 20.26k prefill tokens per condition, compared with 129.56 / 49.22 / 87.56k for the same PCRC schedule without caching, with identical quality. The prefill savings are 72.85% / 72.10% / 76.86%, compared with 71.71% / 70.98% / 75.65% in the original Table 4 (computed as 100% minus the reported R.R.).

PCRC therefore retains both its quality lead and substantial cache savings when source-annotated boxes replace detector proposals. These advantages are not confined to the detector-generated candidate set used in the original experiments.

2. Fine-grained comparison of No reliability gating and PCRC

Table 14’s “No reliability gating” control reports only aggregate scores, without measuring erroneous early stops or how often later evidence could correct them.

We want to clarify that the original Table 14 already reports judgment F1, composite F1, prefill consumption, and R.R. for both variants. To show how continued evidence collection affects precision and recall, Table R3c adds the previously unreported judgment P / R and composite Pc / Rc for No reliability gating (marked †), alongside the corresponding PCRC scores from the original Tables 11 and 3.

Both variants use the same inputs, selected evidence pool, and cache-reuse policy under the original Table 14 protocol. No reliability gating continues through all selected exclusive paragraphs regardless of an earlier definite judgment.

Table R3c. Fine-grained comparison of PCRC and No reliability gating (No gating). † marks newly reported values; all other entries are reproduced from the original manuscript. P, R, and F1 are judgment-only micro scores; Pc, Rc, and F1c are composite scores. Quality scores are displayed to three decimal places; differences below are calculated from the displayed values. Prefill is measured per condition in thousands of tokens.

MetricFIN: PCRCFIN: No gatingPUB: PCRCPUB: No gatingDEG: PCRCDEG: No gating
PP0.8880.889†0.8840.898†0.8990.891†
RR0.9140.925†0.9190.935†0.9060.905†
F1F_{1}0.9010.9070.9010.9160.9020.898
PcP_{c}0.7130.716†0.7410.744†0.7910.780†
RcR_{c}0.7340.745†0.7710.775†0.7970.792†
F1cF_{1c} (F1-TEDS)0.7230.7310.7560.7590.7940.786
Prefill tokens (k)32.9540.1913.4317.4519.5723.96

The following changes are No reliability gating minus PCRC, with quality differences expressed in percentage points:

  • FIN: additional evidence gives a small benefit, mainly in recall. Judgment P / R / F1 increase by 0.1 / 1.1 / 0.6 points, and composite Pc / Rc / F1c increase by 0.3 / 1.1 / 0.8 points. Continuing evidence collection therefore improves aggregate coverage, with a 0.8-point gain in composite F1.
  • PUB: judgment improves more than the joint judgment-and-reconstruction score. Judgment P / R / F1 increase by 1.4 / 1.6 / 1.5 points, while composite Pc / Rc / F1c increase by only 0.3 / 0.4 / 0.3 points. The additional evidence improves final judgment, but its net benefit to composite F1 is only 0.3 points.
  • DEG: consuming all remaining evidence lowers all six quality scores. Judgment P / R / F1 decrease by 0.8 / 0.1 / 0.4 points, and composite Pc / Rc / F1c decrease by 1.1 / 0.5 / 0.8 points. Continued evidence collection therefore provides no aggregate quality gain on this dataset.
  • The additional evidence has a consistent token cost. Removing gating increases prefill from 32.95 / 13.43 / 19.57k to 40.19 / 17.45 / 23.96k on FIN / PUB / DEG: an extra 7.24 / 4.02 / 4.39k tokens, or 22.0% / 29.9% / 22.4%. This additional processing yields composite F1 changes of +0.8 / +0.3 / -0.8 points, respectively.

PCRC offers a favorable quality-cost trade-off: it saves 18.0% / 23.0% / 18.3% of prefill relative to No reliability gating, with composite F1 differences within 0.8 percentage points and higher quality on DEG.

Insisting on using all selected exclusive related paragraphs as evidence for judgment and reconstruction incurs additional prefill-token consumption: full evidence consumption and the savings from skipping later evidence cannot be obtained simultaneously. PCRC does not force the model to make an early binary decision; the uncertain option allows it to progressively request more evidence before committing. Our experiments demonstrate that this design maintains comparable task quality while avoiding the additional prefill required to consume all remaining evidence unconditionally.

3. The counterexample does not contradict our composite evaluation

The composite metric can reward missing difficult tables. In §A.3 (lines 1251–1258), F1-TEDS = F1 × EQ uses cell-weighted EQ only over true positives. With two equally sized positives, TEDS 1.0 and 0.1, and no false positives, selecting both scores 0.55; missing the second scores 0.667. Higher scores need not mean more complete extraction.

We want to clarify that our composite evaluation comprises three jointly reported scores: composite precision PcP_{c}, composite recall RcR_{c}, and composite F1 F1cF_{1c}. To assess the combined outcome of condition judgment and table reconstruction, these three scores should be considered together. Table 3 in the paper follows this reporting protocol; the evaluation is not based on F1cF_{1c} alone.

