TabReason: A Reinforcement Learning-Enhanced Reasoning LLM for Explainable Tabular Data Prediction

Greg Mori; Hossein Hajimirsadeghi; Lauren Kelly Zung; Tommy Xu; Xiangyu Sun; Zhitian Zhang

REVIEW 3 major objections 5 minor 22 references

TabReason: A Reinforcement Learning-Enhanced Reasoning LLM for Explainable Tabular Data Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 1.5-billion-parameter reasoning LLM, trained with group-relative policy optimization and three simple rewards, claims the highest weighted F1 score on 7 of 9 financial tabular benchmarks while generating natural-language explanations.

desk verdict A working RL-for-tabular recipe whose headline F1 result is mostly a majority-class artifact; the credible signal is on two balanced datasets. read the letter →

arxiv 2505.21807 v3 pith:XUJYHVDL submitted 2025-05-27 cs.LG

Tommy Xu , Zhitian Zhang , Xiangyu Sun , Lauren Kelly Zung , Hossein Hajimirsadeghi , Greg Mori This is my paper

classification cs.LG

keywords tabulardatareinforcementlearningreasoningLLMgrouprelativepolicyoptimizationexplainabilityfinancialpredictionclassimbalanceweightedF1

verification ladder T0 review T1 audit T2 compute T3 formal

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 proposes TabReason, a method that turns a small instruction-tuned LLM into an explainable tabular-data predictor by training it with reinforcement learning. The claim is that a 1.5-billion-parameter model trained with group-relative policy optimization and rewards for format, answer validity, and correctness can reach the highest weighted F1 score on seven of nine financial benchmarks, outperforming established and much larger LLMs. The point of coupling RL with generated reasoning is that the same model both predicts and explains, which addresses a gap left by gradient boosting and by earlier LLM tabular predictors that optimize accuracy alone. The paper presents this as the first integration of RL-trained reasoning LLMs for tabular data prediction, and it is explicit in the appendix that on several heavily imbalanced datasets the weighted-F1 advantage coincides with convergence to majority-class predictions.

What carries the argument

The object that carries the argument is the GRPO training loop combined with a textual reasoning format. For each prompt, a group of outputs is sampled from the current policy; each output receives a numeric reward (format 0.5, answer validity 0.5, correctness 1.0); the group mean and standard deviation convert these raw rewards into a relative advantage; and a PPO-style clipped objective with a KL penalty makes above-average outputs more likely. The reasoning text between the <reasoning> and </reasoning> tags is part of the output that reinforcement learning reshapes, so the explanation emerges from policy optimization rather than from a separately trained explanation module.

What would settle it

Compare every baseline and TabReason under Matthews correlation or balanced accuracy on all nine datasets; if TabReason no longer leads on ccf, ccfraud, Polish, Taiwan, Porto Seguro, and Travel Insurance, the central claim is reduced to matching a majority-class baseline on those datasets.

Watch

Extended reading notes

Core claim

On the paper's own terms, TabReason establishes that a relatively small reasoning LLM can be taught to predict tabular outcomes and to say why, using group-relative policy optimization rather than supervised fine-tuning alone. Starting from a 1.5-billion-parameter instruction-tuned model, the method samples groups of responses for each tabular prompt, gives each response a reward (0.5 for following the required XML format, 0.5 for a valid answer, 1.0 for a correct prediction), normalizes rewards within the group, and updates the policy so that outputs scoring above the group average become more likely. Across nine financial datasets covering credit scoring, fraud detection, financial distress, and claim analysis, TabReason attains the top weighted F1 score on seven datasets, and on the more balanced datasets its margins are large. The authors also report in Appendix C that on several imbalanced datasets the same training drifts toward predicting the majority class, with a Matthews correlation coefficient near zero (a balanced measure of binary classification quality), so the weighted-F1 results come with the paper's own warning.

Load-bearing premise

The load-bearing premise is that the paper's headline metric, weighted F1 (an averaging that gives more weight to the frequent class), is a meaningful summary of predictive quality on these heavily skewed datasets; if a model can score high simply by predicting the majority class, the claimed wins over other LLMs lose their force.

