Pith. sign in

REVIEW 3 major objections 5 minor 82 references

Semantic relevance does not ensure a table can answer the query: dense retrievers repeatedly pick the wrong sibling table, and explicit answerability scoring largely fixes it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 17:03 UTC pith:6BBTN3LL

load-bearing objection Solid diagnostic benchmark showing dense retrievers can't pick the answerable sibling table; the AAR attribution claim overreaches and needs a matched-capacity control. the 3 major comments →

arxiv 2607.17742 v1 pith:6BBTN3LL submitted 2026-07-20 cs.AI

Semantically Similar, Logically Distinct: Diagnosing the Semantic-Answerability Gap in Table RAG

Jiaming Tian , Liyao Li , Wentao Ye , Haobo Wang , Lihua Yu , Zujie Ren , Gang Chen , Junbo Zhao This is my paper
classification cs.AI
keywords table retrievalretrieval-augmented generationanswerabilitysemantic relevancesibling tablesdense retrievalrerankingbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that table retrieval for RAG has been optimizing the wrong target: semantic relevance, not answerability. On a benchmark built from sibling tables—tables with nearly identical schemas but subtly different contents—the dense retrievers tested routinely find the right sibling group yet fail to pick the one table that actually contains the evidence, retrieving it at the top rank only 18.2% of the time. That failure is not cosmetic: feeding the top five retrieved tables to a QA model drops its F1 from 0.755 (oracle) to 0.330, and adding more tables only adds interference. A lightweight answerability-aware reranker that scores each candidate against the query raises top-1 retrieval to 57.4%, which the authors take as evidence that the bottleneck is a missing verification step, not model capacity.

Core claim

The paper's central claim is that the Semantic-Answerability Gap—the distance between coarse semantic relevance and fine-grained evidence sufficiency—is measurable, systematic, and not explained by surface variation. On its TCR-Bench benchmark, every dense retriever evaluated achieves strong group-level recall (GR@1 up to 0.670) while within-group discrimination stays at or below the random baseline (DS@1 around 0.298 or lower), so the answerable table ranks first only 18.2% of the time. Controlled perturbations show the gap persists across serialization formats and query phrasings, and trace it to three mechanisms: retrieval reward tracks token density rather than logical sufficiency, model

What carries the argument

The load-bearing instrument is TCR-Bench, a diagnostic benchmark in which each query has exactly one answerable Target Table and several Sibling Distractor Tables that preserve schema and topical overlap while deleting or resampling away the rows and columns that would satisfy the query. The metric that isolates the gap is DS@1 = R@1/GR@1, the ratio of exact top-1 target retrieval to group recall; a value near the random baseline of 0.298 means the retriever finds the neighborhood but cannot tell which member answers. The diagnostic probe is AAR, a two-stage reranker (cross-encoder or LLM judge) that scores each candidate jointly with the query; the large gains it produces identify missing i

Load-bearing premise

The load-bearing premise is that the synthetic sibling distractors really are unanswerable—that column removal, row deletion, resampling, and global verification strip out all query-satisfying evidence while preserving topical similarity, leaving exactly one answerable target per query; if hidden answerability remains, the measured gap and the reranking gains are artifacts of the construction.

