Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A generate-and-judge SQL loop plus column-level partial-match rewards lifts execution accuracy to state-of-the-art on Spider, Spider 2.0, and BIRD.

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

2026-08-04 22:44 UTC pith:RKSZP246

load-bearing objection Useful RL-for-text-to-SQL engineering, but the SOTA headline does not survive its own tables. the 4 major comments →

arxiv 2509.07159 v1 pith:RKSZP246 submitted 2025-09-08 cs.AI

PaVeRL-SQL: Text-to-SQL via Partial-Match Rewards and Verbal Reinforcement Learning

Heng Hao , Wenjun Hu , Oxana Verkholyak , Davoud Ataee Tarzanagh , Baruch Gutow , Sima Didari , Masoud Faraki , Hankyu Moon
show 1 more author
Seungjai Min
This is my paper
classification cs.AI
keywords Text-to-SQLreinforcement learningGRPOpartial-match rewardsverbal reinforcement learningexecution accuracySpider 2.0cross-dialect transfer
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 aims to show that Text-to-SQL systems can be improved by replacing sparse 0/1 execution rewards with a column-level fractional match signal and by using a generate-and-judge loop in which the same LLM that writes candidate queries also scores them. It presents two pipelines: a verbal self-evaluation loop that samples ten executable SQL candidates per question and picks the highest-scoring one, and a chain-of-thought reinforcement-learning pipeline that trains a small 7B model with GRPO using the fractional match reward. On the industrial Spider 2.0-SQLite benchmark, the verbal pipeline reports 37% execution accuracy, 7.4 points above the reproduced CHESS baseline; the CoT pipeline also improves over prior RL systems on Spider, BIRD, and Spider 2.0. The paper further reports that mixing SQL dialects during RL training roughly triples execution accuracy on a low-resource MySQL test set without hurting the rich SQLite dialect. If these results hold, they suggest that denser execution feedback and verbal self-scoring are practical levers for making Text-to-SQL work on real enterprise databases.

Core claim

The central claim is that treating SQL correctness as a graded, column-level property rather than a binary pass/fail makes reinforcement learning for Text-to-SQL both more sample-efficient and more accurate. The paper defines EX_f, the proportion of golden result-table columns matched by the generated query's result, and uses it as the reward for GRPO training, alongside EX_b, a lenient binary match. It also shows that a training-free verbal approximation—sampling K=10 executable candidates and asking the same LLM to score them—can outperform established pipelines on the hardest public industrial benchmark. The reported numbers on Spider2.0-SQLite are 37.0% EX for the verbal pipeline versus

What carries the argument

The load-bearing mechanism is the partial-match execution reward EX_f: each generated SQL is executed, and its result table is compared column-by-column to the golden result table, giving a fractional score in [0,1]. This converts a sparse 0/1 signal into a dense reward that GRPO can optimize. The second mechanism is the verbal generate-and-judge loop: sample executable candidates, have the same LLM score each candidate 20 times, take the mean, and return the top-scoring SQL. This is a training-free approximation of the group-relative preference signal in GRPO. The two-stage GRPO schedule—restart from the best stage-one checkpoint, then cosine decay—is what keeps training stable within about

Load-bearing premise

The verbal pipeline assumes the LLM's self-scores on candidate SQL queries track true execution correctness; if the judge has a systematic bias (say, preferring verbose SQL), the reported 37% accuracy on Spider2.0-SQLite could be driven by the candidate sampling distribution rather than by valid selection.

What would settle it

Compare the LLM judge's mean score per candidate against actual column-level execution match on a held-out set of Spider2.0-SQLite candidates. If the correlation is near zero or negative, the selection step adds no signal; equally, if random selection among the ten executable candidates yields the same 37% EX, the judge is not the cause of the gain.

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 gains hold, a training-free verbal loop could become a standard inference-time recipe: sample until K executable candidates, self-score, select, with no gradient updates.
  • The 37% execution accuracy on Spider2.0-SQLite would raise the practical ceiling for enterprise Text-to-SQL, where prior methods stalled below 30% with the same backbone.
  • Column-level fractional rewards should allow small on-prem models (7B) to reach competitive accuracy with far fewer training epochs, lowering the cost of deployment.
  • Mixed-dialect training would let organizations with scarce dialect-specific data borrow supervision from a data-rich dialect (SQLite) instead of collecting new annotations.

