Pith. sign in

REVIEW 4 major objections 6 minor 5 cited by

An 8-billion-parameter agent trained to plan, write code, and self-correct outperforms larger and prior table-reasoning systems on WikiTQ, TabMWP, and TabFact.

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

2026-08-04 23:51 UTC pith:NCJJGQU2

load-bearing objection A plausible, well-ablated agent for tool-augmented table reasoning whose headline SOTA margins are inflated by test-set hyperparameter selection; the method deserves referee time but not blind acceptance. the 4 major comments →

arxiv 2509.06278 v4 pith:NCJJGQU2 submitted 2025-09-08 cs.AI

TableMind: An Autonomous Programmatic Agent for Tool-Augmented Table Reasoning

Chuang Jiang , Mingyue Cheng , Xiaoyu Tao , Qingyang Mao , Jie Ouyang , Qi Liu This is my paper
classification cs.AI
keywords table reasoningprogrammatic agenttool-augmented agentreinforcement fine-tuningRAPOplan-action-reflectcode executionLLM agent
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 claims that a lightweight 8-billion-parameter model can be turned into an autonomous table-reasoning agent by training it to iterate through a plan–write-code–execute–reflect loop, instead of by prompting a large model once to flatten a table. TableMind first fine-tunes Qwen3-8B on 200 distilled, answer-correct tool-use trajectories, then applies reinforcement fine-tuning with a multi-objective reward (format, final-answer accuracy, and strategic tool use) and a new rank-aware objective, RAPO, that strengthens the update for trajectories whose model confidence disagrees with their reward. On WikiTQ, TabMWP, and TabFact the trained agent reports 76.82%, 99.27%, and 91.85%, surpassing prior baselines including Table-R1. The authors present these results as evidence that a two-stage imitation-then-exploration recipe can give a small open model the flexibility and self-correction that previously required larger, workflow-driven systems.

Core claim

TableMind is an 8B-parameter agent that solves table questions by running a multi-turn plan–action–reflect loop: it decomposes the task, writes and executes Python in a sandbox, reads the output or error, and adjusts. Training is two-stage: supervised fine-tuning on 200 distilled, answer-filtered trajectories establishes the tool-calling format; reinforcement fine-tuning maximizes a reward of format, exact-match accuracy, and a decaying tool-use term. The proposed RAPO objective, a rank-aware GRPO variant, multiplies each trajectory's group advantage when model confidence disagrees with reward. On WikiTQ, TabMWP, and TabFact it reports 76.82%, 99.27%, and 91.85%, beating six baselines includ

What carries the argument

RAPO (Rank-Aware Policy Optimization): a policy-gradient objective that multiplies each trajectory's group-normalized advantage by a pairwise weight γ_w,l = 1 + α·I[logP(o_w) < logP(o_l)] when a high-reward trajectory is less confident than a low-reward one, averaging those flags per trajectory; this makes confidence–reward misalignment the extra learning signal. Around it stand the SFT warm-up on 200 distilled, answer-filtered trajectories and the multi-objective reward (format + accuracy + decaying tool-use curriculum), which define what RAPO optimizes.

Load-bearing premise

The result rests on the 200 SFT trajectories being clean teaching examples: they come from an unnamed expert model and are kept only when the final answer matches ground truth, so if that expert had already seen the test tables, or the kept samples encode test answers, the benchmark numbers can be inflated by memorization rather than by the training method.

What would settle it

Check whether the 200 SFT trajectories or the unnamed teacher model overlap with the WikiTQ, TabMWP, and TabFact test sets; then retrain with a teacher that provably never saw those test tables and see whether the reported margins survive. A second control is to keep SFT but replace the teacher's final-answer filter with an execution-correctness filter; if the gap collapses, the gains are memorization rather than learning to reflect.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the reported numbers hold, a single 8B open model can match or beat agent pipelines that rely on much larger proprietary models, lowering the compute and API barriers for table reasoning.
  • The two-stage recipe—SFT on a small distilled set, then RL with a verifiable reward—becomes a template for teaching other tool-augmented skills (SQL generation, data plotting, spreadsheet manipulation) to small models.
  • RAPO's rank-aware weighting implies that where model confidence is poorly calibrated, extra gradient pressure on misaligned winner-loser pairs stabilizes training and accelerates convergence; this should transfer to other group-based policy optimization settings.
  • The finding that three interaction turns are optimal indicates that multi-step reflection helps but error propagation hurts beyond a point, so agent designs should cap interaction depth rather than maximize it.

