Pith. sign in

REVIEW 3 major objections 5 minor 30 references

CORE-T: one LLM pass plus a join-compatibility cache retrieves coherent table sets for text-to-SQL.

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

2026-08-03 09:37 UTC pith:KCEXT7O4

load-bearing objection Solid, clearly-specified open-book table retriever; benchmark gains likely hold, but the compatibility cache's coverage of realistic join paths is unquantified and the abstract overclaims. the 3 major comments →

arxiv 2601.13111 v2 pith:KCEXT7O4 submitted 2026-01-19 cs.CL cs.AIcs.IR

CORE-T: COherent REtrieval of Tables for Text-to-SQL

Hassan Soliman , Vivek Gupta , Dan Roth , Iryna Gurevych This is my paper
classification cs.CL cs.AIcs.IR
keywords text-to-SQLmulti-table retrievaljoin-aware retrievalopen-book settingdense retrievalLLM table selectioncompatibility cacheexecution accuracy
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.

The paper argues that open-book text-to-SQL — answering a question from a large pool of unrelated tables with no database identifiers — fails with standard dense retrieval because high recall comes with many distractors, while join-aware baselines assume clean scoping or spend many LLM calls. CORE-T is a training-free pipeline that enriches each table offline with an LLM-written purpose, precomputes a compatibility score for table pairs using key-like column constraints, then at query time runs one LLM selection pass over the top-K candidates and adds back strongly compatible tables. The claim is that this returns fewer, more coherent, join-consistent table sets, improving table-selection F1 by up to 22.7 points over dense retrieval and raising multi-table SQL execution accuracy while cutting LLM token usage by 1.64–4.2×. If correct, the main practical consequence is that schema-agnostic retrieval over pooled enterprise tables can be made both cheaper and more accurate.

Core claim

On the paper's own terms, the contribution is a specific decomposition: move schema understanding offline by writing a short purpose description for every table and caching a compatibility score for each table pair; keep online inference to a single LLM call over only the top-K retrieved candidates, using the cached compatibility evidence as guidance; then restore any strongly compatible tables the LLM pruned. The compatibility cache is the load-bearing new object: it scores pairs only when at least one column is unique and the other column's values are a subset of it, combining header similarity and value overlap, and it records the best join columns. The reported experiments show that this

What carries the argument

The central mechanism is the offline table-compatibility cache, a pairwise score CS(t_i, t_j) computed from column headers (exact and semantic) and value overlap, gated by a hard key–foreign-key-like constraint: at least one column must be unique and the other must be a subset. The table–table score is the best valid column match, and the argmax join columns are cached. This cache supplies the join evidence that a single LLM selection call consumes, and it drives the additive restoration step that re-adds strongly compatible tables. Around it, the pipeline also relies on LLM-generated purpose metadata to disambiguate near-duplicate tables and a dense-retrieval top-K stage to guarantee recall

Load-bearing premise

The claim collapses if the compatibility cache cannot recognize the join paths a query needs: it only scores pairs where one column is unique and the other's values are a subset, and it estimates value overlap from five randomly sampled rows, so non-key joins, self-joins, many-to-many bridge joins, and multi-hop joins can receive a zero score and the pipeline can silently drop a required table.

What would settle it

Take the BIRD, Spider, or MMQA query set and partition the multi-table queries by the type of join required (key–foreign-key vs. non-key/bridge/multi-hop). If CORE-T's perfect recall on the non-key partition is no better than plain dense retrieval's, the compatibility cache — not the LLM selection — is the binding constraint. A cheaper probe: remove the hard uniqueness/subset constraint and rerun the joinability-accuracy comparison; the reported gap would be the measure of how much the constraint buys.

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

If this is right

  • If the central claim holds, open-book retrieval can be made join-aware with a single LLM call, removing the need for database identifiers or gold foreign keys.
  • Multi-table execution accuracy should improve most where queries require two or more tables, because coherent schema slices reduce the chance of missing a required join edge.
  • Token cost should stay roughly constant relative to the top-K candidate size, not the corpus size, making the method applicable to large pooled corpora.
  • The compatibility cache yields a free side product: inferred join columns for any table pair, reusable outside text-to-SQL in data-lake analytics.
  • The error analysis suggests that gains come primarily from removing distractor tables while holding recall, so the method's value should grow as corpora become more heterogeneous.

