Pith. sign in

REVIEW 3 major objections 5 minor 72 references

TableZoomer claims that schema-based zooming plus code execution lets a small, unfine-tuned LLM beat much larger baselines on very large tables.

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 ↗

TableZoomer compresses tables into schemas, zooms to query-relevant regions, and executes generated Python to answer questions, lifting Qwen3-8B accuracy by 19.34 points on DataBench and 25 points on TableBench Fact Checking.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely promising agent framework with large reported gains, but the results rest on small test sets and an untested schema-sampling assumption; worth reviewing with requests for robustness checks. the 3 major comments →

arxiv 2509.01312 v1 pith:AEDNCSO3 submitted 2025-09-01 cs.CL

TableZoomer: A Collaborative Agent Framework for Large-scale Table Question Answering

Sishi Xiong , Ziyang He , Zhongjiang He , Yu Zhao , Changzai Pan , Jie Zhang , Zhenhe Wu , Shuangyong Song
show 1 more author
Yongxiang Li
This is my paper
classification cs.CL
keywords table question answeringLLM agent frameworkprogram-of-thoughtstable schemaquery-aware table zoomingentity linkinglarge-scale tablesReAct paradigm
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 reading

This paper tries to establish that table question answering over industrial-scale tables does not require a large fine-tuned model or a fully verbalized table. The framework first distills each table into a structured schema, with column statistics, exemplar cell values, and semantic descriptions, reducing input complexity from rows-times-columns to just columns. A query planner then identifies relevant columns and rows, a refiner zooms the schema down to a query-focused sub-schema, and a code generator answers by writing and executing Python rather than reasoning in natural language. On a million-cell benchmark the framework lifts a small 8-billion-parameter model by 19.34 accuracy points over program-of-thought prompting, and by 25 points on a fact-checking benchmark, while using a fraction of the tokens. If right, the result means small general-purpose models become practical for large-table question answering without fine-tuning.

Core claim

The central claim is that the main bottleneck in LLM table reasoning is how the table is presented, not the model's parametric scale. TableZoomer replaces the fully verbalized table with a structured table schema built once per table, then runs query-aware zooming: column selection removes irrelevant features, and entity linking aligns query entities to cell values by longest-common-subsequence matching over a threshold of 0.6. The resulting sub-schema feeds a code-generating step that writes executable Python, with execution errors fed back for self-correction, inside a ReAct-style think-act-observe loop. The paper argues this combination explains its measured gains: ablation shows the sche

What carries the argument

The load-bearing object is the table schema: a JSON description built once per table by a Table Describer that records, for each column, its data type, statistical summaries (max, min, mean, median, or category frequencies), K randomly sampled cell values, J full-row records, and an LLM-written semantic annotation of the table and each column. The argument moves through three operations on this schema: the Query Planner decomposes the question and classifies it as column-only or row-column; the Table Refiner performs column selection and LCS-based entity linking to compress the schema into a query-relevant sub-schema; and the Code Generator turns that sub-schema into executable Python under

Load-bearing premise

The global schema is built from random samples and statistical summaries, so if the sample misses a distinctive value or a column's relevance, the zooming step can delete the column the answer needs, and later code generation cannot recover it.

What would settle it

Take a large table where the target row's key entity appears only once, in a cell value that the sampled cells and sampled full rows do not contain. Ask a row-column question that hinges on that entity; if TableZoomer fails to link it and returns a wrong or empty answer, the schema-sampling sufficiency premise is false.

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

If this is right

  • Very large row-count tables become queryable by ordinary instruction-tuned LLMs, because token complexity follows column count rather than cell count.
  • Smaller models can close much of the accuracy gap with larger ones on large-table QA, lowering deployment cost; the 8-billion-parameter variant outperforms the unaided 32-billion-parameter baseline on the large-table benchmark.
  • The framework is insensitive to table-serialization format for program-guided QA, varying by at most 0.57 accuracy points across four formats, whereas plain PoT varies by up to 9.18 points.
  • A fixed five-step collaborative workflow with conditional reflection achieves near-peak gains per LLM call, suggesting most of the benefit comes from schema representation and zooming rather than added iterations.

Where Pith is reading between the lines

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

  • Editorial inference: the schema-sampling design implies a sufficiency guarantee the paper does not state—if the sampled cell values and row records are too small to expose a rare value that the answer depends on, entity linking will miss it and no later step can recover it; this failure mode could be tested by sweeping K and J.
  • Editorial inference: the same zooming idea transfers to semi-structured extraction outside QA, such as fact-checking or report generation over spreadsheets, where the relevant columns are not known in advance.
  • Editorial inference: because the framework never needs full-table text, it may compose naturally with retrieval over many tables, treating each table's schema as an index entry—a direction the paper does not explore.
  • Editorial inference: the authors' future-work plan for a learned central scheduler choosing among built-in roles suggests the fixed workflow measured here may be a lower bound on what an adaptive version could achieve; that adaptive claim is not tested in the paper.
Share X Bluesky LinkedIn Reddit HN

