Pith. sign in

REVIEW 3 major objections 5 minor 35 references

This paper claims that a text-to-SQL model can be trained to decide when to reason step-by-step and when to answer directly, and that doing so cuts inference cost by about a quarter without lowering accuracy.

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

2026-08-02 10:55 UTC pith:FBVEL5VD

load-bearing objection Adaptive reasoning for text-to-SQL is real and the SFT-stage results hold up, but the DPO-stage efficiency numbers rely on an unexplained baseline collapse that needs to be fixed. the 3 major comments →

arxiv 2607.22622 v1 pith:FBVEL5VD submitted 2026-06-17 cs.CL cs.AI

Learning When to Reason for Text-to-SQL via SFT and DPO

Soohyuk Jang , Jiheum Yeom , Nohil Park , Sang Hun Kim , Yoonyoung Choi , Kiwook Bae , Sungroh Yoon This is my paper
classification cs.CL cs.AI
keywords text-to-SQLchain-of-thoughtadaptive reasoningsupervised fine-tuningdirect preference optimizationinference efficiencyquery difficulty routing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a text-to-SQL model can be trained to make its own decision about whether a query needs multi-step reasoning or can be answered directly, rather than being forced into one mode. The authors label each training query by running the base model 16 times under both a no-reasoning prompt and a chain-of-thought prompt, then fine-tune with supervised learning and preference optimization under a prompt that asks the model to judge complexity itself. On two standard benchmarks, the resulting model matches or beats baselines that always reason or never reason, while cutting average output tokens by roughly a quarter and latency by 11–17%. If correct, this offers a practical way to lower the cost of SQL generation without sacrificing accuracy, and a template for adaptive reasoning in other structured tasks.

Core claim

The central claim is that reasoning effort in text-to-SQL can be treated as a learned behavior rather than a fixed prompting choice. AutoThinkSQL assigns each training instance a ground-truth mode by running 16 generation rollouts under a no-chain-of-thought prompt and a chain-of-thought prompt: the query is labeled no-reasoning only if every no-reasoning rollout already produces a correct SQL, and as reasoning if any reasoning rollout succeeds after the no-reasoning ones fail. The model is then fine-tuned on these labels using the auto-thinking prompt, which asks it to evaluate complexity and either emit SQL directly or produce a reasoning chain, followed by a DPO stage that builds preferen

What carries the argument

The central object is the auto-thinking mechanism: a format prompt that instructs the model to first assess query complexity, then either output SQL directly or produce a reasoning chain before the SQL. Learning is carried by a rollout-based mode-labeling procedure: 16 independent responses under no-reasoning and reasoning prompts determine whether an instance is labeled no-CoT (only if all 16 no-reasoning responses are correct) or CoT (if reasoning helps), and these labels become the SFT targets with the auto-thinking prompt as input. A subsequent DPO stage builds preference pairs from new rollouts, favoring no-reasoning responses when the no-reasoning mode is always correct and reasoning r

Load-bearing premise

The method assumes that 16 sampled responses per query from the base model reliably reveal whether that query can be answered correctly without chain-of-thought reasoning; if those samples misrepresent a query's true difficulty, the learned router will be calibrated to the wrong signal.

What would settle it

Run the same SFT+DPO pipeline with a much larger rollout budget (e.g., 64 or 128 per mode) and compare the mode labels: if a substantial fraction of queries change their assigned mode, the routing policy trained on 16-rollout labels is not learning true query difficulty, and the token savings would not generalize to that ground truth.