Where Pith is reading between the lines

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

  • The paper does not test it, but the same offline-cache plus one-shot selection scheme should transfer to retrieving other schema-agnostic artifacts (views, JSON collections, document stores) whenever a compatibility signal can be precomputed.
  • If the method is correct, the practical bottleneck shifts from retrieval architecture to the quality of the compatibility cache; a testable prediction is that retrieval F1 will degrade on query types whose join paths are non-key, self-joins, bridge tables, or multi-hop, exactly the cases the paper flags as limitations.
  • The 5-row sampling for value overlap is a tunable knob: with minimal added cost, sampling more rows or using schema-only signals could reduce the chance of missing shared keys, and an ablation varying sample size would reveal the sensitivity.
  • The reported gains concentrate in removing distractors while holding recall, so an enterprise data lake with cross-domain duplicates is the natural stress test for whether CORE-T's advantage survives in real deployments.

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 studies open-book text-to-SQL over pooled multi-source table collections without db_ids or gold foreign keys. It proposes CORE-T, a training-free pipeline that (i) offline enriches each table with LLM-generated purpose metadata and a dense index, (ii) precomputes a table–table compatibility cache using column uniqueness/subset constraints and value/header similarity, and (iii) online retrieves top-K tables by dense retrieval, performs a single LLM selection pass conditioned on relevance and cached compatibility edges, and then additively restores strongly compatible tables pruned by the LLM. The authors report table-selection F1 gains over DR/DRR/ReAct/JAR/ARM on pooled BIRD, SPIDER, and MMQA, multi-table execution-accuracy improvements across three SQL generators, token-efficiency gains, ablations, oracle bounds, and an error analysis.

Significance. If the experimental results are accepted, the paper makes a useful practical contribution: it shows that a lightweight, training-free, cache-based compatibility signal combined with a single LLM selection call can substantially improve the precision–recall balance of multi-table retrieval and translate into downstream execution gains, while reducing LLM token usage compared to agentic and MIP-based baselines. The work is clearly presented, ships prompts and implementation details, provides ablations against DR@10, includes oracle upper bounds, and makes code/data available. The main caveat is that the core compatibility mechanism is restricted to key–foreign-key-like equality joins; the paper's own Limitations section acknowledges this, and the broader claim of applicability to realistic enterprise open-book corpora needs additional evidence.

major comments (3)
  1. [Eq. (2) and §3.1/§3.2.3] The compatibility cache is load-bearing: it is the only join evidence given to the LLM selector and the exclusive trigger for the additive restoration step. Eq. (2) requires at least one unique column and a subset relation between columns, so non-equi joins, self-joins, many-to-many bridge joins, and multi-hop joins all receive CS=0. The Limitations section acknowledges this ('may miss other connections common in practice...'), but the paper does not quantify how often such join patterns occur in the gold multi-table queries of BIRD/SPIDER/MMQA, nor how often a necessary table is missed because its only edge to the selected set fails the validity constraint. Without this analysis, the strong F1 and EM gains may be an artifact of benchmark join simplicity, and the paper's claim of improvement for realistic open-book/enterprise corpora is not established. I request an empirical breakdown o
  2. [§4.1.4, §5, Appendix F.2] The pipeline has several hand-set free parameters — K=10, τ_comp=0.3, the 5-row sample size, the column-score weights in Eq. (3), and the 0.5 joinability threshold — and the text states only that they are 'kept fixed across all datasets,' not how they were chosen. If these values were selected after inspecting the same dev sets on which all results are reported, the headline numbers are selected-on-test. The paper should either specify a validation protocol (e.g., tune on one dataset and evaluate on the others, or report performance across a plausible range of τ_comp and K) or justify the hand-set choices by sensitivity analysis. This is important because the central claims are empirical and several compared baselines are evaluated with their own tuned hyperparameters.
  3. [Abstract] The first abstract states the method is evaluated 'Across Bird, Spider, MMQA, and Beaver' and claims 'up to 24.4 points in multi-table execution accuracy,' while Section 1, Section 4, and Tables 2–3 report only BIRD, SPIDER, and MMQA, with multi-table execution gains of up to 5.0 points on BIRD and 6.9 points on MMQA. The body also mentions BEAVER only as a future benchmark in Appendix D. The abstract either must be corrected to match the experiments or the BEAVER experiments must be included. This is a factual reporting inconsistency that affects reader trust.
minor comments (5)
  1. [§4.2 / Table 4] The text says CORE-T 'reduces input tokens by 4.0–5.0× and total selection tokens by 3.6–4.2×,' but the abstract claims '1.64–4.20× fewer total selection tokens.' Please unify the numbers and clarify whether the abstract refers to input, output, or total tokens.
  2. [§3.1 / Appendix A] Five rows are 'sampled uniformly at random without replacement,' but no seed or sampling protocol is stated for this step, while all LLM decoding uses seed 42. Please report the sampling seed/procedure for reproducibility; if the code already fixes it, say so explicitly.
  3. [Appendix F.2] Minor typographical/capitalization issues: 'corerationale' in §4.1.2, 'AS these were not available' in Appendix F.2, and inconsistent 'Bird' vs 'BIRD' in captions and text.
  4. [§4.2 / Table 8] The paper says text-embedding-3-large is strongest but chooses UAE-Large-V1 as default for open-source reproducibility; the rationale is understandable, but the sentence could be clearer about the trade-off and about whether any downstream comparisons change if text-embedding-3-large is used.
  5. [§4.2 / Table 3] The MMQA EM=1T column contains only 0.4% of queries, so the Oracle numbers are highly unstable; this is a known small-sample caveat and should be acknowledged in the text, not only in the table.