In your example, the two relevant tables have equal numbers of cells and TEDS scores of 1.0 and 0.1. Recovering both tables gives a lower F1cF_{1c} than returning only the easier table (0.550 versus 0.667), but a higher RcR_{c} (0.550 versus 0.500). The benefit of recovering the difficult table therefore remains visible within the three composite metrics themselves.

Table R3d. Composite scores for the reviewer's two equally sized positive tables, with TEDS scores of 1.0 and 0.1 and no false positives. Values are rounded to three decimals.

Returned tablesPcP_{c}RcR_{c}F1cF_{1c}
Both tables0.5500.5500.550
Only the perfect table1.0000.5000.667

Reading the three composite scores together does not support treating the output that misses the difficult table as uniformly better: its higher PcP_{c} and F1cF_{1c} are accompanied by lower RcR_{c}. Thus, although the scalar F1cF_{1c} reverses the ordering in this example, the composite evaluation still identifies the coverage advantage of recovering both relevant tables. This is why joint judgment-and-reconstruction quality should be assessed through PcP_{c}, RcR_{c}, and F1cF_{1c} together.

Reviewer_FJ7cOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Motivation

Existing table extraction systems follow an "extract-all" paradigm: they rebuild every table in a document, then filter downstream. Real use is query-driven, since users want only tables that meet a specific need. A cropped table image alone also rarely shows whether the table is relevant, because its role depends on captions and surrounding text.

Contributions

QCTE (Query-Conditioned Table Extraction): Given a document and a natural-language condition, output structured HTML only for the tables that satisfy it. This adds a required relevance-judgment step.

PCRC (Progressive Context-Reusable Cache), with three mechanisms:

Table-centric relevance index: Each paragraph–table pair is scored as semantic similarity × spatial proximity. The product means a weak match on either factor pushes the score toward zero. Paragraphs are split into document-level shared ones and table-specific exclusive ones. Progressive evidence injection: The model may answer yes, no, or uncertain. Only when uncertain is the next exclusive paragraph added and the judgment redone, until the answer is certain or evidence runs out (then a yes/no is forced). This limits noise and saves compute. KV-cache reuse: A fixed head (instructions, policy, query, shared paragraphs) is prefilled once and reused across tables. A replaceable tail (current table image, OCR, exclusive paragraphs) is discarded when switching tables. States computed during judgment carry over to HTML reconstruction. Three benchmarks: FinConTab (finance), PubConTab (scientific), and DegConTab (degraded documents). Together they hold about 429K documents, 1.45M tables, and 1.29M conditions, evenly spread over five condition types. Labels were LLM-generated, filtered automatically, and 5% were human-verified (98.5% agreement).

Results

Quality: Judgment F1 rises from about 0.73 (single-pass Qwen2.5-VL-7B) to 0.90. Composite F1-TEDS improves by 13–16 points, and PCRC beats six modular extractor-plus-judge pipelines on all datasets. Efficiency: Prefill compute falls to 24–29% of the no-cache version (3.4–4.1× savings), with identical quality, so caching does not change outputs. Ablations: Random paragraphs give almost no gain, and all paragraphs give only 0.84–0.85, so the gain comes from choosing the right context. Removing the reliability gate leaves quality unchanged but raises cost. Generalization: On GLM-4.5V, judgment F1 reaches 0.95–0.96. Context also improves extraction on degraded documents (TEDS +9.6–20.8) but barely on clean ones (+0.3–0.8).

Critical strengths
  • Clear task definition and boundaries. The paper recasts table extraction from "extract-all, then filter" into query-conditioned selective extraction, which fits how people actually use documents. It separates itself from DSEBench, nlcTables, and UniDisc, which retrieve already-structured tables, whereas this work jointly selects and reconstructs tables from rendered document images. Table 1 (input, output, relevance judgment) states the positioning concisely.

  • Rigorous evaluation protocol. Judgment uses micro P/R/F1. Extraction uses TEDS on true positives only. A cell-count-weighted EQ is multiplied in to give a composite (equivalent to F1·TEDS), which prevents systems from scoring well by judging accurately but extracting poorly. Appendix A.3 fixes the ambiguous details: one-to-one matching at IoU ≥ 0.5, with the threshold untuned; fixed rules for nine alignment cases (Table 9); unmatched candidates not counted as negatives; task failures not treated as "no" predictions; quality and cost reported separately.

  • Thorough ablations, including unfavorable results. Seven context-selection variants were all budget-matched to the same paragraph budget. The results show that quality comes from selecting the right context, not from adding more text: random gets 0.71–0.73, close to the no-context baseline, and all context gets 0.84–0.85, below PCRC's 0.90. Cost savings come from early stopping, not retrieval. Sweeps over four hyperparameters and over the decay rate λ are also included (Tables 17–18). The authors also admit that: a generic prefix cache trails by only 0.7 points; a single-pass judgment protocol scores higher judgment F1 on Pub (0.918 vs 0.901); the default hyperparameters are not optimal, since K_excl=2 does better and uses 5–14% fewer tokens.