Where Pith is reading between the lines

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

  • An immediate test: replace the LLM judge with random selection among the K executable candidates; if accuracy stays near 37%, the reported gain is attributable to sampling breadth, not judge quality. The paper does not report this ablation.
  • The column-level EX_f reward could generalize to other structured-generation tasks (e.g., data wrangling, code transformation, API generation) where exact-match rewards are sparse and execution feedback is available.
  • The verbal loop's 20-score mean per candidate is essentially a calibration-free ensemble judge; a calibrated judge or execution-based reranker might push accuracy further, but the paper does not explore this.
  • The two-stage schedule's rule 'if the curve plateaus raise LR, if it fluctuates halve it' is heuristic; a systematic study of when to restart could turn it into a general recipe for GRPO training.

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 / 5 minor

Summary. The paper proposes PaVeRL-SQL, a Text-to-SQL framework with two pipelines: (i) a training-free 'verbal RL' pipeline that samples K executable SQL candidates and selects via LLM self-scoring, and (ii) a CoT RL pipeline using GRPO with a new partial-match reward (EX_b/EX_f), a two-stage learning-rate schedule, and mixed-dialect training. The authors evaluate on Spider, BIRD, and Spider 2.0-SQLite, and claim state-of-the-art results on all three benchmarks, plus a threefold improvement on a low-resource in-house MySQL dialect.

Significance. If fully supported, the paper would make useful contributions: denser column-level execution rewards, a practical generate-and-judge verbal RL workflow, a two-stage GRPO schedule, and evidence for cross-dialect transfer. The authors provide concrete prompts, dataset splits, and training hyperparameters, and the code is made available, which aids reproducibility. However, the central SOTA claim is contradicted by the paper's own tables, and several load-bearing assertions about the evaluation and mixed-dialect results are internally inconsistent or lack supporting calibration evidence.

major comments (4)
  1. [Abstract; §IV-B; §IV-C (Tables II and III)] The abstract and introduction state that the pipelines 'achieve state-of-the-art (SOTA) results on popular Text-to-SQL benchmarks -- Spider, Spider 2.0, and BIRD.' This is contradicted by the paper's own results. Table III reports PaVeRL-BIRD majority EX 69.3, lower than Arctic-Text2SQL-R1's 69.4, and PaVeRL-Spider majority EX 86.6, lower than SQL-R1-7B's 87.6. Table II reports PaVeRL GPT-5-mini BIRD EX 55.8 on the dev subsample, while the reproduced CHESS baseline scores 61.2. The only benchmark where the reported number exceeds cited systems is Spider2.0-SQLite (37.0 vs. CHESS 29.6). The unqualified SOTA claim is therefore not supported by the provided evidence and must be substantially revised or restricted.
  2. [§IV-D, Table IV] The text states that mixed-dialect training improves the low-resource MySQL set 'without degrading' performance on the rich SQLite test set, but Table IV shows the opposite on greedy decoding: OmniSQL-7B achieves 15.6 EX on Spider2.0-SQLite, while PaVeRL-Mixed Gre achieves 13.3, a 2.3-point drop. The majority-vote result (17.8) improves, but the unconditional 'does not degrade' claim is false as stated. This needs to be corrected and qualified to the specific decoding/aggregation setting.
  3. [§III-B, Appendix A] The verbal RL pipeline's final selection relies on the backbone LLM's mean of 20 self-scores after merely filtering for executability. No calibration of these scores against true execution correctness is reported, and no comparison with alternative selection rules (e.g., random selection among executable candidates, or oracle selection) is given. Since the reported Spider2.0-SQLite gain of 37.0% over the zero-shot baseline (33.1%) is small in absolute terms (about five questions) and the selection mechanism is central to the verbal-RL contribution, the absence of any judge-accuracy correlation or ablation leaves the mechanism unvalidated.
  4. [§III-A] The formal definition of EX_b is inconsistent with the preceding prose. The prose says a result is correct if it 'contains all the information from the golden table' and has fewer than τ extra columns, but the formal condition given is only |Cols(Ť) \ Cols(T)| < τ. As written, a generated table that is missing many golden columns but has no extra columns would satisfy the formal condition and be counted correct. This affects the validity of EX_b as reported in Tables II and IV, where EX_b is used to support central claims. The formal definition needs to include the subset/containment condition and a precise statement of row/value matching.
minor comments (5)
  1. [§IV-A, Table II] BIRD results are reported on a 147-question subsample, but the abstract and conclusion refer to 'BIRD' without this qualification. Please state the evaluation subset explicitly wherever the benchmark is named.
  2. [Table III] The table reports majority-vote results for several baselines but does not clearly indicate whether the same decoding/temperature settings were used. Please add protocol details for each compared system.
  3. [Figure 3] The 'optimal' majority-vote size of about 32 is read from a figure with only a few points and no error bars. Given the small Spider2.0-SQLite test set (135 questions), the differences among group sizes may be within sampling noise; a brief statistical qualification would be helpful.
  4. [§IV-C] The learning-rate values are given as ranges ('1e-7 or 5e-7', '1e-5 or 5e-5'), but the final results are not tied to specific runs. Please report the exact schedule and hyperparameters used for each reported checkpoint.
  5. [III-A] The threshold τ is described as 'tunable per application,' but no sensitivity analysis is provided. Since EX_b is a proposed metric and is used in the mixed-dialect evaluation, reporting how results vary with τ (e.g., τ=0, 1, 5) would strengthen the presentation.