Circularity Check

0 steps flagged

No circularity: CORE-T's claims rest on external benchmark comparisons; the compatibility cache is a hand-defined heuristic, not a fitted target.

full rationale

The paper's derivation chain is empirical rather than definitional. The offline compatibility cache (Eqs. 2-4) is a hand-defined heuristic used as evidence in the single LLM selection call and as the trigger for the additive restoration step; however, the headline table-selection F1 and SQL execution EM (Tables 2-3) are measured against gold table sets and gold SQL execution, not against CS values. There is no equation in which the reported prediction is, by construction, equal to a fitted parameter or to the cache score. The only fixed constants are K=10, tau_comp=0.3, and the embedding choice, and they are fixed across datasets rather than fitted to test labels (Section F.2). The compatibility score is itself evaluated against gold joinability annotations in a separate diagnostic (Table 5), and the paper explicitly acknowledges its restrictive key-FK assumption: 'Our compatibility cache focuses on key-foreign-key-like joins... It may miss other connections common in practice, including non-equi joins, self-joins, many-to-many joins via bridge tables, and multi-hop joins that require intermediate tables.' That is a scope limitation, not a circular reduction. Self-citations (JAR, ARM, EnrichIndex) are used as baselines or prior building blocks; JAR/ARM are compared against rather than invoked as evidence for CORE-T's superiority, so no load-bearing self-citation chain exists. Thus the central claim is self-contained given its stated assumptions.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The method's effectiveness rests on several domain assumptions rather than hidden mathematical axioms. The important free parameters are K, tau_comp, the row-sample size, and the fixed score weights, all set by hand on the same dev benchmarks used for evaluation.

free parameters (6)
  • K (dense retrieval cutoff) = 10
    Selected after a DR@K sweep on dev (Table 8); all CORE-T results use K=10.
  • tau_comp (restoration threshold) = 0.3
    Hand-fixed across datasets 'to balance recall and number of tables'; no separate tuning split is reported.
  • Markdown row sample size = 5 rows
    Uniform random sample of 5 rows per table is used for all value-overlap and subset signals; its representativeness is assumed.
  • Column-pair score weights = 0.5 * Jaccard + 0.5 * name similarity
    The weights in Eq. 3 are hand-chosen; no sensitivity analysis is reported.
  • Joinable prediction threshold = CS > 0.5
    Used in Appendix B to binarize compatibility scores for the cache-quality evaluation.
  • Embedding model = UAE-Large-V1
    Chosen after comparing on MTEB and dev retrieval sweeps; text-embedding-3-large had higher perfect recall but was not used.
axioms (4)
  • domain assumption A 5-row uniform random sample is representative enough for Jaccard and subset signals.
    Used throughout for compatibility scoring; if sampled rows miss join keys, the cache underestimates joinability.
  • domain assumption Key-foreign-key-like joins (one unique column, one subset column) are the join structure needed to answer open-book text-to-SQL queries.
    The hard validity constraint in Eq. 2 excludes non-equi, many-to-many, and multi-hop joins; the Limitations admit this.
  • domain assumption Dropping db_id and pooling closed-book benchmark tables reproduces the open-book enterprise setting.
    The entire evaluation is constructed this way; the Limitations note real deployments are noisier and evolving.
  • domain assumption The instruction-tuned LLM follows the recall-oriented group-selection policy and emits parseable JSON.
    If the LLM over-prunes or fails to parse, the method falls back to DR@10; the paper reports <1% parse failures.