  • Large, well-documented dataset. The benchmarks hold about 429K documents, 1.45M tables, and 1.29M conditions across finance, scientific, and degraded settings, with the five condition types balanced at 20% each (Table 7). Human verification is substantial: 64,385 conditions were sampled, covering 217,229 condition–table pairs, and 12,877 conditions were annotated independently by a second person, with a third arbitrating. The paper reports: automatic–human agreement of 98.5%; false-positive label rate of 2.4%; missed-positive rate of 0.9%; inter-annotator agreement of 98.5%; counts of undecidable ("U") items removed. It also states that 95% of conditions were only automatically checked, avoiding overclaiming.

  • Practical, loosely coupled design. The "fixed head + replaceable tail" cache frame (Eq. 4) is simple to implement. Document-level evidence persists, table-level evidence is discarded after use, and judgment states carry into extraction, with all three reuse boundaries made explicit in Algorithm 1. Layout detection and OCR are treated as swappable front-end components (citing PP-DocLayout-L at 90.4% mAP), so the method is not tied to a specific parser. Results on GLM-4.5V show the same pattern (judgment F1 0.88–0.90 → 0.95–0.96, reuse ratio 22–31%), indicating the framework is not tuned to a single model.

Critical weaknesses
  • The method's name does not match its real contribution; the KV-cache mechanism adds little on its own. The paper's own Table 14 undermines the headline "Progressive Context-Reusable Cache."

  • Running the PCRC schedule with vLLM's off-the-shelf prefix cache gives reuse ratios of 28.66/29.13/25.01, versus PCRC's 28.29/29.02/24.35. The gap is under 0.7 points, and quality is identical by construction. The authors concede that "the reuse follows from the schedule." The real gains come from retrieval-based context selection and progressive judgment scheduling, a more modest contribution. The wall-clock gain from progressive scheduling is only 3–5% (Table 16, 20 documents), an order of magnitude below the 22–28% prefill-token savings. The main baseline is weak, so the headline gain may not generalize.

  • The main tables (3 and 11) compare only against the 7B Qwen2.5-VL, given no context. The most direct control, the same Qwen with PCRC-selected paragraphs in a single pass, is in the appendix, and it scores higher judgment F1. Appendix Table 19 shows that Claude Opus 4.5, GPT-5.2, and Qwen3-VL-Plus reach 0.927–0.967 judgment F1 and 0.846–0.871 composite in a single pass with no context. This beats PCRC on 7B (0.901 and 0.723–0.794), so the "0.73→0.90" story would shrink sharply with stronger models. PCRC was never tested on a frontier model. The stated reason is that APIs do not expose KV state, but the "select context, then judge once" variant could have been run through APIs.

  • The benchmarks are less novel and less difficult than claimed. FinConTab and PubConTab reuse existing page and table annotations and add only LLM-generated conditions. DegConTab is the only new set, and it is fully synthetic, so its degradations likely differ from real scans and photos. Yet this set shows PCRC's largest relative gain (composite 0.794) and the largest context benefit for extraction (TEDS +9.6–20.8, Table 20). The strongest evidence therefore comes from the least realistic data. Conditions were generated from reference HTML. This avoids leaking context needs into labels, but it means the task can in principle always be decided from the table alone, which weakens the claim that context is necessary. 95% of conditions were not human-verified, and the 2.4%/0.9% error rates hold only for the 5% sample. The evaluation favors the method, and tuning on the test set is a concern.

  • All quality metrics are conditional on the upstream detector's candidates. Source tables with no matched candidate (Fin 11,562, Pub 34,407, Deg 25,804) are excluded from the recall denominator, and 27.5% of FinConTab candidates are unmatched. The counts are disclosed, but the main tables have no end-to-end metric, so real deployment performance is unclear. Section 5.1 states there is no train/dev/test split and all experiments run on the full datasets. The hyperparameter and λ sweeps (Tables 15 and 18) were run on the same data. The defaults are said to be within 1 judgment-F1 point of the best, but tuning on the evaluation set is still a methodological flaw. Metrics switch between abstract and body, and key analyses are missing.

The abstract reports judgment F1 (0.73→0.90), while the body's Table 3 reports composite F1-TEDS (0.597→0.723, etc.). The judgment-only view is in appendix Table 11. Mixing the two overstates the gain (+17 points vs +12–15). There are no qualitative examples or error analysis. Figure 1 is only a pipeline diagram, with no case showing how context changes a judgment and no failure cases. There is no empirical comparison with CacheBlend or RAG-Cache, only a conceptual one in Related Work, though these are the most relevant baselines. Memory overhead, the decline in prefix-reuse benefit under large-batch serving, and head-cache eviction under concurrent multi-document requests are not discussed. Table 16's timing uses only 20 documents per dataset and one condition per document, a small sample.

Questions for authors

None

5 comments

We thank the reviewer for the positive assessment and constructive suggestions. We clarify the following points below.