Circularity Check

0 steps flagged

No significant circularity: rewards and evaluation are external execution comparisons; the SOTA-claim inconsistency is a factual/consistency issue, not circular derivation.

full rationale

The paper's central derivations are empirical pipelines, not analytical reductions. The CoT RL reward is defined from EXf, which compares rollout execution results against golden SQL execution tables (Section III-C2): 'we execute each golden SQL to get the golden result table before the GRPO training starts. Then during the training we only compare the rollout SQL execution result table to the golden result table.' This is external, verifiable supervision, not a fitted input called prediction — the golden tables are fixed and independent of the policy. The verbal RL pipeline samples executable candidates and selects the highest LLM-judge score; this is a selection heuristic, and the paper itself acknowledges in the Limitations 'residual sensitivity of judge-style scoring to backbone biases.' No equation defines a predicted quantity in terms of a fitted parameter, and no load-bearing result is justified solely by a self-citation: the base model OmniSQL-7B, GRPO, the two-stage schedule, and the CHESS subsample are all external resources with their own stated constructions. The abstract's unqualified SOTA claim is contradicted by the paper's own Tables II and III on BIRD and Spider majority EX (e.g., PaVeRL BIRD EX 55.8 vs CHESS 61.2; PaVeRL-BIRD Maj 69.3 vs Arctic-Text2SQL-R1 69.4), but that is an internal consistency/accuracy problem, not circular reasoning under the enumerated patterns. The self-defined EXb/EXf metrics are metric choices; using the same EXf as reward and evaluation is standard objective alignment, not a derivation that reduces to its inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim relies on several hand-tuned free parameters (tau, reward weights, K, number of judge scores, LR schedule), and on domain assumptions about the validity of execution-based rewards, LLM self-judging, and the meaningfulness of column-level partial match. No new physical or conceptual entities are introduced.