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

If this is right

  • AutoThinkSQL SFT+DPO reports the highest execution accuracy in three of four evaluation settings on Spider and BIRD, while matching the best baseline within one standard deviation in the fourth.
  • Average output tokens drop by 24.6% on Spider and 18.3% on BIRD compared to CoT-only SFT+DPO, with latency down 17.1% and 11.5%, so the efficiency gain does not come with an accuracy penalty in these experiments.
  • No-CoT-only training underperforms under majority voting because it lacks reasoning diversity, and CoT-only training becomes unstable under DPO; auto-thinking avoids both failure modes by letting the model choose per query.
  • The CoT trigger rate increases monotonically from Easy to Extra Hard on the Spider difficulty split, indicating the model's routing decisions align with query difficulty rather than being arbitrary.
  • Because auto-thinking uses rollouts from both modes, it retains more SFT and DPO training instances than either single-mode baseline, which may contribute to its consistent gains.

Where Pith is reading between the lines

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

  • A natural extension is to replace the hard 16-rollout threshold with a continuous difficulty score: models trained on finer-grained labels might route more precisely and retain even more training data.
  • The auto-thinking prompt is task-agnostic in spirit; the same SFT/DPO labeling scheme could be applied to other structured-output tasks such as code repair or information extraction, where only a fraction of inputs need deep reasoning.
  • The monotonic routing curve suggests the model learns an implicit difficulty classifier; probing which input features (schema size, join count, filter complexity) drive the CoT trigger could make the routing decision interpretable and testable.
  • If the 16-rollout labels are noisy, the reported gains may depend on the base model's sampling behavior; a sensitivity analysis with different rollout counts would show how stable the routing supervision is.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes AutoThinkSQL, a Text-to-SQL framework that trains a model to decide per query whether to generate a Chain-of-Thought (CoT) reasoning trace or directly output SQL. The method is implemented in two stages: a supervised fine-tuning (SFT) stage using mode labels derived from the base model's rollout correctness under CoT and No-CoT prompts, and a Direct Preference Optimization (DPO) stage that constructs preference pairs that encourage skipping reasoning on simple queries and using reasoning on complex ones. Evaluations on Qwen3-Coder-30B-A3B over Spider and BIRD dev sets report that AutoThinkSQL SFT+DPO achieves the highest execution accuracy in three of four settings and reduces output tokens by ~24.6%/18.3% and latency by ~17.1%/11.5% compared to CoT-only SFT+DPO. The paper also presents an analysis showing the model's CoT trigger rate increases with query difficulty.

Significance. If the results hold, the paper makes a practical contribution by showing that adaptive reasoning can be trained into a Text-to-SQL model with both SFT and DPO, yielding inference-cost savings without a consistent accuracy penalty. The methodology is clearly described, uses a strong backbone, and is evaluated under both greedy and majority-vote decoding. The efficiency gains are substantial and directly relevant to deployment. However, the significance is tempered by an unexplained collapse of the CoT-only SFT+DPO baseline on Spider Greedy, which the headline efficiency numbers are measured against, and by the lack of sensitivity analysis for the mode-assignment threshold. The routing analysis, while suggestive, is partly a consistency check of the training labels rather than an independent validation of difficulty alignment.