1. Clarifying the role of PCRC's cache management

Running the PCRC schedule with vLLM's off-the-shelf prefix cache gives reuse ratios of 28.66/29.13/25.01, versus PCRC's 28.29/29.02/24.35. The gap is under 0.7 points, and quality is identical by construction. The authors concede that "the reuse follows from the schedule." The real gains come from retrieval-based context selection and progressive judgment scheduling, a more modest contribution.

We would like to clarify how PCRC handles cache boundaries and retains shared context across requests. In the evaluated generic configuration, repeated prefill can arise when tables are replaced or cached prefixes are evicted, as explained below.

Block boundaries and repeated prefill

  • The generic cache manages KV storage in blocks. The evaluated vLLM cache allocates and recycles cache blocks, and reuses a cached prefix only when complete blocks match. Our control uses 16-token blocks (vLLM Project, n.d.-a, n.d.-b).
  • The shared head and table content can occupy the same block. If their boundary falls inside a block, that block contains both shared-head KV states and table-specific KV states. When the table tokens in this block change, the old block no longer matches. The next request must therefore prefill the shared-head tokens in that block again, even though those tokens are unchanged. Complete-block matching cannot reuse the unchanged head portion.
  • PCRC preserves the shared head when replacing the table. It retains all head KV states up to the exact head boundary and replaces only the table tail, avoiding this repeated prefill. This mechanism provides an explanation for PCRC's consistently lower prefill token consumption compared with the generic cache across all three evaluated datasets, as shown in Table R2a, which compares the two cache implementations under the same PCRC schedule using the original results from Tables 4 and 14.

Table R2a. Prefill tokens per condition (k) under the same PCRC schedule, from Tables 4 and 14. Lower is better.

Cache managementFINPUBDEG
Generic33.3813.4720.11
PCRC32.9513.4319.57

Retaining the shared head throughout the task

  • PCRC retains the head for the entire document–condition task. The head remains live while individual table tails are replaced. Its lifetime therefore spans judgment and extraction requests and successive tables.
  • Default generic caching does not reserve an idle prefix for the remaining document task. Once a request releases its references, unrelated requests sharing the cache pool can displace the inactive prefix under capacity pressure. The next table then requires prefill again, despite the head being unchanged (vLLM Project, n.d.-b).
  • A concrete example: in a four-block LRU pool, task A leaves two head blocks idle between table requests. If other requests insert four distinct blocks without touching A's head, both head blocks can be evicted. A must recompute them when it resumes.
  • PCRC keeps the shared head allocated until the task finishes. Its unchanged KV states are retained between requests instead of being released as idle cache entries that other tasks can replace. This keeps the shared head available for subsequent requests and prevents repeated prefill caused by other tasks evicting it. Default generic prefix caching does not provide this retention guarantee for the entire document–condition task.

Reference

vLLM Project. (n.d.-a). Automatic Prefix Caching. vLLM Design Documentation.

vLLM Project. (n.d.-b). BlockPool (vllm.v1.core.block_pool). vLLM API Reference.

2. Quality and prefill savings, with supplementary timing results

The wall-clock gain from progressive scheduling is only 3–5% (Table 16, 20 documents), an order of magnitude below the 22–28% prefill-token savings.

We want to clarify that our principal contributions concern condition-judgment and table-reconstruction quality with reduced prefill-token consumption; we report execution time as a supplementary measure. In response to your request, we supplement the timing results using 500 documents per dataset. Table R2b reports inference times for PCRC and single judgment, both using generic prefix caching and excluding preprocessing. It also compares PCRC's time including preprocessing against single judgment's inference-only time, showing whether PCRC remains faster even when preprocessing is counted for PCRC alone.

Table R2b. Expanded timing comparison using 500 documents per dataset, with one condition per document. Times are mean seconds per condition; model loading is excluded, and the remaining measurement settings follow Table 16. Both schedules use generic prefix caching. The first two timing columns exclude preprocessing. The additional PCRC timing column includes detection, text acquisition, embedding, and relevance-index construction. Ratio 1 is single-judgment inference time divided by PCRC inference time. Ratio 2 is single-judgment inference time divided by PCRC time including preprocessing. A Ratio 2 above 1 means that PCRC remains faster even with preprocessing included.

DatasetPCRC + generic, inference only (s)Single judgment + generic, inference only (s)Ratio 1: single / PCRCPCRC + generic, including preprocessing (s)Ratio 2: single / PCRC including preprocessing
FIN61.366.21.0864.71.02
PUB45.848.11.0546.11.04
DEG59.761.81.0461.21.01

As shown in Table R2b, even with preprocessing included, PCRC remains faster on average than single judgment without preprocessing on all three datasets. These timing results complement our main findings on quality and prefill savings.

3. Additional quality evaluation on frontier models

PCRC was never tested on a frontier model. The stated reason is that APIs do not expose KV state, but the "select context, then judge once" variant could have been run through APIs.