free parameters (6)
  • Threshold tau in EXb = 5
    In Section III-A, EXb counts a result as correct if the count of extra columns is less than tau; tau=5 is chosen as default, tunable per application.
  • Reward weights for CoT RL = 10*EX_f, 0.5, 0
    Section III-C2 defines R=10*EX_f if SQL executes and returns results; 0.5 if executes but incorrect; 0 otherwise. These values are hand-tuned and no ablation is shown.
  • K executable candidates and sampling cap = 10, 200
    Section III-B: sample until K=10 executable SQLs or 200 attempts. These numbers are chosen without sensitivity analysis.
  • Number of judge scores per candidate = 20
    Section III-B: each candidate is scored 20 times and the mean is used as the final score. No rationale is given for 20.
  • Learning rate schedule hyperparameters = warm-up 1e-7 or 5e-7; max 1e-5 or 5e-5
    Section IV-C reports values that vary per run; the stage-two restart rule depends on observing the accuracy curve (plateau vs. fluctuate), a hand-tuned heuristic.
  • Majority voting group size = ~32
    Section IV-C and Figure 3 test 8, 16, 32, 64, 128 and find 32 'typically optimal.' This is a post-hoc selection from evaluation data.
axioms (5)
  • domain assumption Execution accuracy against golden SQL result is a valid reward signal for text-to-SQL.
    Used throughout for reward computation and evaluation; standard for the task.
  • domain assumption The same backbone LLM's self-scores correlate with execution correctness in the verbal-RL pipeline.
    Section III-B selects the final SQL by the LLM judge's score; no calibration against execution accuracy is reported.
  • domain assumption Column-wise comparison (EX_f) captures partial correctness meaningfully.
    Section III-A defines EX_f and uses it as the training reward; it ignores row ordering and tuple-set equivalence.
  • ad hoc to paper Non-negative rewards stabilize GRPO for text-to-SQL.
    Section III-C2 states empirically that mixing positive and negative rewards causes confusion, but no supporting ablation is shown.
  • domain assumption Filtering training data to remove missing-column or incorrect-SQL samples does not bias the result.
    Section IV-A1 filters Spider from 8659 to 8648 pairs and BIRD from 9428 to 9029 pairs, without analyzing what was removed.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of PaVeRL-SQL: Text-to-SQL via Partial-Match Rewards and Verbal Reinforcement Learning." pith.science (2026). https://pith.science/paper/RKSZP246

@misc{pith2026250907159,
  author       = {Pith},
  title        = {Pith review of: PaVeRL-SQL: Text-to-SQL via Partial-Match Rewards and Verbal Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RKSZP246}},
  note         = {Machine review of arXiv:2509.07159}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text-to-SQL models allow users to interact with a database more easily by generating executable SQL statements from natural-language questions. Despite recent successes on simpler databases and questions, current Text-to-SQL methods still suffer from low execution accuracy on industry-scale databases and complex questions involving domain-specific business logic. We present \emph{PaVeRL-SQL}, a framework that combines \emph{Partial-Match Rewards} and \emph{Verbal Reinforcement Learning} to drive self-improvement in reasoning language models (RLMs) for Text-to-SQL. To handle practical use cases, we adopt two pipelines: (1) a newly designed in-context learning framework with group self-evaluation (verbal-RL), using capable open- and closed-source large language models (LLMs) as backbones; and (2) a chain-of-thought (CoT) RL pipeline with a small backbone model (OmniSQL-7B) trained with a specially designed reward function and two-stage RL. These pipelines achieve state-of-the-art (SOTA) results on popular Text-to-SQL benchmarks -- Spider, Spider 2.0, and BIRD. For the industrial-level Spider2.0-SQLite benchmark, the verbal-RL pipeline achieves an execution accuracy 7.4\% higher than SOTA, and the CoT pipeline is 1.4\% higher. RL training with mixed SQL dialects yields strong, threefold gains, particularly for dialects with limited training data. Overall, \emph{PaVeRL-SQL} delivers reliable, SOTA Text-to-SQL under realistic industrial constraints. The code is available at https://github.com/PaVeRL-SQL/PaVeRL-SQL.

Figures

Figures reproduced from arXiv: 2509.07159 by Baruch Gutow, Davoud Ataee Tarzanagh, Hankyu Moon, Heng Hao, Masoud Faraki, Oxana Verkholyak, Seungjai Min, Sima Didari, Wenjun Hu.