Editorial extensions

If this is right

If the claim holds, a 1.5-billion-parameter reasoning LLM trained with RL is a viable alternative to much larger LLMs for tabular prediction, at least when label distributions are reasonably balanced.
Free-text explanations can be obtained as a byproduct of reward-based training, which removes the need for curated rationale datasets.
Because the reward scheme is modular, the framework can be extended to consistency checks, cost-sensitive weighting, or external evaluators, all of which the paper identifies as open directions.
On the balanced datasets the method's edge is substantial, so the effect is not purely a metric artifact even though it is one on several imbalanced sets.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A correctness reward of 1.0 combined with group-normalized advantages will actively select majority-class predictions on imbalanced data, and the reported near-zero Matthews correlations are the visible signature of that collapse.
A direct next test is to replace or augment the correctness reward with a balanced metric such as Matthews correlation or balanced accuracy and see whether the seven-of-nine weighted-F1 wins survive.
The generated reasoning is currently unverified; a natural extension is to hold out predictions and measure whether the rationales distinguish correct from incorrect answers or merely rationalize a fixed prediction.
The paper's own table offers a falsifier: if every winning dataset outside Australian and LendingClub has near-zero Matthews correlation, the outperformance claim reduces to matching a majority-class baseline on those datasets.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this paper deserves a serious referee but not acceptance as is. The recipe is real—GRPO with format, validity, and correctness rewards on a 1.5B Qwen model, applied to financial tabular prediction—and as far as I can tell it is the first RL-plus-reasoning-LLM combination for tabular data. The training setup is described cleanly, and the appendix is refreshingly honest: they admit RL collapses to majority-class guessing on imbalanced data, and they report MCC alongside weighted F1. On balanced datasets, Australian and LendingClub, the MCC is 0.66 and 0.89 and the RL training clearly beats the base model. That part is credible and worth building on.

The problem is the central claim as printed. Abstract and Section 3.3 say "highest weighted F1 on 7/9 datasets" and "potential to outperform well-established LLMs." Their own Table 5 shows six of those datasets (ccf, ccfraud, Polish, Taiwan, Porto Seguro, Travel Insurance) with TabReason MCC at or near 0.00. Weighted F1 on a 95%-majority test set is nearly 1.0 for a constant "no" predictor, so a high weighted F1 with zero MCC is evidence of majority-class guessing, not predictive quality. The stress-test note is right: the outperform claim reduces to 2/9 datasets once you use a class-balanced metric. The paper is transparent about the MCC numbers, which is good, but the main text still overstates the headline.

Other soft spots: no comparison against XGBoost or CatBoost, which are the actual strong baselines for tabular data; no error bars or multiple seeds; no code release; and the explainability claim is unvalidated. They even report reasoning/answer inconsistencies in Appendix C, so the "human-understandable reasons" part needs evaluation, not just examples.

That said, I don't think this is a sloppy paper. The authors know exactly where the weaknesses are. The issue is a mismatched summary: the abstract sells a result that the authors themselves undercut. For peer review: send it out. A serious referee can push them to put MCC front and center, add GBDT baselines, run multiple seeds, and evaluate explanation faithfulness. That is a tractable revision path and the core idea—RL rewards beyond correctness for tabular reasoning—is worth exploring. I would not cite it in its current form.

Referee Report

3 major / 5 minor

Summary. The paper proposes TabReason, a 1.5B-parameter Qwen2.5-Instruct model fine-tuned with Group Relative Policy Optimization (GRPO) for binary classification on nine financial tabular datasets. The reward is a sum of format, validity, and correctness rewards, and the model is prompted to output a reasoning chain and a final answer. The authors report the highest weighted F1 on 7 of 9 datasets compared with FinBen baselines and provide qualitative examples of the reasoning. They also report Matthews Correlation Coefficients in Table 5, and Appendix C acknowledges that on imbalanced datasets the model collapses to the majority class and that reasoning can be inconsistent with the final answer.

Significance. The paper explores a timely combination of RL-trained reasoning LLMs and tabular prediction, and the idea of using GRPO with custom rewards for this setting is novel. If the claims were fully supported, the contribution could be useful to the community. The authors are transparent about the main failure mode (majority-class collapse) and include MCC, which is a strength. However, as it stands, the headline accuracy claim is an artifact of metric choice on imbalanced data, and the comparison with FinBen baselines is not controlled; the explainability contribution is also unevaluated. Hence the significance is not yet established.

major comments (3)