Where Pith is reading between the lines

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

  • A direct test would run RAPO against a plain GRPO with identical SFT and reward on a non-table reasoning benchmark (math or code) to see whether confidence-reward misalignment is the general bottleneck RAPO targets.
  • The same two-stage recipe could be applied to other tool-augmented tasks—SQL generation, spreadsheet manipulation, data-viz code—where a verifiable reward exists, to test whether the plan-action-reflect loop is the transferable ingredient.
  • Because the SFT set is only 200 samples, scaling it (e.g., 500–1000 trajectories) might shift the RL ceiling; the paper's ablations hold sample count fixed, so the interaction between SFT size and RAPO gain is untested.

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

4 major / 6 minor

Summary. The manuscript proposes TableMind, an 8B-parameter table-reasoning agent built on Qwen3-8B, which operates in a multi-turn plan-action-reflect loop with code execution in a sandbox. Training is two-stage: supervised fine-tuning on 200 distilled trajectories, followed by reinforcement fine-tuning with a multi-objective reward (format, answer accuracy, and tool-use) and a proposed Rank-Aware Policy Optimization (RAPO) objective. On WikiTQ, TabMWP, and TabFact, the reported accuracies are 76.82%, 99.27%, and 91.85%, surpassing baselines including Table-R1. Ablations attribute the gains to SFT, RFT, the auxiliary tool reward, RAPO, instruction tuning, and model scaling.

Significance. If the empirical results hold, the paper makes a useful contribution: it demonstrates that a relatively small 8B model can be trained to autonomously write and execute code for table reasoning, and the two-stage SFT+RFT recipe with a fine-grained reward is clearly described. The release of code is a positive step for reproducibility. However, the headline SOTA claims are currently not fully supported because key hyperparameters were selected by evaluating directly on the test benchmarks, the distillation teacher is not identified, and all results are single-run point estimates. The method itself is interesting and the ablations are internally consistent, but the evaluation protocol needs substantial strengthening before the absolute performance claims can be accepted.