pith-pipeline@v1.3.0-alltime-deepseek · 25176 in / 15000 out tokens · 156199 ms · 2026-08-03T09:37:17.297048+00:00 · methodology

0 comments
read the original abstract

Realistic text-to-SQL workflows often require joining multiple tables. As a result, accurately retrieving the relevant set of tables becomes a key bottleneck for end-to-end performance. We study an open-book setting where queries must be answered over large, heterogeneous table collections pooled from many sources, without clean scoping signals such as database identifiers. Here, dense retrieval (DR) achieves high recall but returns many distractors, while join-aware alternatives often rely on extra assumptions and/or incur high inference overhead. We propose CORE-T, a scalable, training-free framework that enriches tables with LLM-generated purpose metadata and pre-computes a lightweight table-compatibility cache. At inference time, DR returns top-K candidates; a single LLM call selects a coherent, joinable subset, and a two-step additive adjustment stage restores strongly compatible tables. Across Bird, Spider, MMQA, and Beaver, CORE-T improves over DR by up to 22.7 points in table-selection F1 while returning up to 40% fewer tables, and by up to 24.4 points in multi-table execution accuracy, and uses 1.64-4.20x fewer total selection tokens than LLM-intensive baselines.

Figures

Figures reproduced from arXiv: 2601.13111 by Dan Roth, Hassan Soliman, Iryna Gurevych, Vivek Gupta.

Figure 1
Figure 1. Figure 1: Querying setups. Top: Closed-book: each query targets a known database. Bottom: Open-book: queries must be answered over integrated clusters of tables spanning various domains pooled from multiple database schemas. is confined to a small schema (Yu et al., 2018; Lee et al., 2021; Li et al., 2023; Zhong et al., 2017; Herzig et al., 2021; Chen et al., 2020; Wang et al., 2022). In contrast, data-lake and sema… view at source ↗
Figure 2
Figure 2. Figure 2: provides a high-level overview of CORE￾T. We delegate schema understanding into offline enrichment and caching, and keep online inference lightweight. 2 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation (EM≥2T ). We compare using only the first stage of our pipeline (DR@10) vs. full CORE￾T pipeline. All settings use UAE-Large-V1 embeddings and Llama-3.1-8B-Instruct as the table selector, varying only the SQL generator. 6 Conclusion We introduced CORE-T, a scalable, training￾free framework for open-book multi-table retrieval in text-to-SQL over pooled multi-source tables, where db_id and gold fore… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of table-table compatibility scoring. Given two candidate tables (e.g., Car Makers and Countries), we compute similarity based on column headers (exact and semantic) and column values (Jaccard overlap). Additional constraints (e.g., uniqueness or subset relations) refine the compatibility assessment. The maximum similarity score under these constraints is taken as the overall table-table compa… view at source ↗
Figure 5
Figure 5. Figure 5: Initial dense retrieval with enriched table embeddings. Offline, each table is serialized into Markdown (with 5-row samples) and augmented with an LLM-generated purpose description; the concatenated text is embedded and indexed. At inference time, the query is embedded in the same space and the top-K tables are retrieved by cosine similarity, forming the high-recall candidate set TK(q). Method Est. LLM cal… view at source ↗
Figure 6
Figure 6. Figure 6: Ablation: execution match on all queries (EMall). We compare using only the first stage of our pipeline (DR@10) against the full CORE-T pipeline. All settings use UAE-Large-V1 for retrieval and Llama-3.1-8B-Instruct as the table selector, varying only the SQL generator. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_6.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

30 extracted references · 1 canonical work pages

  1. [1]

    Peter Baile Chen, Fabian Wenz, Yi Zhang, Devin Yang, Justin Choi, Nesime Tatbul, Michael Cafarella, Çağatay Demiralp, and Michael Stonebraker. 2025 a . https://arxiv.org/abs/2409.02038 Beaver: An enterprise benchmark for text-to-sql . Preprint, arXiv:2409.02038