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. TableZoomer is a multi-role LLM agent for table question answering. Instead of verbalizing the entire table, it builds a global schema from column statistics and sampled values, uses a Query Planner and Table Refiner to select query-relevant columns and link entities, and then generates and executes Python code under a ReAct loop. The paper reports large gains over PoT baselines: with Qwen3-8B, it improves by 19.34 points on DataBench, 25 points on TableBench Fact Checking, and 11.84 points on TableBench Numerical Reasoning, while on WikiTableQA it reaches 76.52% accuracy, slightly below TableMaster’s 77.95% and above PoTable’s 65.56%. The authors attribute the gains to schema-based complexity reduction, query-aware zooming, and program-of-thoughts with iterative reflection.

Significance. If the reported results are reliable, this is a practically valuable framework: a small, unfine-tuned LLM can handle very large tables by replacing full-table verbalization with a sampled schema and zooming into query-relevant substructures. The paper is also transparent in several respects: it uses official evaluation metrics, provides a public code link, and the ablation in Table 6 supports the incremental contribution of schema representation, column selection, entity linking, and ReAct. The claimed token-complexity reduction from O(M×N) to O(N) is definitionally correct for fixed sample sizes. However, the empirical evidence has important gaps—most notably the untested reliance on K randomly sampled cell values per column, the small test sets behind several headline numbers, and an internal inconsistency in the published Algorithm 1. These issues need to be addressed before the central claims can be accepted with confidence.

major comments (3)
  1. [§3.2, Algorithm 1, Tables 2–3] The global schema is built from K random cell values per column and J random full-row records. Neither K nor J is reported anywhere, and no ablation varies these values or compares against a schema containing all unique values. The Query Planner and Table Refiner can only select columns that are recognizable from this schema; if a distinctive value needed by a question is absent from the sample, the relevant column may be dropped before code generation. Because each ReAct iteration in Algorithm 1 replans against the same global schema, this evidence defect cannot be repaired in later iterations. The claimed 19.34-point and 25-point gains therefore rest on an untested data-sufficiency premise. An oracle-schema comparison and a sensitivity study over K and J would establish how much of the gain depends on this sampling choice.
  2. [Algorithm 1, lines 16–23] The termination logic is inverted. Line 17 tests whether the ReAct response contains “I have completed the task”; if true, the algorithm sets a further query and increments the round counter, while if false it breaks out of the loop. This is the opposite of the ReAct loop described in §3.3, where a completed task should terminate and an incomplete task should trigger another iteration. As written, the algorithm would either loop again after completion or break immediately on the first “not completed” response, making it impossible to run the published procedure faithfully. This must be corrected; it also makes the reported five-call minimum hard to reconcile with the pseudocode.
  3. [§4.1, §5.1.1, Tables 2–3] The headline numbers are point estimates on small test sets: DataBench has 522 questions over 15 tables, and TableBench Fact Checking has only 96 entries. No error bars, per-table standard deviations, or significance tests are reported. With 15 tables, per-table variance is likely substantial, and several category-level cells in Table 2 (e.g., list[category]) are based on fewer than 100 questions. The phrase “consistently enhances performance across all LLMs significantly” in §5.1.1 exceeds what the evidence supports. I would ask for bootstrap confidence intervals or per-table error bars, and at least a qualitative statement of variability across the 15 DataBench tables.
minor comments (5)
  1. [§3.2] The LCS “overlap rate” used for entity linking is not defined. Please specify the formula (e.g., LCS length normalized by the maximum of the two string lengths) and state whether the 0.6 threshold was tuned or chosen a priori.
  2. [§3.2 and Algorithm 1] The symbol K is used both for the number of random cell values per column in the Table Describer and for the ReAct round cap. This is confusing; consider renaming the ReAct cap (e.g., R) in the pseudocode and implementation details.
  3. [§5.2, Table 6] The text says that introducing “Table Zooming” improves accuracy from 74.33% to 86.40%, but that jump actually combines Column Selection and Entity Linking (Ala2 to Ala4); the individual contribution of Column Selection is 84.67% (Ala3). Please rephrase the decomposition to avoid over-attributing the gain to a single component.
  4. [§5.3.2] The module is called “Query Planner” in §3.2 but “Query Runner” in the efficiency discussion; also “TablaRAG” and “Datebench” are typos. Please align terminology and fix typos.
  5. [Figure 6] The caption mentions “column features contained in the table schema,” but the figure lacks a clear y-axis label. Clarify what is plotted (e.g., number of columns kept after zooming).