major comments (4)
  1. [§5.1.1, §5.6, Tables 2–3] Hyperparameters are selected on the test sets. Section 5.1.1 states 'All evaluations are conducted on the test sets,' and Section 5.6 uses Table 3 to choose Max Turns=3 and Temperature=1.0 because they perform best on WikiTQ, TabMWP, and TabFact. Likewise, Table 2 selects the SFT sample size and epoch by test-set performance. The final Table 1 numbers are therefore the best of a 3×3 grid plus the best SFT configuration, not an unbiased evaluation of a pre-specified model. This is load-bearing because the margins over the strongest baselines are only 1.96 points on WikiTQ and ~3 points on the other tasks; test-set selection of this size can plausibly account for such gaps. The authors should use a held-out validation split or pre-register the configuration, and report results for all configurations or at least for a fixed configuration chosen without test-set feedback.
  2. [§4.3, §5.1.3] The distillation teacher is unnamed and no contamination audit is provided. Section 4.3 says the SFT data are generated by an 'expert model' through knowledge distillation, filtered only by final-answer correctness, but the model's identity is never disclosed. Because only 200 trajectories are retained and the evaluation benchmarks are widely used in pretraining, the possibility of test-answer leakage cannot be ruled out. The authors should name the teacher, describe its training data, and report a contamination check (e.g., n-gram overlap between the 200 SFT trajectories and the test tables/questions/answers). Without this, the absolute numbers cannot be separated from possible leakage.
  3. [§4.4.1, Eq. (3)] The claimed 'implicit curriculum that encourages exploration early and efficiency later' is not implemented by Eq. (3). R_tool = e^{-rho s}(beta * I_success - C * N_turns^2) decays the entire auxiliary reward, including the efficiency penalty, as the global step s grows. Thus the relative weight of the turn penalty does not increase over time; both the success bonus and the penalty shrink together. If the intended behavior is to gradually shift from exploration to efficiency, the coefficient on the penalty needs its own time dependence, or the authors should clarify what 'curriculum' means. As written, the reward at best down-weights tool shaping late in training, which is the opposite of 'efficiency later.'
  4. [Tables 1–2, Figure 4] All reported results are single runs with no error bars or multiple seeds. The differences in Table 1 are small (e.g., 76.82 vs. 74.86 on WikiTQ), and the ablation differences in Figure 4 are also point estimates. Without variance or significance testing, the reader cannot judge whether the reported improvements are real or due to noise. The authors should report means and standard deviations over at least three seeds, or provide bootstrap confidence intervals, for both the main results and the ablations.
minor comments (6)
  1. [§5.1.3] 'Verl' should be written as 'veRL' (or the official framework name).
  2. [Table 1] The 'Relative Improvement' row mixes absolute percentage-point differences and relative percentages without labeling. For WikiTQ, 76.82−74.86 = 1.96 points, but the table reports +2.61, which appears to be the relative percentage. Clarify the unit.
  3. [Throughout] The benchmark name is inconsistent: 'WikiTQ' in most places but 'Wikitq' in Tables 1–3 and in Section 5.2. Use one spelling consistently.
  4. [Title page] The ACM template placeholder text ('Make sure to enter the correct conference title...') and the '©2018' copyright line remain in the manuscript; these should be removed or corrected before submission.
  5. [§4.4.2] Minor wording issues: 'removes the KL divergence' should be 'removing the KL divergence'; 'it employs' is used several times with unclear antecedents. Please proofread.
  6. [§5.5, Figure 5] The axes of Figure 5 are not labeled. The left panel appears to show 'Tool Calls Ratio' over training steps and the right 'Pass Ratio,' but the units and the meaning of the percentages should be stated in the caption.

Circularity Check

0 steps flagged

No circular derivation: TableMind's claims are empirical and self-contained; self-citations are not load-bearing.

full rationale

The paper's central claims are empirical: a two-stage SFT+RFT pipeline with the RAPO objective is trained on benchmark training sets and evaluated on held-out test sets. No target result is used as an input to the derivation. The SFT stage uses expert-model trajectories filtered by ground-truth correctness (Section 4.3), but this filtering is performed on training data and is a standard data-quality step, not a fitted parameter that reappears as the test result. RAPO (Eqs. 4-7) modifies GRPO/DAPO with a rank-aware advantage weighting; its claimed benefit is supported by ablations (Figure 3b, Figure 4), not by definition. The self-citations (PoTable [22], TableTime [27], and the survey [4]) appear as related work or baselines, and none carries a load-bearing assumption, uniqueness theorem, or ansatz that the present paper relies on. The test-set hyperparameter search in Section 5.6/Table 3 and the SFT-size selection in Section 5.3.1/Table 2 are genuine benchmark-validity concerns (model selection directly on test data, no error bars), but these are not circularity: the reported accuracies are empirical measurements, not quantities derived by construction from the tuned hyperparameters. Under the specified circularity criteria, no step in the derivation reduces to its own input.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central system depends on five hand-set or test-selected hyperparameters, an unnamed distillation teacher, and the assumption that execution feedback is trustworthy. No new physical or conceptual entity is introduced beyond the RAPO weighting scheme, which is an algorithmic construction rather than an invented entity.