[Section 3.3 / Table 5 / Appendix C] The headline claim of highest weighted F1 on 7 of 9 datasets does not establish better predictive quality. On the six imbalanced datasets (ccf, ccfraud, Polish, Taiwan, Porto Seguro, Travel Insurance), TabReason's MCC in Table 5 is 0.00, while its weighted F1 is 0.91–1.00. Appendix C explicitly states that RL fine-tuning 'tends to converge on predicting the majority class' on these datasets. Weighted F1 is dominated by majority-class accuracy on such imbalanced test sets, so an MCC of zero means the predictions are uncorrelated with the true labels. The superiority claim is therefore supported only on the two more balanced datasets (Australian MCC 0.66, LendingClub MCC 0.89). The abstract and Section 3.3 should state this qualification, or the claim should be restricted to balanced datasets.
[Section 3.3 / comparison protocol] The comparison against FinBen baselines is not on equal footing. The paper states that scores for other LLMs were 'extracted from FinBen results' (Section 3.3), but it does not specify whether those baselines were fine-tuned on the same training splits or evaluated zero-shot/few-shot. Since TabReason is RL fine-tuned on the training data, the improvement may be due to additional training rather than the proposed RL-reasoning approach. The authors should either include fine-tuned baselines under an identical protocol or explicitly state the baseline protocol and qualify the claim of outperforming well-established LLMs.
[Abstract / Section 1 / Appendix C] The explainability claim is not empirically supported. The abstract and Section 1 state that the model provides 'human-understandable reasons' and 'inherently provides explainability,' but no evaluation of the reasoning quality is provided. Appendix C admits 'inconsistencies between the reasoning component and the final answer,' and the Conclusion defers evaluation of explanation quality to future work. The presence of reasoning text is not evidence of explainability; without a faithfulness or consistency analysis, or a human study, the contribution remains unsubstantiated.

minor comments (5)

[Table 3 / Table 5] In the ccf row, the entry '0.001.00' appears to be a formatting error; it should likely read '0.00 1.00' for Gemini and FinMA, respectively.
[Table 1 / Table 2 / Table 3] The fraud datasets are referred to inconsistently as 'Credit Card Fraud'/'ccFraud' in Table 1 and 'ccf'/'ccfraud' in Tables 2 and 3; please unify the names.
[Appendix A] In the Taiwan example, the prompt contains 'Operating Expense Rate: 9290000000.000', which appears to be a data error (probably 0.929), and the reasoning says '687%' instead of 0.687; please correct the example or its transcription.
[References] References Yang et al. 2024b and Yang et al. 2025 have the same arXiv identifier (2403.20208); please consolidate to avoid duplication.
[Section B] The exact number of RL epochs and the GRPO group size are not reported; add these hyperparameters for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the RL training loop and held-out evaluation are self-contained, and cited baselines are external.

full rationale

No circularity found. TabReason's training pipeline is a standard supervised-plus-RL loop: the correctness reward (1.0) is computed from known labels on the training partition, group-relative advantage comes from GRPO, and epoch selection uses weighted F1 on the validation split; the reported weighted F1 and MCC numbers are evaluated on held-out test splits taken from FinBen. The method does not fit any parameter to test outcomes, and no load-bearing premise is justified by self-citation: GRPO is cited to DeepSeekMath (Shao et al., 2024), the base model to Qwen2.5 (Yang et al., 2024a), and baseline scores to FinBen (Xie et al., 2024a), none of which are authored by the present paper's authors. The acknowledged limitation in Appendix C — that on highly imbalanced datasets (ccf, ccfraud, Polish, Taiwan, Porto Seguro, Travel Insurance) RL fine-tuning converges to majority-class prediction, as reflected by MCC ≈ 0.00 — is a validity concern for the weighted-F1 claim, not a circularity: the paper transparently reports both the metric and its failure mode. The observed reasoning/answer inconsistencies are similarly a quality gap, not a circular step. The derivation chain from rewards to test metrics is therefore self-contained, and the central claim, though weakened by the MCC evidence, is not circular.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper's central results rest on three soft assumptions: the RL algorithm works as advertised, the tabular serialization preserves the relevant signal, and the hand-designed reward is sufficient. The paper itself provides evidence against the third assumption. There are no invented physical or mathematical entities, but the reward scheme and metric choices are free parameters that materially affect the conclusions.

free parameters (3)

reward weights (format, validity, correctness) = 0.5, 0.5, 1.0
Chosen by hand in Section 2.2; no sensitivity analysis or ablation is reported.
inference temperature = 0.1
The training temperature is 0.7, but inference uses 0.1 after the authors found it more effective; a post hoc tuning decision.
best epoch by validation F1 = up to 50 epochs; best selected on validation
Appendix B reports selecting the epoch with best validation weighted F1; no multiple seeds or variance reported.