major comments (3)
  1. [Table 1, Section 4.2, Table 2, Abstract] The CoT-only SFT+DPO baseline collapses on Spider Greedy to 69.14±0.54 from 84.59±0.58 at the SFT stage, with a similar but smaller drop on BIRD Greedy (54.02 vs 56.88). This collapse is never explained in the main text, appendix, or Limitations. The efficiency savings highlighted in the abstract (24.6% tokens, 17.1% latency on Spider; 18.3%, 11.5% on BIRD) are computed relative to this same degraded model (Table 2), and the conclusion's claim that 'single-mode training degrades' under DPO is based on it. As written, the Pareto improvement over a well-trained CoT model is not established. Please analyze the cause of the collapse (e.g., DPO instability, preference construction, evaluation artifact), provide a stronger CoT-only DPO baseline (e.g., with tuned β or a more robust preference setup), and recompute the efficiency savings relative to a healthy CoT baseline. Without this, the cent
  2. [Section 3.2, Appendix B, Table 3] The mode assignment in SFT labels a query as No-CoT only if all 16 No-CoT rollouts are correct (N_NC=16). This threshold is a free parameter, and the paper provides no sensitivity analysis. The learned routing policy—and hence the reported token/latency savings—depends directly on this choice. If 16 samples are noisy or the threshold is too strict/loose, the routing behavior could be miscalibrated. Please report results for at least one or two alternative thresholds (e.g., N_NC=12 or a majority criterion) or otherwise justify why N_NC=16 is a stable and appropriate choice. This is important to assess whether the method's success is contingent on an untested hyperparameter.
  3. [Section 4.3, Figure 1] The routing analysis in Figure 1 is presented as evidence that the model 'learns to align its reasoning decisions with query difficulty.' However, the training procedure explicitly labels queries based on the base model's ability to solve them without CoT (Section 3.2), so a correlation with Spider difficulty is largely a reflection of the training signal rather than an emergent, independent behavior. The analysis would be more convincing if it were framed as a consistency check, or if an independent measure (e.g., human-annotated complexity, or accuracy of the base model on held-out queries not used for labeling) were used. Please reframe or strengthen this analysis.
minor comments (5)
  1. [Abstract, Table 1] The abstract claims 'consistent gains compared to the best counterpart baseline on both Spider and BIRD.' The paper's own Table 1 shows AutoThinkSQL SFT+DPO is 0.29 points below CoT-only SFT on Spider Maj@8 (86.78 vs 87.07), a gap that is not clearly within one standard deviation if the given error bars are taken at face value. Suggest rewording to 'matches or exceeds' or 'consistent gains in most settings.'
  2. [Section 4.3 vs Abstract/Table 2] Section 4.3 states AutoThinkSQL SFT+DPO uses '24.9% fewer tokens' on Spider, but the abstract and Table 2 round to 24.6%. Please reconcile the numbers.
  3. [Table 2 and Table 1 captions] The efficiency table reports 'Lat.' and 'Tok.' but does not explicitly define these as per-generation latency and output token count. Clarify in the caption. Also, consider reporting total end-to-end time if different from generation-only latency.
  4. [Appendix B, Table 3] AutoThinkSQL retains more SFT training instances (5981) than CoT-only (5873) and No-CoT-only (5352). The paper does not discuss how this data-size imbalance might affect the comparison. A matched-data ablation (e.g., training on equal-sized subsets) would strengthen the claim that the gains come from the auto-thinking mechanism rather than more data.
  5. [General] The paper does not report statistical significance tests (e.g., paired bootstrap) for the accuracy differences. Given the small margins in some settings, such tests would help the reader assess whether the improvements are reliable.

Circularity Check

0 steps flagged

No significant circularity; central accuracy and efficiency claims are measured on held-out dev sets and do not reduce to the training labels.

full rationale

The paper's central claims are empirical and held-out. AutoThinkSQL is trained on BIRD training data using mode labels derived from 16 base-model rollouts per query, and then evaluated on Spider and BIRD dev sets. Accuracy and token/latency numbers in Tables 1 and 2 are measurements of the resulting model, not derivations from the labeling rule. The routing analysis in Figure 1 is somewhat self-confirmatory, since the model is explicitly trained to skip reasoning when No-CoT rollouts were correct, but it is evaluated on Spider difficulty levels that were not used in training and therefore retains independent content as a generalization check rather than a circular prediction. No load-bearing self-citation appears: the references to prior adaptive-reasoning work (AdaCoT, Thinkless, AdaptThink) are background, not justifications of this paper's results. The unexplained collapse of the CoT-only SFT+DPO baseline on Spider Greedy (69.14 vs 84.59 at SFT stage) is a serious correctness and reproducibility concern, and the abstract's 'consistent gains' is overstated given that AutoThinkSQL SFT+DPO trails CoT-only SFT on Spider Maj@8 (86.78 vs 87.07). However, these are issues of baseline health, evaluation reporting, and claim calibration, not circularity: the reported numbers are not constructed to equal their inputs. Accordingly, no circular step satisfying the evidence bar is present, and the score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim relies on hyperparameters (threshold, rollouts, beta) that are set by hand, and on domain assumptions about the validity of EX and the transferability of the routing policy. No new physical or theoretical entities are introduced.