Circularity Check

0 steps flagged

No significant circularity: TableZoomer's gains are measured against external benchmarks and no fitted parameter is renamed as a prediction or derived by construction.

full rationale

The paper's central claims are empirical accuracy improvements on DataBench, TableBench, and WikiTableQA using official evaluation implementations and Exact Match accuracy. These benchmarks are external, and there is no indication that answers or test labels were used to set prompts, thresholds, or hyperparameters. The schema-sampling step (K cell values per column, J full-row records) is a data-sufficiency assumption, but it is a robustness/coverage concern rather than a circular derivation: the framework's outputs are not constructed from the reference answers, and a missed column would cause errors rather than force the reported numbers. The token-complexity reduction from O(M×N) to O(N) is a definitional property of representing a table by column-wise schema statistics and exemplars; it is not an empirical prediction derived from the data and therefore is not circular. The paper cites some prior work by its own authors (e.g., TeleChat, Table-R1), but these citations are used for baseline model selection and related-work context, not as load-bearing justification for the framework's claimed results. No step reduces to its own input by construction, and no fitted input is relabeled as a prediction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The framework introduces no new physical or conceptual entities. It relies on five software roles that are all LLM prompts, plus a set of hyperparameters (LCS threshold, sample sizes, round caps) whose values are chosen without sensitivity analysis. The central assumption is that an LLM can perform reliable schema-based planning and code generation, a premise that is not formally verified.

free parameters (4)
  • Entity linking overlap threshold = 0.6
    Used in Table Refiner and Algorithm 1 line 9 to accept LCS matches; no sensitivity analysis is provided, and results may change with this value.
  • K (random cell examples per column) = not specified
    K randomly selected cell values per column are used to build the schema (Section 3.2); the value is not reported and may affect schema quality.
  • J (random full-row examples) = not specified
    J full-row records are included in the schema (Section 3.2); the value is not reported.
  • ReAct reasoning round cap = 5
    Section 4.4 caps reasoning rounds at 5; this directly limits how many refinement iterations are available for hard questions.
axioms (4)
  • domain assumption LLMs can faithfully decompose questions and generate correct code from a schema-based description
    The whole framework depends on the Query Planner, Code Generator, and ReAct loop producing reliable outputs when prompted with schema and zoomed sub-schema (Sections 3.2-3.3).
  • domain assumption LCS overlap above 0.6 is a sufficient proxy for entity matching
    The Table Refiner uses this threshold to link query entities to table cells; no evidence is given that this holds across the evaluated domains.
  • domain assumption Pandas parsing preserves the structure of heterogeneous table files
    Table Describer parses spreadsheets with Pandas; the paper assumes this is lossless enough for schema construction and code execution.
  • standard math Official benchmark evaluation implementations are correct and comparable
    Results are computed with the official evaluation scripts of DataBench, TableBench, and WikiTableQA; the paper does not re-verify them.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TableZoomer: A Collaborative Agent Framework for Large-scale Table Question Answering." pith.science (2026). https://pith.science/paper/AEDNCSO3

@misc{pith2026250901312,
  author       = {Pith},
  title        = {Pith review of: TableZoomer: A Collaborative Agent Framework for Large-scale Table Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AEDNCSO3}},
  note         = {Machine review of arXiv:2509.01312}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While large language models (LLMs) have shown promise in the table question answering (TQA) task through prompt engineering, they face challenges in industrial applications, including structural heterogeneity, difficulties in target data localization, and bottlenecks in complex reasoning. To address these limitations, this paper presents TableZoomer, a novel LLM-powered, programming-based agent framework. It introduces three key innovations: (1) replacing the original fully verbalized table with structured table schema to bridge the semantic gap and reduce computational complexity; (2) a query-aware table zooming mechanism that dynamically generates sub-table schema through column selection and entity linking, significantly improving target localization efficiency; and (3) a Program-of-Thoughts (PoT) strategy that transforms queries into executable code to mitigate numerical hallucination. Additionally, we integrate the reasoning workflow with the ReAct paradigm to enable iterative reasoning. Extensive experiments demonstrate that our framework maintains the usability advantages while substantially enhancing performance and scalability across tables of varying scales. When implemented with the Qwen3-8B-Instruct LLM, TableZoomer achieves accuracy improvements of 19.34% and 25% over conventional PoT methods on the large-scale DataBench dataset and the small-scale Fact Checking task of TableBench dataset, respectively.