What would settle it

Apply the same retrieval protocol to naturally occurring sibling tables in a real data lake—filtered, truncated, or revised versions of one source table—with answerability labels assigned by executing the query's SQL against each version. If a dense retriever's DS@1 rises well above the random baseline, or if AAR's gain shrinks substantially, the Semantic-Answerability Gap is at least partly a product of the synthetic construction rather than a general property of dense retrieval.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Top-rank precision, not recall, governs generation quality: adding more retrieved tables to the prompt degrades QA efficiency, so RAG systems need the answerable table at rank one.
  • Retrieval pipelines should include an explicit query-table verification stage; interaction-based scoring can recover a large share of performance without larger embedding models.
  • Surface-level fixes—changing serialization, paraphrasing queries, shuffling rows or attributes—will not close the gap because the failure sits in the retrieval objective itself.
  • Scaling embedding model size alone is insufficient: moving from 0.6B to 8B parameters improved recall modestly but left within-group discrimination near chance.
  • Table retrieval benchmarks should report answerability metrics such as DS@1 alongside relevance metrics, or they will overstate practical RAG readiness.
  • The same evidence-sufficiency failure should appear in passage-level RAG, where a retrieved chunk can be on-topic yet lack the specific facts needed; the sibling-table setting is a controlled instance of a broader retrieval-objective mismatch.
  • The token-density bias predicts a concrete, testable failure: retrievers should systematically favor longer tables or tables with more matching tokens, so inserting irrelevant rows into an answerable table should raise its score.
  • If the gap is objective-level rather than capacity-level, training a first-stage retriever with answerability-aware hard negatives—the sibling distractors themselves—should natively close much of the gap, making the reranking stage less necessary.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same evidence-sufficiency problem should appear in passage-level RAG, where a retrieved chunk can be on-topic yet lack the specific facts needed; the paper's sibling-table construction is a controlled instance of a broader retrieval-objective mismatch.
  • Editorial inference: the token-density bias predicts a concrete, testable failure—retrievers should systematically favor longer tables or tables with more matching tokens, so inserting irrelevant rows into an answerable table should raise its score; this can be checked without rebuilding the benchmark.
  • Editorial inference: if the gap is objective-level rather than capacity-level, training a first-stage retriever with answerability-aware hard negatives (the sibling distractors themselves) should natively close much of the gap, making the reranking stage less necessary—a direct follow-up the paper leaves implicit.
  • Editorial inference: AAR's gain from 18.2% to 57.4% suggests a practical two-stage recipe for existing RAG systems—keep the cheap dense retriever for candidate generation and add a small cross-encoder answerability scorer—but only if first-stage recall is high enough to include the target, so the approach inherits the benchmark's controlled conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper defines a property called 'answerability' for table retrieval: a retrieved table must contain sufficient, constraint-consistent evidence to answer a query, not merely be semantically similar. It introduces TCR-Bench, a diagnostic benchmark built from Spider and BIRD by constructing 'sibling' tables that share schemas and topical content but differ in whether they contain the answer. Dense retrievers are evaluated with two complementary metrics: group recall (GR@k) for coarse semantic relevance and top-k recall (R@k) for exact answerable-table retrieval. Across seven dense models and multiple serialization formats, the paper reports a large Semantic-Answerability Gap: GR@1 can reach 0.67 while R@1 stays near 0.18, and downstream QA F1 drops from 0.755 under an oracle table to 0.330 with top-5 retrieval. Controlled experiments argue that surface variations do not explain the gap, and diagnostic probes attribute it to semantic volume bias, schema-level cue dependence, and weak row-column binding. Finally, the paper proposes Answerability-Aware Reranking (AAR), a two-stage pipeline using either a Qwen3 cross-encoder (AAR-CE) or an LLM judge (AAR-Judge), which raises R@1 from 0.182 to 0.574 and 0.536, respectively. The authors interpret this large gain as evidence that the bottleneck is a missing answerability-verification step rather than a limitation of model capacity.

Significance. The paper identifies and operationalizes an important retrieval failure mode that existing table RAG benchmarks do not isolate: the distinction between retrieving a semantically relevant table and retrieving the uniquely answerable table. The main empirical phenomenon—a consistent gap between group-level and exact retrieval across architecturally diverse dense encoders—is well supported by the reported data, including bootstrap confidence intervals in Appendix O, and the release of code and data is a concrete reproducibility asset. The DS@1 metric, while dependent on sibling-group construction, is a useful diagnostic quantity. If the benchmark construction and the AAR attribution are adequately validated, the paper would make a solid contribution to table retrieval and RAG diagnostics. The main weakness is that the paper's central causal claim about the source of the gap is not established by the AAR experiments as reported, because the reranking baselines are not controlled for cross-encoder interaction and model capacity. With additional controls and a more careful framing, the paper's core diagnostic finding would remain significant.