free parameters (4)
  • mode assignment threshold N_NC=16 = 16 (exact)
    A query is labeled No-CoT only if all 16 rollouts under pNC are correct. This hand-chosen threshold determines the routing policy; a lower threshold would yield more no-CoT labels and more token savings but potentially lower accuracy on harder queries.
  • rollout count per mode = 16
    Number of independent rollouts used to estimate mode success; affects label reliability.
  • DPO beta = 0.3
    Standard DPO hyperparameter controlling deviation from the reference policy, chosen by hand.
  • LoRA rank and alpha = r=16, alpha=32
    Standard LoRA hyperparameters, chosen by hand.
axioms (4)
  • domain assumption Execution accuracy (EX) on Spider and BIRD dev is a valid measure of Text-to-SQL quality
    The central claim is measured in EX; if EX is not representative of real-world utility, the claim weakens.
  • domain assumption The base model's 16-rollout correctness is a reliable proxy for query difficulty under the two modes
    Mode labels (Section 3.2) are derived from whether 16 sampled responses are correct. This assumes sampling noise is small enough that the labels are stable.
  • domain assumption Training on BIRD and evaluating on Spider/BIRD dev is a fair transfer test
    The method assumes the learned routing policy transfers across domains; Spider is cross-domain, which is standard practice.
  • ad hoc to paper The prompt templates in Appendix C are sufficient to elicit the intended behavior (complexity assessment)
    The pAuto prompt is designed by the authors; if it fails to elicit true complexity assessment, the learned routing may be an artifact of the prompt.

pith-pipeline@v1.3.0-alltime-deepseek · 9848 in / 12001 out tokens · 101233 ms · 2026-08-02T10:55:56.284383+00:00 · methodology

0 comments
read the original abstract

Recent Text-to-SQL methods rely heavily on reasoning-centric paradigms such as Chain-of-Thought (CoT), achieving substantial gains on complex benchmarks at the cost of high inference-time overhead. However, a large fraction of real-world queries are simple lookups or aggregations that can be resolved without multi-step deduction, making forced reasoning wasteful. Thus, we propose AutoThinkSQL, a framework that integrates an auto-thinking mechanism into both Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) on Text-to-SQL. Our approach enables the model to dynamically bypass reasoning for simple queries while invoking deep CoT for complex queries. On Qwen3-Coder-30B-A3B, our method achieves consistent gains compared to the best counterpart baseline on both Spider and BIRD benchmarks while simultaneously reducing average output tokens by 24.6% and 18.3%, and average latency by 17.1% and 11.5% compared to CoT-only generation. Further analysis indicates that the model learns to align its reasoning decisions with query difficulty.

Figures

Figures reproduced from arXiv: 2607.22622 by Jiheum Yeom, Kiwook Bae, Nohil Park, Sang Hun Kim, Soohyuk Jang, Sungroh Yoon, Yoonyoung Choi.

Figure 1
Figure 1. Figure 1: Comparison of Chain-of-Thought (CoT) and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: No-CoT-only format instruction. Format Instruction (AutoThinkSQL) Format Instructions: Analyze the question, database schema, and external knowledge to determine how to write the SQL query. First, evaluate the complexity of the problem. If it is an easy problem: Do not provide any reasoning or explanation. Directly output the final SQL query enclosed within [SQL] and [/SQL] tags. If it is a difficult probl… view at source ↗
Figure 5
Figure 5. Figure 5: Auto-thinking format instruction. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

35 extracted references · 1 canonical work pages

  1. [1]

    2023 , url=

    Mohammadreza Pourreza and Davood Rafiei , booktitle=. 2023 , url=

    2023

  2. [2]