To address your question about frontier models, we additionally evaluated PCRC's progressive judgment and extraction schedule through the APIs of Claude Opus 4.5, GPT-5.2, and Qwen3-VL-Plus. Table R2c reports the complete quality results, and Table R2d compares them with the original Table 19 no-context baselines.

Table R2c. Frontier models with the PCRC progressive schedule through APIs. P, R, and F1 are judgment scores; Pc, Rc, and F1-TEDS are composite scores.

DataModel + PCRCPRF1PcRcF1-TEDS
FINClaude Opus 4.50.9720.9590.9660.8940.8820.888
FINGPT-5.20.9630.9950.9780.8660.8940.880
FINQwen3-VL-Plus0.9410.9570.9490.8640.8780.871
PUBClaude Opus 4.50.9560.9740.9650.8740.8900.882
PUBGPT-5.20.9920.9610.9760.9110.8830.897
PUBQwen3-VL-Plus0.9720.9350.9530.9140.8780.896
DEGClaude Opus 4.50.9741.0000.9870.8570.8800.868
DEGGPT-5.20.9440.9600.9520.8590.8740.866
DEGQwen3-VL-Plus0.9590.9710.9650.8430.8530.848

Table R2d. Gains over the original Table 19 no-context baselines, in percentage points. Each cell gives judgment F1 / composite F1-TEDS improvement. Gains are calculated from the new scores and the reported baseline scores.

ModelFINPUBDEG
Claude Opus 4.5+2.770 / +2.900+2.560 / +2.092+2.001 / +2.045
GPT-5.2+2.336 / +3.378+3.413 / +2.577+2.913 / +1.717
Qwen3-VL-Plus+2.172 / +2.384+1.990 / +2.670+2.892 / +3.196

All nine model–dataset combinations improve on both judgment F1 and composite F1-TEDS. Judgment F1 increases by 1.990–3.413 percentage points, and composite F1-TEDS by 1.717–3.378 points. The gains are smaller than on the 7B backbone, but remain positive across all three frontier models and datasets.

4. Synthetic degradation and the larger context benefit on DEG

DegConTab is the only new set, and it is fully synthetic, so its degradations likely differ from real scans and photos. Yet this set shows PCRC's largest relative gain (composite 0.794) and the largest context benefit for extraction (TEDS +9.6–20.8, Table 20).

  • Why we use synthetic degraded documents. DEGCONTAB is synthetic, as you noted. Collecting and openly releasing real degraded documents with reliable annotations is difficult: official and business records bearing stamps may contain sensitive information and may not be available for unrestricted redistribution. DocILE identifies sensitive information in business documents as an obstacle to creating large public datasets (Šimsa et al., 2023). Documents may also undergo cleaning or removal of material that obscures information before digitization, as described in Library of Congress preservation guidance (Library of Congress, n.d.). Publicly released copies therefore need not preserve the original damage. Controlled synthesis lets us study these degradations.

  • Why the benefit is largest on DEG. The task requires condition judgment and reconstruction of tables whose visual evidence may be obscured by watermarks, stains, or stamps. Related paragraphs can supply information missing from the visible table, making their support especially valuable for both judgment and reconstruction. The larger improvement observed on DEG is consistent with this role of contextual information.

References

Library of Congress. (n.d.). Stabilization Treatments and Storage Enclosures. In Preservation Actions for Collection Digitization. Library of Congress Research Guides.

Šimsa, Š., Šulc, M., Uřičář, M., Patel, Y., Hamdi, A., Kocián, M., Skalický, M., Matas, J., Doucet, A., Coustaty, M., and Karatzas, D. (2023). DocILE Benchmark for Document Information Localization and Extraction. In Document Analysis and Recognition – ICDAR 2023, Part II. Lecture Notes in Computer Science, vol. 14188, pp. 147–166. Springer, Cham.

Reviewer_uzsbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Query-Conditioned Table Extraction, which identifies tables satisfying a natural-language condition and reconstructs the selected tables as HTML. The proposed PCRC framework combines table-centered context retrieval, progressive evidence incorporation, and KV-cache reuse across table judgments and reconstruction. Three benchmarks cover financial, scientific, and synthetically degraded documents. With the same Qwen backbone, PCRC improves judgment F1 by approximately 17 percentage points and composite F1-TEDS by 12.6 to15.5 points.

Critical strengths
  1. PCRC jointly addresses table selection and HTML reconstruction, improving judgment F1 by approximately 17 points and composite F1-TEDS by 12.6–15.5 points with the same Qwen backbone. Besides, context ablations support the importance of relevant evidence, while early-stopping experiments distinguish accuracy gains from computational savings.

  2. The datasets cover three document domains and five condition categories, with human checks of over 217,000 condition–table pairs.

Critical weaknesses
  1. DoTabler already supports query-driven PDF table retrieval, requiring sharper positioning of QCTE’s additional contribution. Generic caching nearly reproduces PCRC’s savings, while a single judgment with the same selected context achieves comparable quality.