Figure 1
Figure 1. Figure 1: Verbal RL Pipeline: A generate-and-judge workflow that approxi￾mates GRPO without gradient updates. For each natural language question, the pipeline (1) samples SQL candidates from the backbone LLM using the generation prompt until collecting K = 10 executable SQLs or reaching 200 attempts, (2) executes each candidate to verify validity, (3) uses the same LLM to score all executable candidates via the scor… view at source ↗
Figure 2
Figure 2. Figure 2: CoT RL training dynamics across two-stage GRPO optimization. Left: Greedy decoding execution accuracy (EX) progression showing model [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Impact of majority voting ensemble size on execution accuracy for CoT RL models. Results show official execution accuracy (EX, blue) and binary [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example schema string for the Spider school_budget database used in both training and inference. The string includes primary/foreign keys and lightweight profiles (MIN/MAX for numeric columns and top-3 modes for text) to guide joins and filters without leaking answers [PITH_FULL_IMAGE:figures/full_fig_p010_4.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 1 Pith paper

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

  1. Progress-SQL: Improving Reinforcement Learning for Text-to-SQL via Progressive Rewards

    cs.CL 2026-06 unverdicted novelty 5.0

    Progress-SQL introduces a multi-turn RL framework with ODT-based structural alignment and progressive rewards that measure improvement across refinement turns, yielding gains on BIRD, Spider, and robustness benchmarks.

Reference graph

Works this paper leans on

48 extracted references · 12 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Rat-sql: Relation-aware schema encoding and linking for text-to-sql parsers,

    B. Wang, R. Shin, X. Liu, O. Polozov, and M. Richardson, “Rat-sql: Relation-aware schema encoding and linking for text-to-sql parsers,” arXiv preprint arXiv:1911.04942, 2019

    Pith/arXiv arXiv 1911

  2. [2]

    Picard: Parsing incremen- tally for constrained auto-regressive decoding from language models,

    T. Scholak, N. Schucher, and D. Bahdanau, “Picard: Parsing incremen- tally for constrained auto-regressive decoding from language models,” arXiv preprint arXiv:2109.05093, 2021

    Pith/arXiv arXiv 2021

  3. [3]

    Din-sql: Decomposed in-context learning of text-to-sql with self-correction,

    M. Pourreza and D. Rafiei, “Din-sql: Decomposed in-context learning of text-to-sql with self-correction,”Advances in Neural Information Processing Systems, vol. 36, 2024

    work page 2024

  4. [4]

    CHESS: Contextual harnessing for efficient SQL synthesis,

    S. Talaei, M. Pourreza, Y .-C. Chang, A. Mirhoseini, and A. Saberi, “CHESS: Contextual harnessing for efficient SQL synthesis,”arXiv preprint arXiv:2405.16755, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task,

    T. Yu et al., “Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task,” 2018. [Online]. Available: https://arxiv.org/pdf/1809.08887

    Pith/arXiv arXiv 2018

  6. [6]

    Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls,

    J. Li, B. Hui, G. Qu, J. Yang, B. Li, B. Li, B. Wang, B. Qin, R. Geng, N. Huoet al., “Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls,”Advances in Neural Information Processing Systems, vol. 36, 2024

    work page 2024

  7. [8]

    CRUSH4SQL: Collective retrieval using schema hallucination for Text2SQL,

    M. Kothyari, D. Dhingra, S. Sarawagi, and S. Chakrabarti, “CRUSH4SQL: Collective retrieval using schema hallucination for Text2SQL,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, H. Bouamor, J. Pino, Fig. 3. Impact of majority voting ensemble size on execution accuracy for CoT RL models. Results show official ex...

    work page 2023

  8. [9]

    Text-to-sql empowered by large language models: A benchmark evaluation,

    D. Gao, H. Wang, Y . Li, X. Sun, Y . Qian, B. Ding, and J. Zhou, “Text-to-sql empowered by large language models: A benchmark evaluation,” 2023. [Online]. Available: https://arxiv.org/abs/2308.15363

    Pith/arXiv arXiv 2023

  9. [10]

    Fundamental challenges in evaluating text2sql solutions and detecting their limitations,

    C. Renggli, I. F. Ilyas, and T. Rekatsinas, “Fundamental challenges in evaluating text2sql solutions and detecting their limitations,” 2025. [Online]. Available: https://arxiv.org/abs/2501.18197

    Pith/arXiv arXiv 2025

  10. [11]