major comments (3)
  1. [§7, Table 6, Abstract] The causal attribution is not supported by the AAR experiments. AAR-CE is described only as a Qwen3-8B cross-encoder reranker initialized from the same family as the retriever; the manuscript does not state that it is fine-tuned on answerability labels or that it differs from an ordinary relevance reranker. A generic cross-encoder would likely improve over single-vector retrieval even without any explicit answerability modeling. AAR-Judge, the only variant with an explicit binary answerability judgment, uses Qwen3-30B-A3B-Thinking-2507, which is far larger than the 8B retriever. Therefore the abstract's claim that the gain is 'itself evidence that much of the observed failure reflects a missing answerability verification step, rather than an inherent limitation of model capacity alone' does not follow. Please add a matched-capacity generic cross-encoder reranker as a control, clarify whe
  2. [§3.3.2, Appendix A.1/A.5] Every headline number in the paper is computed against answerability labels produced by the sibling-distractor construction. The construction relies on automated operations (30–70% row deletion, full resampling, column removal) followed by manual and global verification, but no quantitative validation of the label quality is reported. There is no inter-annotator agreement for the manual inspection, no count of how many cases were removed as ambiguous, and no sensitivity analysis showing that the main gap persists when only the high-confidence subset is used. Since the authors themselves acknowledge in Appendix A.5 that residual false negatives are 'impossible to completely eliminate,' the paper should report the verification statistics and show that the gap and the AAR gains are robust to label noise. Otherwise the transfer of the conclusion from this controlled benchmark to real-world t
  3. [§4.1.2, Figure 3, Appendix O] The random baseline for DS@1 is used inconsistently. The text states that the expected DS@1 is 0.221 on the original dataset and 0.298 excluding column-deletion variants, yet Figure 3 and the main text use 0.298 as the random baseline for the full benchmark. Since DS@1 depends on the distribution of sibling-group sizes, the baseline should be computed on exactly the same query set and metric definition as the reported DS@1 values. The bootstrap p-values in Appendix O should be stated with reference to that same baseline. This matters because the 'near-chance' interpretation is central to the paper's claim that fine-grained answerability discrimination is not better than chance.
minor comments (5)
  1. [Table 1] Eff@1 values of 1.046 and 1.014 are greater than 1, which is counterintuitive because QA@1 should not exceed the oracle score on the same query. Please explain that the top-1-hit subset is not a random sample and may have a higher oracle score than the full benchmark; otherwise readers may suspect a metric error.
  2. [§7] The phrase 'ensuring comparable model capacity' for AAR-CE is imprecise: a cross-encoder reranker and a bi-encoder retriever of the same parameter count do not have comparable inference-time capacity or interaction mechanism. Please rephrase to avoid overstating the control.
  3. [§1, Abstract] The abstract says 'dropping QA performance from 0.755 (oracle) to 0.330 (top-5 retrieved),' but the QA drop is reported for Qwen3-8B specifically and is not uniform across retrievers. Please qualify the summary statistic.
  4. [Appendix D.4] The SA query-type breakdown is reported with the caveat that the effective sample size is small. Consider reporting the number of SA queries prominently in the main text, since 56 SA queries support the aggregation-related conclusions.
  5. [Limitations] The Limitations section appropriately notes that AAR 'mitigates rather than closes' the gap and depends on initial retrieval quality. This caveat should be reflected in the abstract and conclusion, which currently state the causal claim more strongly.

Circularity Check

0 steps flagged

No significant circularity: the retrieval gap and AAR results are empirical measurements, not reductions to the benchmark construction or to self-citations.

full rationale

The paper's central measurements are empirical rather than constructional. TCR-Bench's answerability labels are produced by a deterministic construction (removing condition/target columns, deleting or resampling rows that would answer the query, and global verification that exactly one Target Table is answerable) and are not fitted to any retriever or reranker; the retrievers are pretrained off-the-shelf embedding models, and AAR-CE/AAR-Judge are used as a pretrained cross-encoder and a zero-shot LLM judge rather than trained on TCR-Bench. The DS@1 metric is explicitly a ratio R@1/GR@1 with a defined random baseline, so the near-chance DS@1 finding is a measurement, not a tautology. The AAR intervention is an independent probe: it is not derived from the benchmark labels, even though the benchmark labels are used for evaluation. The only self-citation (LongTableBench) supplies the downstream QA prompt/pipeline, and the QA numbers reported here are freshly measured, not imported from that paper. The skeptic's concern that AAR-CE may be a generic cross-encoder and AAR-Judge uses a larger model is a validity/attribution issue, not a circular reduction: the paper's causal interpretation could be wrong or confounded, but it does not reduce to the benchmark's construction or to a self-citation chain. No load-bearing step is equivalent to its own input by definition.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claims rest on the benchmark construction (answerability labels), the choice of retriever families, and the QA pipeline. No free parameters are fitted to produce the gap, but several construction thresholds shape its size, and the answerability construct itself is introduced and measured entirely within this paper.