  2. Cross-page spatial distances are underspecified, and high average relevance does not guarantee relevance to most tables. Extraction quality is evaluated on each method’s own true-positive set, so composite gains do not independently establish improved reconstruction on comparable tables.

Questions for authors
  1. Would it be better to add a task comparison with nlcTables, UniDisc, and DoTabler?
4 comments

We thank the reviewer for their comments and the opportunity to clarify our contributions. We address the concerns below.

1. DoTabler, nlcTables, and UniDisc address fundamentally different tasks from our proposed Query-Conditioned Table Extraction.

DoTabler already supports query-driven PDF table retrieval, requiring sharper positioning of QCTE's additional contribution. Would it be better to add a task comparison with nlcTables, UniDisc, and DoTabler?

Query-Conditioned Table Extraction (QCTE) takes a PDF document and a user-specified natural-language condition, determines which tables satisfy that condition, and returns reconstructed HTML for the accepted tables. Our contributions are:

  1. Task formulation: joint condition-wise table selection and faithful HTML reconstruction from PDF documents.
  2. Contextual evidence: identifying and experimentally demonstrating how relevant document paragraphs support table-condition judgment and reconstruction.
  3. Efficient execution: a progressive cache-management mechanism that reuses shared evidence, accumulates table-specific evidence when needed, and carries judgment states into reconstruction to reduce prefill-token consumption.

DoTabler (Li et al., 2025) covers a different input–output contract:

  • Input: a PDF document and a natural-language query; preprocessing supplies layout blocks and OCR text.
  • Output: top-k table blocks with associated paragraphs.
  • Missing condition-satisfaction stage: its binary classifier predicts table–paragraph association. Its query module ranks query–table relevance and returns top-k results; it does not formulate or evaluate QCTE's per-table yes/no condition-satisfaction decisions (Sections 3.3 and 3.5).
  • Different use of context: identifying associated paragraphs does not itself determine whether a table satisfies a condition. Its published retrieval score uses the query–table pair; it does not provide a condition judge that progressively consults those paragraphs.
  • Missing reconstruction stage: it does not reconstruct accepted tables as HTML. Its released targets cover association and retrieval, rather than HTML cell/row/span reconstruction.
  • Missing cache mechanism: it does not provide shared-head/table-tail management or judgment-to-reconstruction KV reuse.

DoTabler cannot be directly evaluated as an end-to-end QCTE baseline because its published workflow does not perform condition-wise table judgment or HTML reconstruction. Its benchmark does not evaluate this joint task.

Table R1a. Inputs, outputs, and task coverage of the suggested comparisons.

WorkInputNative task and outputWhat is missing for the complete QCTE task
nlcTables (Cui et al., 2025)Structured candidate tables and natural-language conditions; some settings also include a query tableBenchmark for relevant-table discovery, including NL-only, union, and join settingsIt is a discovery benchmark, not an extraction algorithm; it does not evaluate PDF-to-HTML reconstruction.
UniDisc (Wang et al., 2026)Structured tables and natural-language statements in a data lake; NL or table queriesRelevant-table ranking or binary relevance prediction using graph-based representationsIt assumes structured candidate tables and supplies no visual table-to-HTML reconstruction stage.
DoTabler (Li et al., 2025)PDF document and natural-language query; layout blocks and OCR from preprocessingQuery-based top-k table retrieval and table–paragraph associationsIts published task does not specify per-table condition-satisfaction decisions or conditional HTML reconstruction; it also lacks progressive evidence consumption and judgment-to-reconstruction KV reuse.
QCTE / PCRC (ours)PDF document and a natural-language conditionPer-table condition decisions and reconstructed HTML for accepted tablesPCRC implements the joint task with selected paragraph evidence, progressive judgment, and KV reuse across tables and into reconstruction.

We will add this comparison. Using these works as complete QCTE baselines requires supplying their missing decision, parsing, and/or reconstruction components; their published retrieval results cannot directly evaluate the joint task.

2. PCRC Preserves Quality While Reducing Token Consumption

A single judgment with the same selected context achieves comparable quality.

  • This result supports the role of selected context. Single judgment receives all shared and exclusive paragraphs selected by PCRC in one call. Its strong judgment and reconstruction results are expected: it benefits from the contextual evidence identified by our method. Together with the no-context and random-context controls, this supports our finding that relevant surrounding paragraphs help conditional table extraction. Comparable quality under full evidence does not weaken this contribution.
  • PCRC's advantage over this baseline is lower token consumption. Progressive evidence injection, early stopping, and explicit KV reuse allow PCRC to retain comparable quality while avoiding unnecessary evidence processing and repeated prefill. Table 14 provides the direct comparison below.

Table R1b. Single judgment + generic prefix cache versus full PCRC, reproduced from Table 14. Prefill is measured in thousands of tokens per condition. FIN, PUB, and DEG denote FINCONTAB, PUBCONTAB, and DEGCONTAB.