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

72 extracted references · 48 canonical work pages · 4 internal anchors

  1. [1]

    A Survey on Table Question Answering: Recent Advances

    Jin N, Siebert J, Li D, Chen Q.: A Survey on Table Question Answering: Recent Advances. Available from: https://arxiv.org/abs/2207.05270

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Available from: https: //arxiv.org/abs/2408.10548

    Ruan Y, Lan X, Ma J, Dong Y, He K, Feng M.: Language Modeling on Tabular Data: A Survey of Foundations, Techniques and Evolution. Available from: https: //arxiv.org/abs/2408.10548

    Pith/arXiv arXiv

  3. [3]

    Learn Beneficial Noise as Graph Augmentation

    Huang S, Xu Y, Zhang H, Li X.: Learn Beneficial Noise as Graph Augmentation. Available from: https://arxiv.org/abs/2505.19024

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Enhance vision-language alignment with noise

    Huang S, Zhang H, Li X. Enhance vision-language alignment with noise. In: Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence. AAAI’25/IAAI’25/EAAI’25. AAAI Press; 2025. Avail...

    work page doi:10.1609/aaai.v39i16.33918 2025

  5. [5]

    Available from: https://arxiv.org/abs/2408

    Zhang H, Xu Y, Huang S, Li X.: Data Augmentation of Contrastive Learning is Estimating Positive-incentive Noise. Available from: https://arxiv.org/abs/2408. 09929. 29

    work page

  6. [6]

    Available from: https://arxiv.org/abs/2306.07651

    Zhang H, Huang S, Guo Y, Li X.: Variational Positive-incentive Noise: How Noise Benefits Models. Available from: https://arxiv.org/abs/2306.07651

    Pith/arXiv arXiv

  7. [7]

    Positive-Incentive Noise

    Li X. Positive-Incentive Noise. IEEE Transactions on Neural Networks and Learning Systems. 2024 Jun;35(6):8708–8714. https://doi.org/10.1109/tnnls. 2022.3224577

    arXiv 2024

  8. [8]

    TableLlama: Towards Open Large Generalist Mod- els for Tables

    Zhang T, Yue X, Li Y, Sun H. TableLlama: Towards Open Large Generalist Mod- els for Tables. In: Duh K, Gomez H, Bethard S, editors. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). Mexico City, Mexico: Association for Computational Lingui...

    work page 2024

  9. [9]

    Available from: https://arxiv

    Su A, Wang A, Ye C, Zhou C, Zhang G, Chen G, et al.: TableGPT2: A Large Multimodal Model with Tabular Data Integration. Available from: https://arxiv. org/abs/2411.02059

    Pith/arXiv arXiv

  10. [10]

    StructLM: Towards Building Generalist Models for Structured Knowledge Grounding

    Zhuang A, Zhang G, Zheng T, Du X, Wang J, Ren W, et al. StructLM: Towards Building Generalist Models for Structured Knowledge Grounding. In: First Con- ference on Language Modeling; 2024. Available from: https://openreview.net/ forum?id=EKBPn7no4y

    work page 2024

  11. [11]

    Available from: https://arxiv.org/abs/2403.19318

    Zhang X, Luo S, Zhang B, Ma Z, Zhang J, Li Y, et al.: TableLLM: Enabling Tabular Data Manipulation by LLMs in Real Office Usage Scenarios. Available from: https://arxiv.org/abs/2403.19318

    Pith/arXiv arXiv

  12. [12]

    Available from: https://arxiv

    Wu Z, Yang J, Liu J, Wu X, Pan C, Zhang J, et al.: Table-R1: Region-based Reinforcement Learning for Table Understanding. Available from: https://arxiv. org/abs/2505.12415

    Pith/arXiv arXiv

  13. [13]

    Chain-of- Table: Evolving Tables in the Reasoning Chain for Table Understanding

    Wang Z, Zhang H, Li CL, Eisenschlos JM, Perot V, Wang Z, et al. Chain-of- Table: Evolving Tables in the Reasoning Chain for Table Understanding. In: The Twelfth International Conference on Learning Representations; 2024. Available from: https://openreview.net/forum?id=4L0xnS4GQM

    work page 2024

  14. [14]

    Large Language Models are Versatile Decomposers: Decomposing Evidence and Questions for Table-based Reasoning

    Ye Y, Hui B, Yang M, Li B, Huang F, Li Y. Large Language Models are Versatile Decomposers: Decomposing Evidence and Questions for Table-based Reasoning. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. SIGIR ’23. New York, NY, USA: Association for Computing Machinery; 2023. p. 174–184. Av...

    arXiv 2023

  15. [15]

    Available from: https://arxiv.org/abs/2501.19378

    Cao L, Liu H.: TableMaster: A Recipe to Advance Table Understanding with Language Models. Available from: https://arxiv.org/abs/2501.19378. 30

    Pith/arXiv arXiv

  16. [16]

    Binding Language Models in Symbolic Languages

    Cheng Z, Xie T, Shi P, Li C, Nadkarni R, Hu Y, et al. Binding Language Models in Symbolic Languages. In: The Eleventh International Conference on Learning Representations; 2023. Available from: https://openreview.net/forum? id=lH1PV42cbF

    work page 2023

  17. [17]

    Lever: Learning to verify language-to-code generation with execution

    Ni A, Iyer S, Radev D, Stoyanov V, Yih Wt, Wang SI, et al. Lever: Learning to verify language-to-code generation with execution. In: Proceedings of the 40th International Conference on Machine Learning (ICML’23); 2023

    work page 2023

  18. [18]