    Pith/arXiv arXiv 2025

  2. [2]

    Peter Baile Chen, Tomer Wolfson, Michael Cafarella, and Dan Roth. 2025 b . https://arxiv.org/abs/2504.03598 EnrichIndex : Using LLM s to enrich retrieval indices offline . In Proceedings of the Second Conference on Language Modeling ( COLM ) . COLM 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Peter Baile Chen, Yi Zhang, Mike Cafarella, and Dan Roth. 2025 c . https://doi.org/10.18653/v1/2025.acl-long.1463 Can we retrieve everything all at once? ARM : An alignment-oriented LLM -based retrieval method . In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 30298--30317, Vienna, A...

    work page doi:10.18653/v1/2025.acl-long.1463 2025

  4. [4]

    Peter Baile Chen, Yi Zhang, and Dan Roth. 2024. https://doi.org/10.18653/v1/2024.acl-long.148 Is table retrieval a solved problem? exploring join-aware multi-table retrieval . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2687--2699, Bangkok, Thailand. Association for Computationa...

    doi:10.18653/v1/2024.acl-long.148 2024

  5. [5]

    Zhiyu Chen, Mohamed Trabelsi, Jeff Heflin, Yinan Xu, and Brian D. Davison. 2020. https://doi.org/10.1145/3397271.3401044 Table search using a deep contextualized language model . In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval , pages 589--598, Virtual Event, China. Association for Computi...

    arXiv 2020

  6. [6]

    Tianji Cong, James Gale, Jason Frantz, H. V. Jagadish, and C a g atay Demiralp. 2023. https://www.cidrdb.org/cidr2023/papers/p75-cong.pdf WarpGate : A semantic join discovery system for cloud data warehouses . In Proceedings of the 13th Conference on Innovative Data Systems Research ( CIDR 2023) , Amsterdam, The Netherlands

    2023

  7. [7]

    Jonathan Herzig, Thomas M \"u ller, Syrine Krichene, and Julian Eisenschlos. 2021. https://doi.org/10.18653/v1/2021.naacl-main.43 Open domain question answering over tables via dense retrieval . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 512--519,...

    doi:10.18653/v1/2021.naacl-main.43 2021

  8. [8]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen - tau Yih. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.550 Dense passage retrieval for open-domain question answering . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769--6781, Online. A...

    doi:10.18653/v1/2020.emnlp-main.550 2020

  9. [9]

    Lai, Yeye He, and Surajit Chaudhuri

    Eugenie Y. Lai, Yeye He, and Surajit Chaudhuri. 2025. https://doi.org/10.14778/3734839.3734856 Auto-prep: Holistic prediction of data preparation steps for self-service business intelligence . Proc. VLDB Endow. , 18(7):2212--2225

    arXiv 2025

  10. [10]

    Chia-Hsuan Lee, Oleksandr Polozov, and Matthew Richardson. 2021. https://doi.org/10.18653/v1/2021.acl-long.176 K aggle DBQA : Realistic evaluation of text-to- SQL parsers . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2261--2273, Online. Association for Computational Linguistics

    doi:10.18653/v1/2021.acl-long.176 2021

  11. [11]

    Fangyu Lei, Jixuan Chen, Yuxiao Ye, Ruisheng Cao, Dongchan Shin, Hongjin Su, Zhaoqing Suo, Hongcheng Gao, Wenjing Hu, Pengcheng Yin, Victor Zhong, Caiming Xiong, Ruoxi Sun, Qian Liu, Sida Wang, and Tao Yu. 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/46c10f6c8ea5aa6f267bcdabcb123f97-Abstract-Conference.html Spider 2.0: Evaluating language...

    2025

  12. [12]

    Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, Xuanhe Zhou, Chenhao Ma, Guoliang Li, Kevin Chen-Chuan Chang, Fei Huang, Reynold Cheng, and Yongbin Li. 2023. https://papers.nips.cc/paper_files/paper/2023/hash/83fc8fab1710363050bbd1d4b8cc0021-Abstract-Datasets_and_Benchmarks.html Can LLM alrea...