    Openai o1 system card,

    OpenAI, “Openai o1 system card,” 2024. [Online]. Available: https://arxiv.org/abs/2412.16720

    Pith/arXiv arXiv 2024

  11. [12]

    DeepSeek-R1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo et al., “DeepSeek-R1: Incentivizing reasoning capability in llms via reinforcement learning,” DeepSeek-AI, Tech. Rep., 2025. [Online]. Available: https://arxiv.org/abs/2501.12948

    Pith/arXiv arXiv 2025

  12. [13]

    Challenging big-bench tasks and whether chain-of-thought can solve them,

    M. Suzgun et al., “Challenging big-bench tasks and whether chain-of-thought can solve them,” 2022. [Online]. Available: https: //arxiv.org/abs/2210.09261

    Pith/arXiv arXiv 2022

  13. [14]

    SFT memorizes, RL generalizes: A comparative study of foundation model post-training,

    T. Chu et al., “SFT memorizes, RL generalizes: A comparative study of foundation model post-training,” 2025. [Online]. Available: https://arxiv.org/pdf/2501.17161

    Pith/arXiv arXiv 2025

  14. [15]

    Seq2sql: Generating structured queries from natural language using reinforcement learning,

    V . Zhong, C. Xiong, and R. Socher, “Seq2sql: Generating structured queries from natural language using reinforcement learning,”arXiv preprint arXiv:1709.00103, 2017

    Pith/arXiv arXiv 2017

  15. [16]

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

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.03300

    Pith/arXiv arXiv 2024

  16. [17]

    Resdsql: Decoupling schema linking and skeleton parsing for text-to-sql,

    H. Li, J. Zhang, C. Li, and H. Chen, “Resdsql: Decoupling schema linking and skeleton parsing for text-to-sql,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 2023, pp. 13 067–13 075

    work page 2023

  17. [18]

    Codes: Towards building open-source language models for text-to-sql,

    H. Li, J. Zhang, H. Liu, J. Fan, X. Zhang, J. Zhu, R. Wei, H. Pan, C. Li, and H. Chen, “Codes: Towards building open-source language models for text-to-sql,”Proc. ACM Manag. Data, vol. 2, no. 3, May

    work page

  18. [19]

    PSM-SQL: Progressive Schema Learning with Multi-granularity Semantics for Text-to-SQL

    Z. Yang, Y . Xie, R. Zhong, Y . Tan, E. Liu, Z. Yang, M. Gao, B. Hu, and Z. Li, “Psm-sql: Progressive schema learning with multi-granularity semantics for text-to-sql,”arXiv preprint arXiv:2502.05237, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [20]

    Act-sql: In-context learning for text-to-sql with automatically-generated chain-of-thought,

    H. Zhang, R. Cao, L. Chen, H. Xu, and K. Yu, “Act-sql: In-context learning for text-to-sql with automatically-generated chain-of-thought,” arXiv preprint arXiv:2310.17342, 2023

    Pith/arXiv arXiv 2023

  20. [21]

    CHASE-SQL: Multi-path reasoning and preference optimized candidate selection in text-to-sql,

    M. Pourreza et al., “CHASE-SQL: Multi-path reasoning and preference optimized candidate selection in text-to-sql,” 2024. [Online]. Available: https://arxiv.org/pdf/2410.01943

    Pith/arXiv arXiv 2024

  21. [22]

    Re- flexion: Language agents with verbal reinforcement learning,

    N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao, “Re- flexion: Language agents with verbal reinforcement learning,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  22. [23]

    Text-to-sql calibration: No need to ask – just rescale model probabilities,

    A. Ramachandran and S. Sarawagi, “Text-to-sql calibration: No need to ask – just rescale model probabilities,” 2024. [Online]. Available: https://arxiv.org/abs/2411.16742

    Pith/arXiv arXiv 2024

  23. [24]

    Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback,

    K. Tian, E. Mitchell, A. Zhou, A. Sharma, R. Rafailov, H. Yao, C. Finn, and C. D. Manning, “Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback,” 2023. [Online]. Available: https: //arxiv.org/abs/2305.14975

    Pith/arXiv arXiv 2023

  24. [25]