DatasetJudgment F1: single → PCRCF1-TEDS: single → PCRCSingle prefill (k)PCRC prefill (k)Prefill reduction
FIN0.902 → 0.9010.722 → 0.72342.6832.9522.8%
PUB0.918 → 0.9010.759 → 0.75618.8213.4328.6%
DEG0.905 → 0.9020.784 → 0.79426.4319.5726.0%

3. What generic prefix caching does not guarantee

Generic caching nearly reproduces PCRC's savings.

Compared with PCRC, the evaluated generic cache cannot guarantee full reuse of unchanged context. It can require repeated prefill when tables are replaced or cached prefixes are evicted. The reasons are as follows.

3.1. Replacing a table can force part of the shared head to be recomputed

  • The generic cache manages KV storage in blocks. The evaluated vLLM cache allocates and recycles cache blocks, and reuses a cached prefix only when complete blocks match. Our control uses 16-token blocks (vLLM Project, n.d.-a, n.d.-b).
  • The shared head and table content can occupy the same block. If their boundary falls inside a block, that block contains both shared-head KV states and table-specific KV states. When the table tokens in this block change, the old block no longer matches. The next request must therefore prefill the shared-head tokens in that block again, even though those tokens are unchanged. Complete-block matching cannot reuse the unchanged head portion.
  • PCRC preserves the shared head when replacing the table. It retains all head KV states up to the exact head boundary and replaces only the table tail, avoiding this repeated prefill. This mechanism can explain part of the observed prefill reduction relative to the generic cache; the present comparison does not isolate its individual contribution.

Table R1c. Prefill tokens per condition (k) under the same PCRC schedule, from Tables 4 and 14. Lower is better.

Cache managementFINPUBDEG
Generic33.3813.4720.11
PCRC32.9513.4319.57

3.2. A task-owned head avoids dependence on inter-request cache eviction

  • PCRC retains the head for the entire document–condition task. The head remains live while individual table tails are replaced. Its lifetime therefore spans judgment requests and successive tables.
  • Default generic caching does not reserve an idle prefix for the remaining document task. Once a request releases its references, unrelated requests sharing the cache pool can displace the inactive prefix under capacity pressure. The next table then requires prefill again, despite the head being unchanged (vLLM Project, n.d.-b).
  • A concrete example: in a four-block LRU pool, task A leaves two head blocks idle between table requests. If other requests insert four distinct blocks without touching A's head, both head blocks can be evicted. A must recompute them when it resumes.
  • PCRC keeps the shared head allocated until the task finishes. Its unchanged KV states are retained between requests instead of being released as idle cache entries that other tasks can replace. This design keeps the shared head available for subsequent requests. Unlike default generic prefix caching, it explicitly maintains this state throughout the document–condition task. Our reported comparison does not separately quantify the effect of eviction under concurrent workloads.

4. Cross-page spatial distance

Cross-page spatial distances are underspecified.

We will add the following description of cross-page spatial distance to the revised manuscript:

  • Single-column pages: retain the original layout; no column unfolding is needed.

  • Double-column pages: move the right column below the left column, aligning their vertical edges to form one vertical strip. The positions of blocks within each column are preserved.

  • Join the pages: concatenate the resulting page strips vertically in page order, placing each immediately below the preceding one. This forms a single tall document canvas, like a continuous scroll, on which every table and paragraph has a document-level position.

  • Coordinate origin and bbox centers: use the upper-left corner of the first page on the unfolded canvas as the origin O=(0,0)O=(0,0), with the horizontal axis pointing right and the vertical axis pointing down. For a component XX whose bbox has bounds (xmin⁡,ymin⁡,xmax⁡,ymax⁡)(x_{\min},y_{\min},x_{\max},y_{\max}) in this coordinate system, its center is

    \frac{x_{\min}+x_{\max}}{2},\\; \frac{y_{\min}+y_{\max}}{2} \right).$$ The first page's upper-right corner or either lower corner of the last page could also serve as the origin; a consistent change of origin and axis directions leaves the Euclidean distances unchanged.
  • Compute the distance: represent each table or paragraph by the geometric center of its bounding box on this canvas. The spatial distance between a paragraph and a table is the two-dimensional Euclidean distance between their centers, accounting for both horizontal and vertical separation across columns and pages.

Writing the bounding-box center on the unfolded canvas as c(X)=(xX,yX)c(X)=(x_X,y_X), the distance is

=\sqrt{(x_{P_j}-x_{T_i})^2+(y_{P_j}-y_{T_i})^2}.$$ We then apply the document-level normalization and default spatial scoring from Sections 3.1 and A.4.6: $$\hat d_{ij}=\frac{d_{ij}}{d_{\max}},\qquad s_{\mathrm{spa}}(P_j,T_i)=\exp(-\hat d_{ij}).$$ Here $d_{\max}$ is the largest paragraph–table distance in the document. **Please feel free to raise any further questions or let us know if any clarification would be helpful.** ## References Cui, L., Li, H., Chen, K., Shou, L., and Chen, G. (2025). nlcTables: A Dataset for Marrying Natural Language Conditions with Table Discovery. *Proceedings of SIGIR 2025*. Li, X., Dong, J., and Wong, R. (2025). From Surface to Semantics: Semantic Structure Parsing for Table-Centric Document Analysis. *Proceedings of ECAI 2025*. Wang, T., Huang, S., Bao, Z., Culpepper, J. S., Sadiq, S., Dedeoglu, V., and Arablouei, R. (2026). Unified Data Discovery across Query Modalities and User Intents. *arXiv preprint*, arXiv:2604.27252. vLLM Project. (n.d.-a). Automatic Prefix Caching. *vLLM Documentation*. vLLM Project. (n.d.-b). BlockPool (vllm.v1.core.block_pool). *vLLM API Reference*.