    2023

  13. [13]

    Xianming Li and Jing Li. 2024. https://doi.org/10.18653/v1/2024.acl-long.101 A o E : Angle-optimized embeddings for semantic textual similarity . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1825--1839, Bangkok, Thailand. Association for Computational Linguistics

    doi:10.18653/v1/2024.acl-long.101 2024

  14. [14]

    Llama Team . 2024. https://doi.org/10.48550/arXiv.2407.21783 The L lama 3 herd of models . Preprint, arXiv:2407.21783. Author list credited to Llama Team; detailed contributors listed in the paper appendix

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783 2024

  15. [15]

    Meta AI . 2024. https://ai.meta.com/blog/meta-llama-3-1/ Introducing L lama 3.1: Our most capable models to date

    2024

  16. [16]

    Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. https://doi.org/10.18653/v1/2023.eacl-main.148 MTEB : Massive text embedding benchmark . In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2014--2037, Dubrovnik, Croatia. Association for Computational Linguistics

    doi:10.18653/v1/2023.eacl-main.148 2023

  17. [17]

    Miller, Ken Q

    Fatemeh Nargesian, Erkang Zhu, Ren \'e e J. Miller, Ken Q. Pu, and Patricia C. Arocena. 2019. https://doi.org/10.14778/3352063.3352116 Data lake management: Challenges and opportunities . Proc. VLDB Endow. , 12(12):1986--1989

    arXiv 2019

  18. [18]

    Qwen Team . 2025. https://doi.org/10.48550/arXiv.2412.15115 Qwen2.5 technical report . Preprint, arXiv:2412.15115

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2025

  19. [19]

    Zhiruo Wang, Zhengbao Jiang, Eric Nyberg, and Graham Neubig. 2022. https://aclanthology.org/2022.suki-1.5/ Table retrieval may not necessitate table-specific model design . In Proceedings of the Workshop on Structured and Unstructured Knowledge Integration (SUKI), pages 36--46, Seattle, USA. Association for Computational Linguistics

    2022

  20. [20]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Cl \'e ment Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, and Jamie Brew. 2020. https://doi.org/10.18653/v1/2020.emnlp-demos.6 Transformers: State-of-the-art natural language processing . In Proceedings of the 2020 Conference on Empirical Methods in Natural Lan...

    doi:10.18653/v1/2020.emnlp-demos.6 2020

  21. [21]

    Jian Wu, Linyi Yang, Dongyuan Li, Yuliang Ji, Manabu Okumura, and Yue Zhang. 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/794a425a2e47e05d29d30f79b79a692d-Abstract-Conference.html MMQA : Evaluating LLM s with multi-table multi-hop complex questions . In Proceedings of the Thirteenth International Conference on Learning Representations

    2025

  22. [22]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. https://iclr.cc/virtual/2023/oral/12647 ReAct : Synergizing reasoning and acting in language models . In International Conference on Learning Representations

    2023

  23. [23]

    Puxuan Yu, Luke Merrick, Gaurav Nuti, and Daniel Campos. 2024. https://arxiv.org/abs/2412.04506 Arctic-embed 2.0: Multilingual retrieval without compromise . Preprint, arXiv:2412.04506

    Pith/arXiv arXiv 2024

  24. [24]

    Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. https://doi.org/10.18653/v1/D18-1425 S pider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to- SQL task . In Proceedings of the 2018 Conference on Emp...

    doi:10.18653/v1/d18-1425 2018

  25. [25]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. https://doi.org/10.48550/arXiv.2506.05176 Qwen 3 embedding: Advancing text embedding and reranking through foundation models . arXiv preprint arXiv:2506.05176

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.05176 2025

  26. [26]

    Yi Zhang and Zachary G. Ives. 2020. https://doi.org/10.1145/3318464.3389726 Finding related tables in data lakes for interactive data science . In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data , pages 1951--1966, Portland, OR, USA. Association for Computing Machinery

    arXiv 2020

  27. [27]

    Victor Zhong, Caiming Xiong, and Richard Socher. 2017. https://arxiv.org/abs/1709.00103 Seq2SQL : Generating structured queries from natural language using reinforcement learning . Preprint, arXiv:1709.00103

    Pith/arXiv arXiv 2017

  28. [28]

    Erkang Zhu, Dong Deng, Fatemeh Nargesian, and Ren \'e e J. Miller. 2019. https://doi.org/10.1145/3299869.3300065 JOSIE : Overlap set similarity search for finding joinable tables in data lakes . In Proceedings of the 2019 International Conference on Management of Data, pages 847--864, Amsterdam, The Netherlands. Association for Computing Machinery

    arXiv 2019

  29. [29]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  30. [30]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...