    Acereason-nemotron: Advancing math and code reasoning through reinforcement learning,

    Y . Chen, Z. Yang, Z. Liu, C. Lee, P. Xu, M. Shoeybi, B. Catanzaro, and W. Ping, “Acereason-nemotron: Advancing math and code reasoning through reinforcement learning,” 2025. [Online]. Available: https://arxiv.org/abs/2505.16400

    Pith/arXiv arXiv 2025

  25. [26]

    Qwen3 technical report,

    A. Yang et al., “Qwen3 technical report,” Qwen Team Alibaba Group, Tech. Rep., 2025. [Online]. Available: https://arxiv.org/abs/2505.09388

    Pith/arXiv arXiv 2025

  26. [27]

    Phi-4-reasoning technical report,

    M. Abdin et al., “Phi-4-reasoning technical report,” Microsoft Research, Tech. Rep., 2025. [Online]. Available: https://arxiv.org/abs/2504.21318

    Pith/arXiv arXiv 2025

  27. [28]

    Deepseek llm: Scaling open-source language models with longtermism,

    DeepSeek-AI, “Deepseek llm: Scaling open-source language models with longtermism,” 2024. [Online]. Available: https://arxiv.org/abs/ 2401.02954

    Pith/arXiv arXiv 2024

  28. [29]

    In-context reinforcement learn- ing with retrieval-augmented generation for text-to-sql,

    R. Toteja, A. Sarkar, and P. M. Comar, “In-context reinforcement learn- ing with retrieval-augmented generation for text-to-sql,” inProceedings of the 31st International Conference on Computational Linguistics, 2025, pp. 10 390–10 397

    work page 2025

  29. [30]

    LLM-based SQL generation with reinforcement learning,

    Anonymous, “LLM-based SQL generation with reinforcement learning,” inSubmitted to The First Workshop on Neural Reasoning and Mathematical Discovery at AAAI’2025, 2024, under review. [Online]. Available: https://openreview.net/forum?id=84M0Jaiapl

    work page 2025

  30. [31]

    Star-sql: Self-taught reasoner for text-to-sql,

    M. He, Y . Shen, W. Zhang, Q. Peng, J. Wang, and W. Lu, “Star-sql: Self-taught reasoner for text-to-sql,”arXiv preprint arXiv:2502.13550, 2025

    Pith/arXiv arXiv 2025

  31. [32]

    Deepsql-r1: A quantized llm for high-performance and reinforcement driven nl2sql generation,

    S. Kumar, “Deepsql-r1: A quantized llm for high-performance and reinforcement driven nl2sql generation,”figshare. Journal contribution, 2025. [Online]. Available: https://doi.org/10.6084/m9. figshare.28330301.v1

    doi:10.6084/m9 2025

  32. [33]

    Reasoning-SQL: Reinforcement learning with SQL tailored partial rewards for reasoning-enhanced Text-to-SQL,

    M. Pourreza et al., “Reasoning-SQL: Reinforcement learning with SQL tailored partial rewards for reasoning-enhanced Text-to-SQL,” 2025. [Online]. Available: https://arxiv.org/pdf/2503.23157

    Pith/arXiv arXiv 2025

  33. [34]

    Sql-r1: Training natural language to sql reasoning model by reinforcement learning,

    P. Ma, X. Zhuang, C. Xu, X. Jiang, R. Chen, and J. Guo, “Sql-r1: Training natural language to sql reasoning model by reinforcement learning,” 2025. [Online]. Available: https://arxiv.org/abs/2504.08600

    arXiv 2025

  34. [35]

    Arctic-Text2SQL-R1: Simple rewards, strong reasoning in Text-to-SQL,

    Z. Yao et al., “Arctic-Text2SQL-R1: Simple rewards, strong reasoning in Text-to-SQL,” 2025. [Online]. Available: https://arxiv.org/pdf/2505. 20315

    work page 2025

  35. [36]

    ReAct: Synergizing reasoning and acting in language models,

    S. Yao et al., “ReAct: Synergizing reasoning and acting in language models,” inInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  36. [37]