    TableRAG: Million-Token Table Understanding with Language Models

    Chen SA, Miculicich L, Eisenschlos JM, Wang Z, Wang Z, Chen Y, et al. TableRAG: Million-Token Table Understanding with Language Models. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page

  19. [19]

    Lost in the Middle: How Language Models Use Long Contexts

    Liu NF, Lin K, Hewitt J, Paranjape A, Bevilacqua M, Petroni F, et al. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics. 2024;12:157–173. https://doi.org/10. 1162/tacl a 00638

    work page 2024

  20. [20]

    MAC-SQL: A Multi- Agent Collaborative Framework for Text-to-SQL

    Wang B, Ren C, Yang J, Liang X, Bai J, Chai L, et al. MAC-SQL: A Multi- Agent Collaborative Framework for Text-to-SQL. In: Rambow O, Wanner L, Apidianaki M, Al-Khalifa H, Eugenio BD, Schockaert S, editors. Proceedings of the 31st International Conference on Computational Linguistics. Abu Dhabi, UAE: Association for Computational Linguistics; 2025. p. 540–...

    work page 2025

  21. [21]

    You Only Read Once (YORO): Learning to Internalize Database Knowledge for Text-to-SQL

    Kobayashi H, Lan W, Shi P, Chang S, Guo J, Zhu H, et al. You Only Read Once (YORO): Learning to Internalize Database Knowledge for Text-to-SQL. In: Chiruzzo L, Ritter A, Wang L, editors. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguis- tics: Human Language Technologies (Volume 1: Long ...

    work page 2025

  22. [22]

    T5-SR: A Unified Seq-to- Seq Decoding Strategy for Semantic Parsing

    Li Y, Su Z, Li Y, Zhang H, Wang S, Wu W, et al. T5-SR: A Unified Seq-to- Seq Decoding Strategy for Semantic Parsing. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    work page 2023

  23. [23]

    Available from: https://arxiv.org/abs/2211.12588

    Chen W, Ma X, Wang X, Cohen WW.: Program of Thoughts Prompting: Disen- tangling Computation from Reasoning for Numerical Reasoning Tasks. Available from: https://arxiv.org/abs/2211.12588

    Pith/arXiv arXiv

  24. [24]

    React: Synergizing reasoning and acting in language models

    Yao S, Zhao J, Yu D, Du N, Shafran I, Narasimhan K, et al. React: Synergizing reasoning and acting in language models. In: International Conference on Learn- ing Representations (ICLR); 2023. Available from: https://arxiv.org/pdf/2210. 03629. 31

    work page 2023

  25. [25]

    Neural Enquirer: Learning to Query Tables with Natural Language

    Yin P, Lu Z, Li H, Kao B.: Neural Enquirer: Learning to Query Tables with Natural Language. Available from: https://arxiv.org/abs/1512.00965

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Scalable Database-Driven KGs can help Text-to-SQL

    Li Z, Wu Z, Li M, He Z, Fang R, Zhang J, et al. Scalable Database-Driven KGs can help Text-to-SQL. In: Etcheverry L, Garcia VL, Osborne F, Pernisch R, editors. Proceedings of the ISWC 2024 Posters, Demos and Industry Tracks: From Novel Ideas to Industrial Practice co-located with 23nd International Semantic Web Conference (ISWC 2024), Hanover, Maryland, U...

    work page 2024

  27. [27]

    UCS-SQL: Uniting Content and Structure for Enhanced Semantic Bridging In Text-to- SQL

    Wu Z, Li Z, JieZhangChinaTele J, He Z, Yang J, Zhao Y, et al. UCS-SQL: Uniting Content and Structure for Enhanced Semantic Bridging In Text-to- SQL. In: Che W, Nabende J, Shutova E, Pilehvar MT, editors. Findings of the Association for Computational Linguistics: ACL 2025. Vienna, Austria: Asso- ciation for Computational Linguistics; 2025. p. 8156–8168. Av...

    work page 2025

  28. [28]

    MR-SQL: Multi-Level Retrieval Enhances Inference for LLM in Text-to-SQL

    Wu Z, Li Z, Li M, Zhang J, He Z, Yang J, et al. MR-SQL: Multi-Level Retrieval Enhances Inference for LLM in Text-to-SQL. DASFAA. 2025

    work page 2025

  29. [29]

    LogicalFactChecker: Leveraging Logical Operations for Fact Checking with Graph Module Network

    Zhong W, Tang D, Feng Z, Duan N, Zhou M, Gong M, et al. LogicalFactChecker: Leveraging Logical Operations for Fact Checking with Graph Module Network. In: Jurafsky D, Chai J, Schluter N, Tetreault J, editors. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Online: Association for Computational Linguistics; 2020. p....

    work page 2020

  30. [30]

    Available from: https://arxiv.org/abs/1809.08887

    Yu T, Zhang R, Yang K, Yasunaga M, Wang D, Li Z, et al.: Spider: A Large- Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task. Available from: https://arxiv.org/abs/1809.08887