free parameters (5)
  • Max tool-call turns = 3
    Chosen from Table 3 sensitivity analysis on test benchmarks; the reported 0.7682/0.9903/0.9233 correspond to this value. Test-set tuning makes the headline numbers post-selection.
  • Rollout temperature = 1.0
    Chosen from Table 3 on test benchmarks; higher temperature improves reported scores.
  • SFT sample size and epochs = 200 / 1
    Table 2 compares configurations and picks 200 samples with 1 epoch based on test accuracy; no separate validation split is reported.
  • Tool reward coefficients rho, C, beta = 0.05, 0.01, 0.5
    Set in Section 5.1.3 without sensitivity analysis; the ablation in Section 5.3.3 credits R_tool with performance, but these constants are hand-chosen.
  • RAPO alpha = not reported
    Equation 7 defines alpha as the reweighting intensity, but no value is given in Section 5.1.3, so the method is under-specified as published.
axioms (5)
  • domain assumption Exact match and label accuracy on the three benchmarks fully capture table reasoning quality.
    Used for R_acc in Equation 2 and for all reported metrics; ignores partial credit and semantic equivalence.
  • domain assumption The code interpreter's execution feedback is a reliable training signal without reward hacking.
    Section 4.4.1 assumes pass/fail execution and final-answer matching properly guide the policy; no analysis of degenerate or adversarial code is given.
  • ad hoc to paper The unnamed expert model used for distillation produces trustworthy trajectories and has not memorized benchmark test answers.
    Section 4.3 relies on knowledge distillation from an expert model that is never named; the SFT data is filtered only by final correctness.
  • domain assumption Serialized table content in the prompt preserves all information needed for the task.
    The prompt template in Section 4.2 places table content directly into the context, so large tables that overflow context are not handled; the paper's motivation mentions context overflow but the method does not address it.
  • standard math The GRPO/DAPO-style policy gradient objective in Equation 4 is valid for optimizing the designed reward.
    RAPO builds on cited GRPO and DAPO work; no proof of convergence is given beyond empirical reward curves.

pith-pipeline@v1.3.0-alltime-deepseek · 15240 in / 14233 out tokens · 151316 ms · 2026-08-04T23:51:02.690804+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of TableMind: An Autonomous Programmatic Agent for Tool-Augmented Table Reasoning." pith.science (2026). https://pith.science/paper/NCJJGQU2

@misc{pith2026250906278,
  author       = {Pith},
  title        = {Pith review of: TableMind: An Autonomous Programmatic Agent for Tool-Augmented Table Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NCJJGQU2}},
  note         = {Machine review of arXiv:2509.06278}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Table reasoning requires models to jointly perform comprehensive semantic understanding and precise numerical operations. Although recent large language model (LLM)-based methods have achieved promising results, most of them still rely on a single-turn reasoning paradigm that processes flattened tables in a single forward pass. This paradigm suffers from inherent limitations, including context overflow on large tables, weak sensitivity to continuous numerical values, and the absence of explicit tool-use and reflection. In this paper, we propose TableMind, a tuning-based autonomous programmatic table agent that simulates the human-like cognitive schema of the multi-turn interaction within a lightweight LLM. Instead of adopting a training-free workflow design, TableMind learns to internalize planning, action, and reflection through a principled two-stage training strategy. To bootstrap structured table reasoning capabilities, we construct and filter high-quality reasoning data for the supervised fine-tuning (SFT) stage. To enable precise code generation, we introduce a designed multi-perspective reward scheme and a novel optimization objective in the reinforcement learning (RL) stage. Extensive experiments on diverse benchmarks demonstrate that TableMind consistently outperforms previous baselines, validating the effectiveness of training autonomous agents to improve overall performance.

Figures

Figures reproduced from arXiv: 2509.06278 by Chuang Jiang, Jie Ouyang, Mingyue Cheng, Qi Liu, Qingyang Mao, Xiaoyu Tao.

Figure 1
Figure 1. Figure 1: TableMind mimics the human chain of thought by using a multi-turn plan-action-reflect loop to solve table tasks. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall training pipeline for TableMind. The process begins with Prompt Building and SFT to warm up the model [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of Reward Curves for Key Model Components. (a) SFT initialization provides a significant advantage over [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study of model components. Performance [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evolution of the Tool Calls Ratio (left) and Pass [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A case study demonstrates the model’s plan-action-reflect loop. By breaking down the problem into several steps and [PITH_FULL_IMAGE:figures/full_fig_p009_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.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records

    cs.CL 2026-05 unverdicted novelty 7.0

    Introduces EHR-ReasonCon benchmark with expert annotations and EHR-Inspector LLM framework for reasoning-intensive verification of consistency between clinical notes and structured tables in EHRs.