    Teaching large language models to self-debug,

    X. Chen, M. Lin, N. Sch ¨arli, and D. Zhou, “Teaching large language models to self-debug,”arXiv preprint arXiv:2304.05128, 2023

    Pith/arXiv arXiv 2023

  37. [38]

    Intercode: Standard- izing and benchmarking interactive coding with execution feedback,

    J. Yang, A. Prabhakar, K. Narasimhan, and S. Yao, “Intercode: Standard- izing and benchmarking interactive coding with execution feedback,” Advances in Neural Information Processing Systems, vol. 36, 2024

    work page 2024

  38. [39]

    Coder: Issue resolving with multi-agent and task graphs,

    D. Chen, S. Lin, M. Zeng, D. Zan, J.-G. Wang, A. Cheshkov, J. Sun, H. Yu, G. Dong, A. Alievet al., “Coder: Issue resolving with multi-agent and task graphs,”arXiv preprint arXiv:2406.01304, 2024

    Pith/arXiv arXiv 2024

  39. [40]

    Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models,

    L. Wang, W. Xu, Y . Lan, Z. Hu, Y . Lan, R. K.-W. Lee, and E.-P. Lim, “Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models,”arXiv preprint arXiv:2305.04091, 2023

    Pith/arXiv arXiv 2023

  40. [41]

    Agentless: De- mystifying llm-based software engineering agents,

    C. S. Xia, Y . Deng, S. Dunn, and L. Zhang, “Agentless: De- mystifying llm-based software engineering agents,”arXiv preprint arXiv:2407.01489, 2024

    Pith/arXiv arXiv 2024

  41. [42]

    ReFoRCE: A Text-to-SQL Agent with Self- Refinement, Format Restriction, and Column Exploration,

    M. Deng, A. Ramachandran, C. Xu, L. Hu, Z. Yao, A. Datta, and H. Zhang, “ReFoRCE: A Text-to-SQL Agent with Self- Refinement, Format Restriction, and Column Exploration,”arXiv preprint arXiv:2502.00675, 2025

    Pith/arXiv arXiv 2025

  42. [43]

    Spider 2.0: Evaluating language models on real- world enterprise text-to-sql workflows,

    F. Lei et al., “Spider 2.0: Evaluating language models on real- world enterprise text-to-sql workflows,” 2024. [Online]. Available: https://arxiv.org/pdf/2411.07763

    Pith/arXiv arXiv 2024

  43. [44]

    OmniSQL: Synthesizing high-quality text-to-sql data at scale,

    H. Li et al., “OmniSQL: Synthesizing high-quality text-to-sql data at scale,” 2025. [Online]. Available: https://arxiv.org/pdf/2503.02240

    Pith/arXiv arXiv 2025

  44. [45]

    Qwen2.5-Coder technical report,

    B. Hui et al., “Qwen2.5-Coder technical report,” Qwen Team Alibaba Group, Tech. Rep., 2024. [Online]. Available: https://arxiv.org/abs/ 2409.12186

    Pith/arXiv arXiv 2024

  45. [46]

    Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls,

    J. Li et al., “Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls,”Advances in Neural Information Processing Systems, vol. 36, 2024

    work page 2024

  46. [47]

    gpt-oss-120b & gpt-oss-20b model card,

    OpenAI, “gpt-oss-120b & gpt-oss-20b model card,” 2025. [Online]. Available: https://arxiv.org/abs/2508.10925

    Pith/arXiv arXiv 2025

  47. [48]

    Hybridflow: A flexible and efficient rlhf framework,

    G. Sheng, C. Zhang, Z. Ye, X. Wu, W. Zhang, R. Zhang, Y . Peng, H. Lin, and C. Wu, “Hybridflow: A flexible and efficient rlhf framework,”arXiv preprint arXiv: 2409.19256, 2024. APPENDIXA VERBALRL PROMPTS The SQL generating prompt is the same as Prompt 1. We use the scoring prompt in a lightweight generate–judge loop that approximates GRPO without gradient...

    Pith/arXiv arXiv 2024

  48. [2024]

    Available: https://doi.org/10.1145/3654930

    [Online]. Available: https://doi.org/10.1145/3654930

    doi:10.1145/3654930

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.