    Pith/arXiv arXiv

  31. [31]

    Available from: https://arxiv.org/abs/2407

    Wu Z, Li Z, Zhang J, Li M, Zhao Y, Fang R, et al.: RB-SQL: A Retrieval-based LLM Framework for Text-to-SQL. Available from: https://arxiv.org/abs/2407. 08273

    work page

  32. [32]

    Sentence Segmentation and Punctuation for Ancient Books Based on Supervised In-context Training

    Wang S, Fu W, Li M, He Z, Li Y, Fang R, et al. Sentence Segmentation and Punctuation for Ancient Books Based on Supervised In-context Training. In: Sprugnoli R, Passarotti M, editors. Proceedings of the Third Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA) @ LREC-COLING-2024. Torino, Italia: ELRA and ICCL; 2024. p. 251–255...

    work page 2024

  33. [33]

    Available from: https://arxiv.org/abs/2201.11903

    Wei J, Wang X, Schuurmans D, Bosma M, Ichter B, Xia F, et al.: Chain-of- Thought Prompting Elicits Reasoning in Large Language Models. Available from: https://arxiv.org/abs/2201.11903. 32

    Pith/arXiv arXiv

  34. [34]

    Lemur: Log parsing with entropy sampling and chain-of-thought merging

    Zhang W, Guan X, Yunhong L, Zhang J, Song S, Cheng X, et al. Lemur: Log parsing with entropy sampling and chain-of-thought merging. arXiv preprint arXiv:240218205. 2024

    work page 2024

  35. [35]

    Available from: https://arxiv.org/abs/2205.10625

    Zhou D, Sch¨ arli N, Hou L, Wei J, Scales N, Wang X, et al.: Least-to-Most Prompting Enables Complex Reasoning in Large Language Models. Available from: https://arxiv.org/abs/2205.10625

    Pith/arXiv arXiv

  36. [36]

    icsBERTs: Optimizing Pre-trained Language Models in Intelligent Customer Service

    Liu S, Peng C, Wang C, Chen X, Song S. icsBERTs: Optimizing Pre-trained Language Models in Intelligent Customer Service. Procedia Computer Sci- ence. 2023;222:127–136. International Neural Network Society Workshop on Deep Learning Innovations and Applications (INNS DLIA 2023). https://doi. org/https://doi.org/10.1016/j.procs.2023.08.150

    work page doi:10.1016/j.procs.2023.08.150 2023

  37. [37]

    icsPLMs: Exploring Pre-trained Language Mod- els in Intelligent Customer Service (Student Abstract)

    Liu S, Wang C, Song S. icsPLMs: Exploring Pre-trained Language Mod- els in Intelligent Customer Service (Student Abstract). In: AAAI Conference on Artificial Intelligence; 2024. Available from: https://api.semanticscholar.org/ CorpusID:268717945

    work page 2024

  38. [38]

    Sentiment Analysis Technologies in AliMe - An Intelligent Assistant for E-commerce

    Song S, Wang C, Liu S, Chen H, Chen H, Bao H. Sentiment Analysis Technologies in AliMe - An Intelligent Assistant for E-commerce. Int J Asian Lang Process. 2020;30:2050016:1–2050016:20

    work page 2020

  39. [39]

    Available from: https: //arxiv.org/abs/2302.13971

    Touvron H, Lavril T, Izacard G, Martinet X, Lachaux MA, Lacroix T, et al.: LLaMA: Open and Efficient Foundation Language Models. Available from: https: //arxiv.org/abs/2302.13971

    Pith/arXiv arXiv

  40. [40]

    StructLM: Towards Building Generalist Models for Structured Knowledge Grounding

    Zhuang A, Zhang G, Zheng T, Du X, Wang J, Ren W, et al.: StructLM: Towards Building Generalist Models for Structured Knowledge Grounding. Available from: https://arxiv.org/abs/2402.16671

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [41]

    Available from: https://arxiv.org/abs/2402.14361

    Kong K, Zhang J, Shen Z, Srinivasan B, Lei C, Faloutsos C, et al.: OpenTab: Advancing Large Language Models as Open-domain Table Reasoners. Available from: https://arxiv.org/abs/2402.14361

    Pith/arXiv arXiv

  42. [42]

    Ai flow at the network edge

    Shao J, Li X. Ai flow at the network edge. IEEE Network. 2025

    work page 2025

  43. [43]

    AI Flow: Perspectives, Scenarios, and Approaches

    An H, Huang S, Huang S, Li R, Liang Y, Shao J, et al. AI Flow: Perspectives, Scenarios, and Approaches. arXiv preprint arXiv:250612479. 2025

    work page 2025

  44. [44]

    Question Answering over Tabular Data with DataBench: A Large-Scale Empir- ical Evaluation of LLMs