assumptions (3)

standard math GRPO, as defined in Shao et al. (2024), improves LLM reasoning when trained with outcome rewards.
Taken as established background; the paper does not re-derive the algorithm.
domain assumption Serializing tabular rows into the FinBen natural-language format preserves enough information for a small LLM to make valid predictions.
The prompt format is inherited from FinBen with no ablations on serialization choices.
ad hoc to paper A simple additive reward (format + validity + correctness) is a sufficient learning signal to obtain both accurate and explainable predictions.
The paper's own Appendix C shows that this reward leads to majority-class collapse on imbalanced datasets.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TabReason: A Reinforcement Learning-Enhanced Reasoning LLM for Explainable Tabular Data Prediction." pith.science (2026). https://pith.science/paper/XUJYHVDL

@misc{pith2026250521807,
  author       = {Pith},
  title        = {Pith review of: TabReason: A Reinforcement Learning-Enhanced Reasoning LLM for Explainable Tabular Data Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XUJYHVDL}},
  note         = {Machine review of arXiv:2505.21807}
}

read the original abstract

Predictive modeling on tabular data is the cornerstone of many real-world applications. Although gradient boosting machines and some recent deep models achieve strong performance on tabular data, they often lack interpretability. On the other hand, large language models (LLMs) have demonstrated powerful capabilities to generate human-like reasoning and explanations, but remain under-performed for tabular data prediction. In this paper, we propose a new approach that leverages reasoning-based LLMs, trained using reinforcement learning, to perform more accurate and explainable predictions on tabular data. Our method introduces custom reward functions that guide the model not only toward better prediction accuracy but also toward human-understandable reasons for its predictions. The proposed method is evaluated on financial benchmark datasets and compared against established LLMs.

Figures

Figures reproduced from arXiv: 2505.21807 by the authors.

**Figure 1.** Examples of model performance over epochs using the proposed RL fine-tuning model. C. Pitfalls We found two main pitfalls in the experiments describe in Section 3: • Imbalanced Labels: For datasets such as CCF, CCFraud, Polish, and Travel Insurance, where the labels are highly imbalanced, RL fine-tuning increases overall accuracy and weighted F1 score but tends to converge on predicting the majority class. We attemp… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 3 canonical work pages

[1]

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 format.date year duplicate empty "emp...
[2]

Arik, S. \"O . and Pfister, T. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 6679--6687, 2021

2021
[3]

Elephants never forget: Memorization and learning of tabular data in large language models

Bordt, S., Nori, H., Rodrigues, V., Nushi, B., and Caruana, R. Elephants never forget: Memorization and learning of tabular data in large language models. arXiv preprint arXiv:2404.06209, 2024

arXiv 2024
[4]

and Guestrin, C

Chen, T. and Guestrin, C. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pp.\ 785--794, 2016

2016
[5]

Empowering many, biasing a few: Generalist credit scoring through large language models

Feng, D., Dai, Y., Huang, J., Zhang, Y., Xie, Q., Han, W., Chen, Z., Lopez-Lira, A., and Wang, H. Empowering many, biasing a few: Generalist credit scoring through large language models. arXiv preprint arXiv:2310.00566, 2023

arXiv 2023
[6]

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

arXiv 2025
[7]

Tabllm: Few-shot classification of tabular data with large language models

Hegselmann, S., Buendia, A., Lang, H., Agrawal, M., Jiang, X., and Sontag, D. Tabllm: Few-shot classification of tabular data with large language models. In International Conference on Artificial Intelligence and Statistics, pp.\ 5549--5581. PMLR, 2023

work page 2023
[8]

Tabpfn: A transformer that solves small tabular classification problems in a second

Hollmann, N., M \"u ller, S., Eggensperger, K., and Hutter, F. Tabpfn: A transformer that solves small tabular classification problems in a second. arXiv preprint arXiv:2207.01848, 2022

arXiv 2022

Show all 22 references

[9]

S., and Lin, M

Liu, Z., Chen, C., Li, W., Qi, P., Pang, T., Du, C., Lee, W. S., and Lin, M. Understanding r1-zero-like training: A critical perspective. 0 (arXiv:2503.20783), March 2025. URL http://arxiv.org/abs/2503.20783. arXiv:2503.20783

2025 arXiv
[10]

V., and Gulin, A

Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V., and Gulin, A. Catboost: unbiased boosting with categorical features. Advances in neural information processing systems, 31, 2018

2018
[11]

Deepseekmath: Pushing the limits of mathematical reasoning in open language models

Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

2024 arXiv
[12]

Internlm: A multilingual language model with progressively enhanced capabilities, 2023

Team, I. Internlm: A multilingual language model with progressively enhanced capabilities, 2023

2023
[13]

Llama 2: Open foundation and fine-tuned chat models

Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

2023 arXiv
[14]