free parameters (4)
  • Row deletion ratio range = 30%–70%
    Appendix A.1: distractor tables delete 30–70% of rows before resampling; this hand-chosen knob shapes how similar siblings are to targets.
  • Token length range = 6,500–7,800 tokens
    Appendix A.1: all sub-tables are sampled to this length; the 'compression pressure' narrative depends on this range.
  • Hybrid dense:BM25 interpolation weight = 9:1
    Appendix D.2: grid-searched best configuration; not central to the main claims but a fitted retrieval hyperparameter.
  • Random-baseline exclusion = exclude column-deletion variants; DS@1 baseline = 0.298
    Section 4.1.2: the DS@1 random baseline excludes column-deletion variants, an analytical choice that raises the bar and shapes the 'near chance' conclusion.
axioms (4)
  • domain assumption A table is answerable iff it provides all required rows, values, and attribute bindings for a non-empty, constraint-consistent result.
    Section 3.1 defines answerability this way; the single-source, all-or-nothing definition excludes partial and compositional answerability, which the Limitations acknowledge.
  • domain assumption Exactly one Target Table per query after sibling construction and global verification.
    Section 3.3.2 and Appendix A.5 rely on manual and global verification to ensure no distractor answers the query; if verification missed semantic duplicates, the labels are wrong.
  • domain assumption Open-source general-purpose dense embeddings are representative of Table RAG retrievers.
    Section 4.1.1/Appendix B evaluate only open-source embedding families; proprietary and table-specific retrievers are excluded, as acknowledged in the Limitations.
  • domain assumption The LongTableBench QA pipeline with Qwen3-30B-A3B is a faithful downstream measure of effective answerability.
    Section 4.2 uses QA F1 as the downstream metric and defines Eff@k assuming Oracle QA is the ceiling and No-Table is near zero.
invented entities (1)
  • Content-level answerability (single-source, binary) no independent evidence
    purpose: Formal target property for table retrieval; used to label Target vs Sibling Distractor tables.
    Introduced as a definition in §3.1 and operationalized only through the authors' own benchmark construction; no external dataset independently validates the construct.

pith-pipeline@v1.3.0-alltime-deepseek · 31103 in / 15881 out tokens · 172954 ms · 2026-08-01T17:03:35.058063+00:00 · methodology

0 comments
read the original abstract

Tables are a critical knowledge source in retrieval-augmented generation (RAG), but a retrieved table may lack sufficient evidence to answer a query, a property we call answerability. While answerability broadly concerns whether a source or collection of sources contains sufficient evidence, retrieval models optimized for semantic relevance do not guarantee it even in the single-source case, creating a fundamental mismatch. To study this, we introduce TCR-Bench, a diagnostic benchmark for Table Content-level Answerability in RAG, built around sibling tables, i.e., tables with highly similar schemas but subtle content differences. On TCR-Bench, the dense retrievers we evaluate persistently exhibit a Semantic-Answerability Gap: they often retrieve the correct sibling group yet struggle to pinpoint the uniquely answerable table within it, dropping QA performance from 0.755 (oracle) to 0.330 (top-5 retrieved). Our analysis suggests this gap is associated with semantic accumulation, schema-level cue dependence, and weak row-column binding. As a diagnostic probe into the source of this gap, we test whether a lightweight two-stage pipeline, Answerability-Aware Reranking (AAR), applying direct query-table answerability judgment, can recover performance: it raises top-1 target retrieval from 18.2% to 57.4%, and this large gain is itself evidence that much of the observed failure reflects a missing answerability verification step, rather than an inherent limitation of model capacity alone.

Figures

Figures reproduced from arXiv: 2607.17742 by Gang Chen, Haobo Wang, Jiaming Tian, Junbo Zhao, Lihua Yu, Liyao Li, Wentao Ye, Zujie Ren.