    Os´ es Grijalba J, Ure˜ na-L´ opez LA, Mart´ ınez C´ amara E, Camacho-Collados J. Question Answering over Tabular Data with DataBench: A Large-Scale Empir- ical Evaluation of LLMs. In: Calzolari N, Kan MY, Hoste V, Lenci A, Sakti S, Xue N, editors. Proceedings of the 2024 Joint International Conference on Com- putational Linguistics, Language Resources an...

    work page 2024

  45. [45]

    Tablebench: A compre- hensive and complex benchmark for table question answering

    Wu X, Yang J, Chai L, Zhang G, Liu J, Du X, et al. Tablebench: A compre- hensive and complex benchmark for table question answering. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39; 2025. p. 25497–25506. Available from: https://arxiv.org/pdf/2408.09174

    Pith/arXiv arXiv 2025

  46. [46]

    Compositional Semantic Parsing on Semi-Structured Tables

    Pasupat P, Liang P. Compositional Semantic Parsing on Semi-Structured Tables. In: Zong C, Strube M, editors. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Con- ference on Natural Language Processing (Volume 1: Long Papers). Beijing, China: Association for Computational Linguistics; ...

    work page 2015

  47. [47]

    Available from: https://arxiv.org/abs/2505.09388

    Yang A, Li A, Yang B, Zhang B, Hui B, Zheng B, et al.: Qwen3 Technical Report. Available from: https://arxiv.org/abs/2505.09388

    Pith/arXiv arXiv

  48. [48]

    Available from: https://arxiv.org/abs/2412.15115

    Qwen, :, Yang A, Yang B, Zhang B, Hui B, et al.: Qwen2.5 Technical Report. Available from: https://arxiv.org/abs/2412.15115

    Pith/arXiv arXiv

  49. [49]

    Available from: https://arxiv.org/abs/2407.21783

    Grattafiori A, Dubey A, Jauhri A, Pandey A, Kadian A, Al-Dahle A, et al.: The Llama 3 Herd of Models. Available from: https://arxiv.org/abs/2407.21783

    Pith/arXiv arXiv

  50. [50]

    Available from: https://qwenlm.github.io/blog/qwq-32b/

    Team Q.: QwQ-32B: Embracing the Power of Reinforcement Learning. Available from: https://qwenlm.github.io/blog/qwq-32b/

    work page

  51. [52]

    TeleChat: An Open- source Billingual Large Language Model

    Wang Z, Yao Y, Mengxiang L, He Z, Wang C, Song S, et al. TeleChat: An Open- source Billingual Large Language Model. In: Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10); 2024. p. 10–20

    work page 2024

  52. [53]

    Available from: https://arxiv.org/abs/2404.16645

    Li X, Yao Y, Jiang X, Fang X, Wang C, Liu X, et al.: Tele-FLM Technical Report. Available from: https://arxiv.org/abs/2404.16645

    Pith/arXiv arXiv

  53. [54]

    TeleChat: An Open- source Billingual Large Language Model

    Wang Z, Liuxz2@chinatelecom cn Lc, Liusx14@chinatelecom cn Lc, Yao Y, Huangyy121@chinatelecom cn Hc, Mengxiang L, et al. TeleChat: An Open- source Billingual Large Language Model. In: Wong KF, Zhang M, Xu R, Li J, Wei Z, Gui L, et al., editors. Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10). Bangkok, Thailand: Associ- a...

    work page 2024

  54. [55]

    TeleChat Technical Report

    Wang Z, Liu X, Liu S, Yao Y, Huang Y, He Z, et al. TeleChat Technical Report. CoRR. 2024;abs/2401.03804

    Pith/arXiv arXiv 2024

  55. [56]

    Available from: https://arxiv.org/abs/2507

    Wang Z, Liu X, Yao Y, Wang C, Zhao Y, Yang Z, et al.: Technical Report of TeleChat2, TeleChat2.5 and T1. Available from: https://arxiv.org/abs/2507. 34 18013

    work page

  56. [57]

    Available from: https://arxiv.org/abs/2412.04272

    Mao Q, Liu Q, Li Z, Cheng M, Zhang Z, Li R.: PoTable: Towards Systematic Thinking via Stage-oriented Plan-then-Execute Reasoning on Tables. Available from: https://arxiv.org/abs/2412.04272

    Pith/arXiv arXiv

  57. [58]

    TabSQLify: Enhancing Reasoning Capabilities of LLMs Through Table Decomposition

    Nahid MMH, Rafiei D. TabSQLify: Enhancing Reasoning Capabilities of LLMs Through Table Decomposition. In: Duh K, Gomez H, Bethard S, editors. Proceedings of the 2024 Conference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). Mexico City, Mexico: Association for Computat...

    work page 2024

  58. [59]

    Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction

    Deng Z, Chan C, Wang W, Sun Y, Fan W, ZHENG T, et al. Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction. ArXiv. 2024;abs/2404.14215