  2. AnomaMind: Agentic Time Series Anomaly Detection with Tool-Augmented Reasoning

    cs.LG 2026-02 reject novelty 6.0

    An agentic framework combining vision-based localization, tool-based evidence checking, and RL-trained final detection reports higher F1 than ten baselines on four benchmarks, under an unfair and possibly circular evaluation.

  3. TabClaw: An Interactive and Self-Evolving Agent for Spreadsheet Manipulation and Table Reasoning

    cs.CL 2026-06 unverdicted novelty 5.0

    TabClaw is an interactive LLM agent for spreadsheets that exposes editable plans, uses parallel specialist agents, streams ReAct loops, and distills skills from user feedback, reporting improved benchmark task completion.

  4. Agent-R1: A Unified and Modular Framework for Agentic Reinforcement Learning

    cs.CL 2025-11 conditional novelty 4.0

    A modular, algorithm-agnostic framework for multi-turn agentic RL that masks policy updates to agent tokens and validates on multi-hop QA with five RL algorithms.

  5. Toward Efficient Agents: Memory, Tool learning, and Planning

    cs.AI 2026-01 conditional novelty 3.0

    A survey that organizes efficiency techniques for LLM agents into memory, tool learning, and planning, and consolidates benchmarks and metrics for measuring cost-performance trade-offs.

Reference graph

Works this paper leans on

46 extracted references · 9 canonical work pages · cited by 5 Pith papers

  1. [1]

    Rami Aly, Zhijiang Guo, Michael Schlichtkrull, James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Oana Cocarascu, and Arpit Mittal. 2021. Feverous: Fact extraction and verification over unstructured and structured information. arXiv preprint arXiv:2106.05707(2021)

    Pith/arXiv arXiv 2021

  2. [2]

    Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2019. Tabfact: A large-scale dataset for table-based fact verification.arXiv preprint arXiv:1909.02164(2019)

    Pith/arXiv arXiv 2019

  3. [3]

    Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. 2022. Convfinqa: Exploring the chain of numerical reasoning in conversational finance question answering.arXiv preprint arXiv:2210.03849 (2022)

    Pith/arXiv arXiv 2022

  4. [4]

    Mingyue Cheng, Qingyang Mao, Qi Liu, Yitong Zhou, Yupeng Li, Jiahao Wang, Jiaying Lin, Jiawei Cao, and Enhong Chen. 2025. A survey on table mining with large language models: Challenges, advancements and prospects.Authorea Preprints(2025)

    work page 2025

  5. [5]

    Mingyue Cheng, Hao Zhang, Jiqian Yang, Qi Liu, Li Li, Xin Huang, Liwei Song, Zhi Li, Zhenya Huang, and Enhong Chen. 2024. Towards personalized evalua- tion of large language models with an anonymous crowd-sourcing platform. In Companion Proceedings of the ACM Web Conference 2024. 1035–1038

    work page 2024

  6. [6]

    Zhoujun Cheng, Haoyu Dong, Zhiruo Wang, Ran Jia, Jiaqi Guo, Yan Gao, Shi Han, Jian-Guang Lou, and Dongmei Zhang. 2021. Hitab: A hierarchical table dataset for question answering and natural language generation.arXiv preprint arXiv:2108.06712(2021)

    Pith/arXiv arXiv 2021

  7. [7]

    Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. 2023. Critic: Large language models can self-correct with tool-interactive critiquing.arXiv preprint arXiv:2305.11738(2023)

    Pith/arXiv arXiv 2023

  8. [8]

    Zihui Gu, Ju Fan, Nan Tang, Preslav Nakov, Xiaoman Zhao, and Xiaoyong Du

  9. [9]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    Pith/arXiv arXiv 2025