Figure 1
Figure 1. Figure 1: Two sibling tables with nearly identical [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the TCR-Bench pipeline and retrieval performance. The left panel shows its construction, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Retrieval performance on TCR-Bench using Mixed Format. Abbreviated model names are used. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Intra-table retrieval by granularity and query [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

82 extracted references · 11 linked inside Pith

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year=

    Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task , author=. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year=

    2018

  10. [10]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models , author=. arXiv preprint arXiv:2506.05176 , year=

    Pith/arXiv arXiv

  11. [11]

    2025 , eprint=

    Jasper and Stella: distillation of SOTA embedding models , author=. 2025 , eprint=

    2025

  12. [12]

    2025 , eprint=

    jina-embeddings-v4: Universal Embeddings for Multimodal Multilingual Retrieval , author=. 2025 , eprint=

    2025

  13. [13]

    arXiv preprint arXiv:2308.03281 , year=

    Towards general text embeddings with multi-stage contrastive learning , author=. arXiv preprint arXiv:2308.03281 , year=

    Pith/arXiv arXiv

  14. [14]

    2024 , eprint=

    BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation , author=. 2024 , eprint=

    2024

  15. [15]

    2009 , isbn =

    The Probabilistic Relevance Framework: BM25 and Beyond , author =. 2009 , isbn =

    2009

  16. [16]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  17. [17]

    Qwen2 Technical Report , year=

  18. [18]

    2025 , eprint=

    gpt-oss-120b & gpt-oss-20b Model Card , author=. 2025 , eprint=

    2025

  19. [19]

    Findings of the Association for Computational Linguistics: EMNLP , volume=

    LongTableBench: benchmarking long-context table reasoning across real-world formats and domains , author=. Findings of the Association for Computational Linguistics: EMNLP , volume=

  20. [20]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Re-rag: Improving open-domain qa performance and interpretability with relevance estimator in retrieval-augmented generation , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  21. [21]

    ICLR 2024 Workshop on Large Language Model (LLM) Agents , year =

    Making Retrieval-Augmented Language Models Robust to Irrelevant Context , author=. ICLR 2024 Workshop on Large Language Model (LLM) Agents , year =

    2024

  22. [22]

    arXiv preprint arXiv:2512.14313 , year=

    Dynamic Context Selection for Retrieval-Augmented Generation: Mitigating Distractors and Positional Bias , author=. arXiv preprint arXiv:2512.14313 , year=

    arXiv

  23. [23]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Hoh: A dynamic benchmark for evaluating the impact of outdated information on retrieval-augmented generation , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  24. [24]

    ACM Transactions on Information Systems (TOIS) , volume=

    A similarity measure for indefinite rankings , author=. ACM Transactions on Information Systems (TOIS) , volume=. 2010 , publisher=

    2010

  25. [25]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Precise zero-shot dense retrieval without relevance labels , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  26. [26]

    arXiv preprint arXiv:2304.14233 , year=

    Large language models are strong zero-shot retriever , author=. arXiv preprint arXiv:2304.14233 , year=

    Pith/arXiv arXiv

  27. [27]

    Proceedings of the ACM on Management of Data , volume=

    Pneuma: Leveraging llms for tabular data representation and retrieval in an end-to-end system , author=. Proceedings of the ACM on Management of Data , volume=. 2025 , publisher=

    2025

  28. [28]

    2024 IEEE 40th International Conference on Data Engineering (ICDE) , pages=

    Datalore: Can a large language model find all lost scrolls in a data repository? , author=. 2024 IEEE 40th International Conference on Data Engineering (ICDE) , pages=. 2024 , organization=

    2024

  29. [29]

    Proceedings of the VLDB Endowment , volume=

    Auto-pipeline: synthesizing complex data pipelines by-target using reinforcement learning and search , author=. Proceedings of the VLDB Endowment , volume=. 2021 , publisher=

    2021

  30. [30]

    Proceedings of the VLDB Endowment , volume=

    Explaining dataset changes for semantic data versioning with explain-da-v , author=. Proceedings of the VLDB Endowment , volume=

  31. [31]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  32. [32]

    arXiv preprint arXiv:2404.10981 , year=

    A survey on retrieval-augmented text generation for large language models , author=. arXiv preprint arXiv:2404.10981 , year=