    Pith/arXiv arXiv 2024

  59. [60]

    5725–5737

    p. 5725–5737. Available from: https://aclanthology.org/2024.naacl-long. 320/

    work page 2024

  60. [61]

    Exploring the Impact of Table-to-Text Methods on Augmenting LLM-based Question Answering with Domain Hybrid Data

    Min D, Hu N, Jin R, Lin N, Chen J, Chen Y, et al. Exploring the Impact of Table-to-Text Methods on Augmenting LLM-based Question Answering with Domain Hybrid Data. In: North American Chapter of the Association for Com- putational Linguistics; 2024. Available from: https://api.semanticscholar.org/ CorpusID:267759856

    work page 2024

  61. [62]

    A Survey of Table Reasoning with Large Language Models

    Zhang X, Wang D, Dou L, Zhu Q, Che W. A Survey of Table Reasoning with Large Language Models. Frontiers Comput Sci. 2024;19:199348

    work page 2024

  62. [63]

    TaPERA: Enhancing Faithfulness and Inter- pretability in Long-Form Table QA by Content Planning and Execution-based Reasoning

    Zhao Y, Chen L, Cohan A, Zhao C. TaPERA: Enhancing Faithfulness and Inter- pretability in Long-Form Table QA by Content Planning and Execution-based Reasoning. In: Annual Meeting of the Association for Computational Linguistics

    work page

  63. [64]

    TeleAI at SemEval-2025 Task 8: Advancing Table Reasoning Framework with Large Language Models

    Xiong S, Li M, Wang D, Zhao Y, Zhang J, Pan C, et al. TeleAI at SemEval-2025 Task 8: Advancing Table Reasoning Framework with Large Language Models. In: Rosenthal S, Ros´ a A, Ghosh D, Zampieri M, editors. Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2025). Vienna, Austria: Association for Computational Linguistics; 2025....

    work page 2025

  64. [65]

    LLMSR@XLLM25: A Language Model-Based Pipeline for Structured Reasoning Data Construction

    Xing H, Liu X, Jiang Z, Yang Z, Yao Y, Wang Z, et al. LLMSR@XLLM25: A Language Model-Based Pipeline for Structured Reasoning Data Construction. In: Fei H, Tu K, Zhang Y, Hu X, Han W, Jia Z, et al., editors. Proceedings 35 of the 1st Joint Workshop on Large Language Models and Structure Modeling (XLLM 2025). Vienna, Austria: Association for Computational L...

    work page 2025

  65. [66]

    Available from: https://api.semanticscholar.org/CorpusID:271923510

    work page

  66. [67]

    Enhancing math reasoning ability of large language models via computation logic graphs

    Zhao D, Han D, Wu J, He Z, Ning B, Yuan Y, et al. Enhancing math reasoning ability of large language models via computation logic graphs. Knowledge- Based Systems. 2025;325:113905. https://doi.org/https://doi.org/10.1016/j. knosys.2025.113905

    arXiv 2025

  67. [68]

    Human Carrying Status in Visual Surveillance

    Tao D, Li X, Maybank SJ, Wu X. Human Carrying Status in Visual Surveillance. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06). vol. 2; 2006. p. 1670–1677

    work page 2006

  68. [69]

    Deep neural networks with Elastic Rectified Linear Units for object recognition

    Jiang X, Pang Y, Li X, Pan J, Xie Y. Deep neural networks with Elastic Rectified Linear Units for object recognition. Neurocomputing. 2018;275:1132–1139. https: //doi.org/https://doi.org/10.1016/j.neucom.2017.09.056

    doi:10.1016/j.neucom.2017.09.056 2018

  69. [70]

    Learning to Rank for Blind Image Quality Assessment

    Gao F, Tao D, Gao X, Li X. Learning to Rank for Blind Image Quality Assessment. IEEE Transactions on Neural Networks and Learning Systems. 2015;26(10):2275–2290. https://doi.org/10.1109/TNNLS.2014.2377181

    work page arXiv 2015

  70. [72]

    Two-Stage Learning to Pre- dict Human Eye Fixations via SDAEs

    Han J, Zhang D, Wen S, Guo L, Liu T, Li X. Two-Stage Learning to Pre- dict Human Eye Fixations via SDAEs. IEEE Transactions on Cybernetics. 2016;46(2):487–498. https://doi.org/10.1109/TCYB.2015.2404432

    work page arXiv 2016

  71. [73]

    Bayesian Tensor Approach for 3-D Face Modeling

    Tao D, Song M, Li X, Shen J, Sun J, Wu X, et al. Bayesian Tensor Approach for 3-D Face Modeling. IEEE Transactions on Circuits and Systems for Video Tech- nology. 2008;18(10):1397–1410. https://doi.org/10.1109/TCSVT.2008.2002825. 36

    work page arXiv 2008

  72. [2024]

    Available from: https://openreview.net/forum?id=41lovPOCo5

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.