    Gao, Dawei and Wang, Haibin and Li, Yaliang and Sun, Xiuyu and Qian, Yichen and Ding, Bolin and Zhou, Jingren , title =. Proc. VLDB Endow. , month = jan, pages =. 2024 , issue_date =. doi:10.14778/3641204.3641221 , abstract =

    arXiv 2024

  3. [3]

    arXiv preprint arXiv:2307.07306 , year=

    C3: Zero-shot text-to-sql with chatgpt , author=. arXiv preprint arXiv:2307.07306 , year=

    Pith/arXiv arXiv

  4. [4]

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

    Wang, Bing and Ren, Changyu and Yang, Jian and Liang, Xinnian and Bai, Jiaqi and Chai, LinZheng and Yan, Zhao and Zhang, Qian-Wen and Yin, Di and Sun, Xing and Li, Zhoujun. MAC - SQL : A Multi-Agent Collaborative Framework for Text-to- SQL. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  5. [5]

    arXiv preprint arXiv:2405.16755 , year=

    Chess: Contextual harnessing for efficient sql synthesis , author=. arXiv preprint arXiv:2405.16755 , year=

    Pith/arXiv arXiv

  6. [6]

    2025 , url=

    Mohammadreza Pourreza and Hailong Li and Ruoxi Sun and Yeounoh Chung and Shayan Talaei and Gaurav Tarlok Kakkar and Yu Gan and Amin Saberi and Fatma Ozcan and Sercan O Arik , booktitle=. 2025 , url=

    2025

  7. [7]

    Synthesizing Text-to- SQL Data from Weak and Strong LLM s

    Yang, Jiaxi and Hui, Binyuan and Yang, Min and Yang, Jian and Lin, Junyang and Zhou, Chang. Synthesizing Text-to- SQL Data from Weak and Strong LLM s. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.425

    doi:10.18653/v1/2024.acl-long.425 2024

  8. [8]

    arXiv preprint arXiv:2502.11741 , year=

    Sql-o1: A self-reward heuristic dynamic search method for text-to-sql , author=. arXiv preprint arXiv:2502.11741 , year=

    Pith/arXiv arXiv

  9. [9]

    and Le, Quoc V

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed H. and Le, Quoc V. and Zhou, Denny , title =. Proceedings of the 36th International Conference on Neural Information Processing Systems , articleno =. 2022 , isbn =

    2022

  10. [10]

    Xingyu Chen and Jiahao Xu and Tian Liang and Zhiwei He and Jianhui Pang and Dian Yu and Linfeng Song and Qiuzhi Liu and Mengfei Zhou and Zhuosheng Zhang and Rui Wang and Zhaopeng Tu and Haitao Mi and Dong Yu , booktitle=. Do. 2025 , url=

    2025

  11. [11]

    arXiv preprint arXiv:2505.11896 , year=

    Adacot: Pareto-optimal adaptive chain-of-thought triggering via reinforcement learning , author=. arXiv preprint arXiv:2505.11896 , year=

    Pith/arXiv arXiv

  12. [12]

    A dapt T hink: Reasoning Models Can Learn When to Think

    Zhang, Jiajie and Lin, Nianyi and Hou, Lei and Feng, Ling and Li, Juanzi. A dapt T hink: Reasoning Models Can Learn When to Think. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.184

    doi:10.18653/v1/2025.emnlp-main.184 2025

  13. [13]

    Thinkless:

    Gongfan Fang and Xinyin Ma and Xinchao Wang , booktitle=. Thinkless:. 2026 , url=

    2026

  14. [14]

    Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage

    Songjun Tu and Jiahao Lin and Qichao Zhang and Xiangyu Tian and Linjing Li and Xiangyuan Lan and Dongbin Zhao , booktitle=. Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage. 2026 , url=

    2026

  15. [15]

    Exploring Chain of Thought Style Prompting for Text-to- SQL

    Tai, Chang-Yu and Chen, Ziru and Zhang, Tianshu and Deng, Xiang and Sun, Huan. Exploring Chain of Thought Style Prompting for Text-to- SQL. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.327

    work page doi:10.18653/v1/2023.emnlp-main.327 2023

  16. [16]

    Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-

    Fangyu Lei and Jixuan Chen and Yuxiao Ye and Ruisheng Cao and Dongchan Shin and Hongjin SU and ZHAOQING SUO and Hongcheng Gao and Wenjing Hu and Pengcheng Yin and Victor Zhong and Caiming Xiong and Ruoxi Sun and Qian Liu and Sida Wang and Tao Yu , booktitle=. Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-. 2025 , url=

    2025

  17. [17]

    2026 , url=

    Jinyang Li and Xiaolong Li and Ge Qu and Per Jacobsson and Bowen Qin and Binyuan Hui and Shuzheng Si and Nan Huo and Xiaohan Xu and Yue Zhang and Ziwei Tang and Yuanshuai Li and Florensia Widjaja and Xintong Zhu and Feige Zhou and Yongfeng Huang and Yannis Papakonstantinou and Fatma Ozcan and Chenhao Ma and Reynold Cheng , booktitle=. 2026 , url=

    2026

  18. [18]

    arXiv preprint arXiv:2401.14196 , year=

    DeepSeek-Coder: when the large language model meets programming--the rise of code intelligence , author=. arXiv preprint arXiv:2401.14196 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2409.12186 , year=

    Qwen2.5-coder technical report , author=. arXiv preprint arXiv:2409.12186 , year=

    Pith/arXiv arXiv

  20. [20]

    Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    2018

  21. [21]

    Advances in Neural Information Processing Systems , volume=

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

  22. [22]

    Qwen3-Coder: Agentic Coding in the World , year =

  23. [23]

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

    Uncovering the impact of chain-of-thought reasoning for direct preference optimization: Lessons from text-to-sql , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  24. [24]

    arXiv preprint arXiv:2505.20315 , year=

    Arctic-text2sql-r1: Simple rewards, strong reasoning in text-to-sql , author=. arXiv preprint arXiv:2505.20315 , year=

    arXiv

  25. [25]

    Advances in Neural Information Processing Systems , volume=

    Sql-r1: Training natural language to sql reasoning model by reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  26. [26]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  27. [27]

    International Conference on Learning Representations , volume=

    To cot or not to cot? chain-of-thought helps mainly on math and symbolic reasoning , author=. International Conference on Learning Representations , volume=

  28. [28]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    Pith/arXiv arXiv

  29. [29]

    2025 , issue_date =

    Li, Haoyang and Wu, Shang and Zhang, Xiaokang and Huang, Xinmei and Zhang, Jing and Jiang, Fuxin and Wang, Shuai and Zhang, Tieying and Chen, Jianjun and Shi, Rui and Chen, Hong and Li, Cuiping , title =. 2025 , issue_date =. doi:10.14778/3749646.3749723 , journal =

    arXiv 2025

  30. [30]

    2024 , issue_date =

    Li, Haoyang and Zhang, Jing and Liu, Hanbing and Fan, Ju and Zhang, Xiaokang and Zhu, Jun and Wei, Renjie and Pan, Hongyan and Li, Cuiping and Chen, Hong , title =. 2024 , issue_date =. doi:10.1145/3654930 , journal =

    doi:10.1145/3654930 2024

  31. [31]

    Reasoning-

    Mohammadreza Pourreza and Shayan Talaei and Ruoxi Sun and Xingchen Wan and Hailong Li and Azalia Mirhoseini and Amin Saberi and Sercan O Arik , booktitle=. Reasoning-. 2025 , url=

    2025

  32. [32]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  33. [33]

    The Eleventh International Conference on Learning Representations , year=

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=. The Eleventh International Conference on Learning Representations , year=

  34. [34]

    Edward J Hu and yelong shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. Lo. 2022 , url=

    2022

  35. [35]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , address=

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , address=. 2024 , url=

    2024