    Pith/arXiv arXiv

  33. [33]

    Compositional semantic parsing on semi-structured tables , author=. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=

  34. [34]

    Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    HybridQA: A dataset of multi-hop question answering over tabular and textual data , author=. Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    2020

  35. [35]

    Proceedings of the VLDB Endowment , volume=

    TableCopilot: A Table Assistant Empowered by Natural Language Conditional Table Discovery , author=. Proceedings of the VLDB Endowment , volume=. 2025 , publisher=

    2025

  36. [36]

    , author=

    THoRR: Complex Table Retrieval and Refinement for RAG. , author=. IR-RAG@ SIGIR , pages=

  37. [37]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Open domain question answering over tables via dense retrieval , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  38. [38]

    Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    Dense passage retrieval for open-domain question answering , author=. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    2020

  39. [39]

    Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Is table retrieval a solved problem? exploring join-aware multi-table retrieval , author=. Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  40. [40]

    NeurIPS 2024 Third Table Representation Learning Workshop , year=

    TARGET: Benchmarking Table Retrieval for Generative Tasks , author=. NeurIPS 2024 Third Table Representation Learning Workshop , year=

    2024

  41. [41]

    Proceedings of the workshop on structured and unstructured knowledge integration (SUKI) , pages=

    Table retrieval may not necessitate table-specific model design , author=. Proceedings of the workshop on structured and unstructured knowledge integration (SUKI) , pages=

  42. [42]

    Intelligent Computing-Proceedings of the Computing Conference , pages=

    Tabular embedding model (tem): Finetuning embedding models for tabular rag applications , author=. Intelligent Computing-Proceedings of the Computing Conference , pages=. 2025 , organization=

    2025

  43. [43]

    2019 IEEE international conference on big data (Big Data) , pages=

    Improved table retrieval using multiple context embeddings for attributes , author=. 2019 IEEE international conference on big data (Big Data) , pages=. 2019 , organization=

    2019

  44. [44]

    arXiv preprint arXiv:2504.09554 , year=

    Mixture-of-RAG: Integrating Text and Tables with Large Language Models , author=. arXiv preprint arXiv:2504.09554 , year=

    arXiv

  45. [45]

    arXiv e-prints , pages=

    HD-RAG: Retrieval-augmented generation for hybrid documents containing text and hierarchical tables , author=. arXiv e-prints , pages=

  46. [46]

    Proceedings of the 2018 world wide web conference , pages=

    Ad hoc table retrieval using semantic similarity , author=. Proceedings of the 2018 world wide web conference , pages=

    2018

  47. [47]

    IEEE Access , year=

    Enhancing rag performance by representing hierarchical nodes in headers for tabular data , author=. IEEE Access , year=

  48. [48]

    Transactions of the association for computational linguistics , volume=

    Lost in the middle: How language models use long contexts , author=. Transactions of the association for computational linguistics , volume=

  49. [49]

    arXiv preprint arXiv:2406.02536 , year=

    Mitigate position bias in large language models via scaling a single dimension , author=. arXiv preprint arXiv:2406.02536 , year=

    Pith/arXiv arXiv

  50. [50]

    arXiv preprint arXiv:2508.07479 , year=

    Positional biases shift as inputs approach context window limits , author=. arXiv preprint arXiv:2508.07479 , year=

    Pith/arXiv arXiv

  51. [51]

    Proceedings of the VLDB Endowment , volume=

    Observatory: Characterizing Embeddings of Relational Tables , author=. Proceedings of the VLDB Endowment , volume=. 2023 , publisher=

    2023

  52. [52]

    NeurIPS 2023 Second Table Representation Learning Workshop , year=

    Tabular Representation, Noisy Operators, and Impacts on Table Structure Understanding Tasks in LLMs , author=. NeurIPS 2023 Second Table Representation Learning Workshop , year=

    2023

  53. [53]

    2025 , url=

    Length-Induced Embedding Collapse in Transformer-based Models , author=. 2025 , url=

    2025

  54. [54]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    TaPas: Weakly supervised table parsing via pre-training , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  55. [55]