Thank you for your detailed and active response.

I think the authors have clearly distinguished the differences between this work and DoTabler, nlcTables, and UniDisc. Also, the definition of cross-page spatial distance is clear now. I thought it would be good to update it in the paper.

But I also have two issues about what the authors want to emphasize as a contribution.

  1. When the schedule remains constant, replacing PCRC native head/tail management with vLLM’s prefix cache results in a performance difference of only 0.3%–2.7%. In contrast, comparing a single judgment + generic prefix cache with full PCRC setting yields savings of 22.8%–28.6%. This indicates that the gains primarily come from the progressive schedule, not from the cache design. Furthermore, the authors agree that there are no significant time savings, emphasizing instead that the primary benefit lies in reduced cache usage.

  2. And my concern about extraction on the method-specific true-positive set was not addressed. The EQ score is a TEDS-based metric, which weighted by the number of cells, calculated exclusively for ture positive sets. It essentially measures the quality of the HTML reconstruction for tables that were correctly selected. By examining the PcP_c, RcR_c, the F1F_1 score, which incorporates EQ, and also the evaluation protocal and formulation provided in the main paper and the appendix, it becomes clear that the performance gains observed in the comparison such as between Qwen and PCRC comes from the correct selection of a greater number of tables, rather than from better HTML reconstruction. However, the authors repeatedly claim in the paper that the context aids both the selection process and the reconstruction. From my perspective, this is an overclaim that is not supported by the numerical results.

Overall, I thought this task and data are useful, but the experiment results does not support the cache and evaluation claim in the paper. So I would keep my current rating.

Comments on the paper

1 comment

Thank you to all reviewers for the careful reading and constructive feedback. We note that the reviewers' concerns mainly center on the following three points. We address them together here for ease of reading and reference.

  1. Does PCRC lack robustness that it only has good performance on detector-generated component blocks?

    We directly evaluated condition judgment and HTML reconstruction using the datasets' source-annotated component bounding boxes, replacing detector-generated component proposals. PCRC achieves composite F1 scores of 0.713 / 0.750 / 0.811 on FINCONTAB / PUBCONTAB / DEGCONTAB, close to the original 0.723 / 0.756 / 0.794, and retains its lead over the non-PCRC baselines. Its prefill savings relative to the same schedule without caching remain 72.85% / 72.10% / 76.86%. These additional experiments show that PCRC's quality and token-saving advantages persist when inputs are supplied directly by source annotations instead of the upstream detector tools.

  2. Why use PCRC's explicit cache management rather than generic prefix caching alone?

    PCRC explicitly controls which KV states are retained and when table-specific states are replaced. Two distinctions matter:

    • Preserving the complete shared head. In the evaluated generic configuration, prefix matching operates on complete cache blocks. When shared-head and table tokens occupy the same block, changing the table prevents that block from matching and requires unchanged shared tokens to be prefilled again. PCRC preserves the shared head at its exact token boundary and replaces only the table tail. This maximizes reuse of the unchanged shared-head states and avoids unnecessary prefill.
    • Retaining shared states throughout the task. PCRC keeps the shared head allocated across successive requests within the document-condition task. Default generic caching does not reserve an idle prefix for that task; other requests can evict it, requiring repeated prefill.

    Under the same PCRC schedule, the original Tables 4 and 14 report lower prefill consumption for native PCRC on all three datasets: 32.95 / 13.43 / 19.57k, compared with 33.38 / 13.47 / 20.11k for generic caching.

  3. How should the timing results be interpreted?

    Our principal efficiency contribution is reduced prefill-token consumption, rather than wall-clock acceleration. The original Table 14 reports 22.8% / 28.6% / 26.0% less prefill than single-pass inference.

    We have additionally expanded the timing evaluation to 500 documents per dataset, with one condition per document. With generic caching used for both schedules, PCRC takes 64.7 / 46.1 / 61.2 seconds including preprocessing, compared with 66.2 / 48.1 / 61.8 seconds for single judgment excluding preprocessing. These are mean times per condition; model loading is excluded. Thus, even when preprocessing is counted for PCRC alone, its mean processing time remains lower on all three datasets.

Thank you again for the helpful feedback. We hope these clarifications and additional experiments address the shared concerns, and we welcome further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.