  10. [10]

    Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno, and Julian Martin Eisenschlos. 2020. TaPas: Weakly supervised table parsing via pre-training.arXiv preprint arXiv:2004.02349(2020)

    Pith/arXiv arXiv 2020

  11. [11]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531(2015)

    Pith/arXiv arXiv 2015

  12. [12]

    Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. 2025. Accurate predictions on small data with a tabular foundation model.Nature637, 8045 (2025), 319–326

    2025

  13. [13]

    Jingcheng Hu, Yinmin Zhang, Qi Han, Daxin Jiang, Xiangyu Zhang, and Heung- Yeung Shum. 2025. Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model.arXiv preprint arXiv:2503.24290 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of LLM agents: A survey.arXiv preprint arXiv:2402.02716(2024)

    Pith/arXiv arXiv 2024

  15. [15]

    Ziqi Jin and Wei Lu. 2023. Tab-cot: Zero-shot tabular chain of thought.arXiv preprint arXiv:2305.17812(2023)

    Pith/arXiv arXiv 2023

  16. [16]

    Mikhail V Koroteev. 2021. BERT: a review of applications in natural language processing and understanding.arXiv preprint arXiv:2103.11943(2021)

    Pith/arXiv arXiv 2021

  17. [17]

    Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, and Jian- Guang Lou. 2021. TAPEX: Table pre-training via learning a neural SQL executor. arXiv preprint arXiv:2107.07653(2021)

    Pith/arXiv arXiv 2021

  18. [18]

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025. Understanding r1-zero-like training: A critical perspective.arXiv preprint arXiv:2503.20783(2025)

    Pith/arXiv arXiv 2025

  19. [19]

    Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. 2022. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning.arXiv preprint arXiv:2209.14610(2022)

    Pith/arXiv arXiv 2022

  20. [20]

    Weizheng Lu, Jing Zhang, Ju Fan, Zihao Fu, Yueguo Chen, and Xiaoyong Du

    work page

  21. [21]

    Xinyuan Lu, Liangming Pan, Qian Liu, Preslav Nakov, and Min-Yen Kan. 2023. SCITAB: A challenging benchmark for compositional reasoning and claim verifi- cation on scientific tables.arXiv preprint arXiv:2305.13186(2023)

    Pith/arXiv arXiv 2023

  22. [22]

    Qingyang Mao, Qi Liu, Zhi Li, Mingyue Cheng, Zheng Zhang, and Rui Li. 2024. PoTable: Towards Systematic Thinking via Stage-oriented Plan-then-Execute Reasoning on Tables.arXiv preprint arXiv:2412.04272(2024)

    Pith/arXiv arXiv 2024

  23. [23]

    Anusri Pampari, Preethi Raghavan, Jennifer Liang, and Jian Peng. 2018. emrqa: A large corpus for question answering on electronic medical records.arXiv preprint arXiv:1809.00732(2018)

    Pith/arXiv arXiv 2018

  24. [24]

    Panupong Pasupat and Percy Liang. 2015. Compositional semantic parsing on semi-structured tables.arXiv preprint arXiv:1508.00305(2015)

    Pith/arXiv arXiv 2015

  25. [25]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

    Pith/arXiv arXiv 2024

  26. [26]

    Aofeng Su, Aowen Wang, Chao Ye, Chen Zhou, Ga Zhang, Gang Chen, Guangcheng Zhu, Haobo Wang, Haokai Xu, Hao Chen, et al. 2024. Tablegpt2: A large multimodal model with tabular data integration.arXiv preprint arXiv:2411.02059(2024)

    Pith/arXiv arXiv 2024

  27. [27]

    Jiahao Wang, Mingyue Cheng, Qingyang Mao, Yitong Zhou, Feiyang Xu, and Xin Li. 2024. TableTime: Reformulating Time Series Classification as Training-Free Ta- ble Understanding with Large Language Models.arXiv preprint arXiv:2411.15737 (2024)

    arXiv 2024

  28. [28]