    ACM SIGMOD Record , volume=

    Turl: Table understanding through representation learning , author=. ACM SIGMOD Record , volume=. 2022 , publisher=

    2022

  56. [56]

    , author=

    WebTables: Exploring the power of tables on the web. , author=. Proc. VLDB Endow. , volume=

  57. [57]

    , author=

    Data curation at scale: the data tamer system. , author=. Cidr , volume=

  58. [58]

    Proceedings of the VLDB Endowment , volume=

    Table union search on open data , author=. Proceedings of the VLDB Endowment , volume=. 2018 , publisher=

    2018

  59. [59]

    arXiv preprint arXiv:2203.16714 , year=

    End-to-end table question answering via retrieval-augmented generation , author=. arXiv preprint arXiv:2203.16714 , year=

    Pith/arXiv arXiv

  60. [60]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Tablerag: A retrieval augmented generation framework for heterogeneous document reasoning , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  61. [61]

    Strich, Jan and Isgorur, Enes Kutay and Trescher, Maximilian and Biemann, Chris and Semmann, Martin , journal=. T

  62. [62]

    Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    T2-RAGBench: Text-and-Table Aware Retrieval-Augmented Generation , author=. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  63. [63]

    International Semantic Web Conference , pages=

    mmRAG: A modular benchmark for retrieval-augmented generation over text, tables, and knowledge graphs , author=. International Semantic Web Conference , pages=. 2025 , organization=

    2025

  64. [64]

    arXiv preprint arXiv:2504.01346 , year=

    RAG over Tables: Hierarchical Memory Index, Multi-Stage Retrieval, and Benchmarking , author=. arXiv preprint arXiv:2504.01346 , year=

    arXiv

  65. [65]

    Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

    Colbert: Efficient and effective passage search via contextualized late interaction over bert , author=. Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

  66. [66]

    2026 , note =

    The Geometry of Consolidation , author =. 2026 , note =

    2026

  67. [67]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    Rethinking tabular data understanding with large language models , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2024

  68. [68]

    Proceedings of the 17th ACM International Conference on Web Search and Data Mining , pages=

    Table meets llm: Can large language models understand structured table data? a benchmark and empirical study , author=. Proceedings of the 17th ACM International Conference on Web Search and Data Mining , pages=

  69. [69]

    Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=

    BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models , author=. Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=

  70. [70]

    arXiv preprint arXiv:2605.18007 , year=

    Semantic Reranking at Inference Time for Hard Examples in Rhetorical Role Labeling , author=. arXiv preprint arXiv:2605.18007 , year=

    Pith/arXiv arXiv

  71. [71]

    arXiv preprint arXiv:2508.21038 , year=

    On the theoretical limitations of embedding-based retrieval , author=. arXiv preprint arXiv:2508.21038 , year=

    arXiv

  72. [72]

    arXiv preprint arXiv:2602.05062 , year=

    Scaling Laws for Embedding Dimension in Information Retrieval , author=. arXiv preprint arXiv:2602.05062 , year=

    arXiv

  73. [73]

    arXiv preprint arXiv:2605.01302 , year=

    Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation , author=. arXiv preprint arXiv:2605.01302 , year=

    Pith/arXiv arXiv

  74. [74]

    Proceedings of the 2025 International ACM SIGIR conference on innovative concepts and theories in information retrieval (ICTIR) , pages=

    W-rag: Weakly supervised dense retrieval in rag for open-domain question answering , author=. Proceedings of the 2025 International ACM SIGIR conference on innovative concepts and theories in information retrieval (ICTIR) , pages=

    2025

  75. [75]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  76. [76]

    International conference on learning representations , volume=

    Self-rag: Learning to retrieve, generate, and critique through self-reflection , author=. International conference on learning representations , volume=

  77. [77]

    International Conference on Learning Representations , year=

    Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering , author=. International Conference on Learning Representations , year=

  78. [78]

    International Conference on Learning Representations , year=

    Representation Degeneration Problem in Training Natural Language Generation Models , author=. International Conference on Learning Representations , year=

  79. [79]

    International Conference on Machine Learning , pages=

    On the Embedding Collapse when Scaling up Recommendation Models , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  80. [80]

    Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Splade: Sparse lexical and expansion model for first stage ranking , author=. Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

Showing first 80 references.