Pixiu: A large language model, instruction data and evaluation benchmark for finance

Xie, Q., Han, W., Zhang, X., Lai, Y., Peng, M., Lopez-Lira, A., and Huang, J. Pixiu: A large language model, instruction data and evaluation benchmark for finance. arXiv preprint arXiv:2306.05443, 2023

2023 arXiv
[15]

Finben: A holistic financial benchmark for large language models

Xie, Q., Han, W., Chen, Z., Xiang, R., Zhang, X., He, Y., Xiao, M., Li, D., Dai, Y., Feng, D., et al. Finben: A holistic financial benchmark for large language models. Advances in Neural Information Processing Systems, 37: 0 95716--95743, 2024 a

2024
[16]

Open-finllms: Open multimodal large language models for financial applications

Xie, Q., Li, D., Xiao, M., Jiang, Z., Xiang, R., Zhang, X., Chen, Z., He, Y., Han, W., Yang, Y., et al. Open-finllms: Open multimodal large language models for financial applications. arXiv preprint arXiv:2408.11878, 2024 b

2024 arXiv
[17]

Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024 a

2024 arXiv
[18]

Yang, H., Liu, X.-Y., and Wang, C. D. FinGPT : Open - Source Financial Large Language Models . June 2023. URL http://arxiv.org/abs/2306.06031. arXiv:2306.06031

2023
[20]

Unleashing the potential of large language models for predictive tabular tasks in data science

Yang, Y., Wang, Y., Li, Y., Sen, S., Li, L., and Liu, Q. Unleashing the potential of large language models for predictive tabular tasks in data science. 0 (arXiv:2403.20208), January 2025. URL http://arxiv.org/abs/2403.20208. arXiv:2403.20208

2025 arXiv
[21]

Finpt: Financial risk prediction with profile tuning on pretrained foundation models

Yin, Y., Yang, Y., Yang, J., and Liu, Q. Finpt: Financial risk prediction with profile tuning on pretrained foundation models. arXiv preprint arXiv:2308.00065, 2023

2023 arXiv
[22]

Dapo: An open-source llm reinforcement learning system at scale

Yu, Q., Zhang, Z., Zhu, R., Yuan, Y., Zuo, X., Yue, Y., Fan, T., Liu, G., Liu, L., Liu, X., Lin, H., Lin, Z., Ma, B., Sheng, G., Tong, Y., Zhang, C., Zhang, M., Zhang, W., Zhu, H., Zhu, J., Chen, J., Chen, J., Wang, C., Yu, H., Dai, W., Song, Y., Wei, X., Zhou, H., Liu, J., Ma...

2025 arXiv
[23]

Tablellama: Towards open large generalist models for tables

Zhang, T., Yue, X., Li, Y., and Sun, H. Tablellama: Towards open large generalist models for tables. arXiv preprint arXiv:2311.09206, 2023

2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

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 format.date year duplicate empty "emp...

[2] [2]

Arik, S. \"O . and Pfister, T. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 6679--6687, 2021

2021

[3] [3]

Elephants never forget: Memorization and learning of tabular data in large language models

Bordt, S., Nori, H., Rodrigues, V., Nushi, B., and Caruana, R. Elephants never forget: Memorization and learning of tabular data in large language models. arXiv preprint arXiv:2404.06209, 2024

arXiv 2024

[4] [4]

and Guestrin, C

Chen, T. and Guestrin, C. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pp.\ 785--794, 2016

2016

[5] [5]

Empowering many, biasing a few: Generalist credit scoring through large language models

Feng, D., Dai, Y., Huang, J., Zhang, Y., Xie, Q., Han, W., Chen, Z., Lopez-Lira, A., and Wang, H. Empowering many, biasing a few: Generalist credit scoring through large language models. arXiv preprint arXiv:2310.00566, 2023

arXiv 2023

[6] [6]

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

arXiv 2025

[7] [7]

Tabllm: Few-shot classification of tabular data with large language models

Hegselmann, S., Buendia, A., Lang, H., Agrawal, M., Jiang, X., and Sontag, D. Tabllm: Few-shot classification of tabular data with large language models. In International Conference on Artificial Intelligence and Statistics, pp.\ 5549--5581. PMLR, 2023

work page 2023

[8] [8]

Tabpfn: A transformer that solves small tabular classification problems in a second

Hollmann, N., M \"u ller, S., Eggensperger, K., and Hutter, F. Tabpfn: A transformer that solves small tabular classification problems in a second. arXiv preprint arXiv:2207.01848, 2022

arXiv 2022