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024. A survey on large language model based autonomous agents.Frontiers of Computer Science18, 6 (2024), 186345

    2024

  29. [29]

    Zhiruo Wang, Haoyu Dong, Ran Jia, Jia Li, Zhiyi Fu, Shi Han, and Dongmei Zhang

  30. [30]

    Zilong Wang, Hao Zhang, Chun-Liang Li, Julian Martin Eisenschlos, Vincent Perot, Zifeng Wang, Lesly Miculicich, Yasuhisa Fujii, Jingbo Shang, Chen-Yu Lee, et al. 2024. Chain-of-table: Evolving tables in the reasoning chain for table understanding.arXiv preprint arXiv:2401.04398(2024)

    Pith/arXiv arXiv 2024

  31. [31]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems35 (2022), 24824–24837

    2022

  32. [32]

    Zheyuan Yang, Lyuhao Chen, Arman Cohan, and Yilun Zhao. 2025. Table-r1: Inference-time scaling for table reasoning.arXiv preprint arXiv:2505.23621(2025)

    arXiv 2025

  33. [33]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR)

    2023

  34. [34]

    Han-Jia Ye, Si-Yang Liu, and Wei-Lun Chao. 2025. A closer look at tabpfn v2: Strength, limitation, and extension.arXiv preprint arXiv:2502.17361(2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Yunhu Ye, Binyuan Hui, Min Yang, Binhua Li, Fei Huang, and Yongbin Li. 2023. Large language models are versatile decomposers: Decomposing evidence and questions for table-based reasoning. InProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval. 174–184

    work page 2023

  36. [36]

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. 2025. Dapo: An open- source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Shiji Song, and Gao Huang. 2025. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?arXiv preprint arXiv:2504.13837(2025)

    Pith/arXiv arXiv 2025

  38. [38]

    Liangyu Zha, Junlin Zhou, Liyao Li, Rui Wang, Qingyi Huang, Saisai Yang, Jing Yuan, Changbao Su, Xiang Li, Aofeng Su, et al . 2023. Tablegpt: Towards unifying tables, nature language and commands into one gpt.arXiv preprint arXiv:2307.08674(2023)

    Pith/arXiv arXiv 2023

  39. [39]

    ChengXiang Zhai. 2024. Large language models and future of information retrieval: opportunities and challenges. InProceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 481– 490

    work page 2024

  40. [40]

    Shuo Zhang, Zhuyun Dai, Krisztian Balog, and Jamie Callan. 2020. Summarizing and exploring tabular data in conversational search. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 1537–1540

    work page 2020

  41. [41]

    Tianshu Zhang, Xiang Yue, Yifei Li, and Huan Sun. 2023. Tablellama: Towards open large generalist models for tables.arXiv preprint arXiv:2311.09206(2023)

    Pith/arXiv arXiv 2023

  42. [42]

    Xiaokang Zhang, Sijia Luo, Bohan Zhang, Zeyao Ma, Jing Zhang, Yang Li, Guan- lin Li, Zijun Yao, Kangli Xu, Jinchang Zhou, et al . 2024. Tablellm: Enabling tabular data manipulation by llms in real office usage scenarios.arXiv preprint arXiv:2403.19318(2024)

    Pith/arXiv arXiv 2024

  43. [43]

    Yilun Zhao, Lyuhao Chen, Arman Cohan, and Chen Zhao. 2024. TaPERA: Enhanc- ing faithfulness and interpretability in long-form table QA by content planning and execution-based reasoning. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 12824–12840

    work page 2024

  44. [2021]

    InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining

    Tuta: Tree-based transformers for generally structured table pre-training. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 1780–1790

    work page

  45. [2022]

    PASTA: table-operations aware fact verification via sentence-table cloze pre-training.arXiv preprint arXiv:2211.02816(2022)

    Pith/arXiv arXiv 2022

  46. [2025]

    Large language model for table processing: A survey.Frontiers of Computer Science19, 2 (2025), 192350

    work page 2025