SEFRQO: A Self-Evolving Fine-Tuned RAG-Based Query Optimizer

Hanwen Liu; Ibrahim Sabek; Qihan Zhang; Ryan Marcus

REVIEW 3 major objections 3 minor 65 references

SEFRQO: A Self-Evolving Fine-Tuned RAG-Based Query Optimizer

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SEFRQO claims a fine-tuned LLM with a retrieval-augmented, self-evolving prompt loop can beat learned query optimizers and PostgreSQL, cutting query latency by up to 65.05% on the CEB workload and 93.57% on Stack.

desk verdict The submission's body is an unrelated Boltzina paper, so there is no SEFRQO to review; the abstract's idea is plausible but unverifiable, and the self-evolution loop raises a real train/test-overlap concern. read the letter →

arxiv 2508.17556 v1 pith:UG45BKGA submitted 2025-08-24 cs.DB

Hanwen Liu , Qihan Zhang , Ryan Marcus , Ibrahim Sabek This is my paper

classification cs.DB

keywords learnedqueryoptimizerLLMretrieval-augmentedgenerationhintsself-evolvingexecutionfeedbackcold-startlatencyreduction

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

SEFRQO is a query optimizer built on a fine-tuned LLM that suggests query hints. The paper claims it overcomes two weaknesses of learned query optimizers—cold start and retraining on workload or schema drift—by continuously absorbing execution feedback through a retrieval-augmented generation (RAG) loop. Instead of retraining, it builds each prompt from similar past queries and the same query's historical execution record, then iteratively revises the prompt to lower actual query latency. On the CEB and Stack benchmarks, the authors report that SEFRQO beats existing learned optimizers and PostgreSQL, cutting query latency by up to 65.05% and 93.57%, respectively. The claim matters because if true, LLM-based query optimization can improve with real workload experience without expensive retraining.

What carries the argument

The central mechanism is a RAG-based prompt construction loop. A fine-tuned LLM, prepared with supervised and reinforcement fine-tuning, generates query hints; a retriever supplies the prompt with both similar past queries and the same query's own historical execution record. The self-evolution step iteratively updates the prompt using the recorded latency of each execution, letting the optimizer improve from feedback without modifying weights. This combination—in-context learning plus execution feedback—is what the paper argues removes cold-start and retraining costs while beating PostgreSQL and existing learned optimizers.

What would settle it

Run SEFRQO on a fresh workload of never-before-seen query templates twice: once with the execution-history store populated only by other queries, and once with the target query's own history included. If the latency drop nearly disappears when the target's own history is withheld, the self-evolution gain is case-specific memorization rather than transferable learning.

Watch

Extended reading notes

Core claim

The center of the paper is a self-evolving generation-retrieval loop rather than a new cost model. SEFRQO starts with a supervised fine-tuned LLM that produces syntactically correct query hints, and a reinforcement fine-tuning stage that pushes hints toward lower measured latency. At inference, a retriever pulls both similar queries from past executions and the execution record of the exact query being optimized, assembling them into the prompt. After each run, the recorded latency feeds back, and the prompt for that query is iteratively refined to minimize latency. The authors claim this loop removes the need for cold-start retraining and adapts to workload shifts, yielding up to a 93.57% l

Load-bearing premise

The reported speedups count only if the queries used for evaluation did not also supply the historical execution records that built their own prompts; otherwise the optimizer is being tested on memorized cases rather than generalizing.

Editorial extensions

If this is right

If the loop generalizes, a database can improve its optimizer hints over time by simply keeping execution logs, with no retraining pipeline needed.
Workload or schema shifts can be absorbed by the retrieval store instead of by weight updates, since the prompt is rebuilt from recent similar executions.
The reported gains on CEB (65.05%) and Stack (93.57%) versus PostgreSQL position SEFRQO as a new reference point for LLM-based query-optimizer research.
Combining supervised fine-tuning with reinforcement fine-tuning offers a practical recipe for making LLM output both syntactically valid and latency-aware.

Reading between the lines

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

The paper leaves implicit a critical separation: the execution records used to evolve prompts must be disjoint from the queries used for evaluation. If they overlap, the reported speedups are fitted rather than generalized; a held-out evaluation would settle this.
The same retrieve-execute-feedback pattern could extend to index selection, join-order enumeration, or cardinality estimation, where execution records are cheap to collect.
Wall-clock latency as the reward signal assumes run-to-run stability; on a busy or cloud database, noisy timings could mislead the evolution loop, so a denoised or stable cost signal would be a natural robustness enhancement.
As the execution-history store grows, prompt retrieval may need staleness-aware ranking so that outdated schema or statistics do not dominate recent experience.

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

Colleague, here's the short version: arXiv:2508.17556 is not a paper about SEFRQO. The metadata and abstract describe an LLM-based query optimizer, but the full text is the preprint of Boltzina, a docking-based virtual screening method with different authors and no shared content. So as submitted, there is no methodology, evaluation, or artifact to inspect for the actual claims.

That said, the abstract itself points at something new: combining supervised and reinforcement fine-tuning with RAG that retrieves similar queries plus the same query's execution history, and then letting the prompt evolve against latency feedback. That is a plausible extension of recent LLM-based optimizers, and the claimed numbers (65.05% and 93.57% better than PostgreSQL on CEB and Stack) would matter if they held up. I can't credit them beyond that, because there is nothing to check.

The main technical worry from the abstract is exactly the one the stress-test flags: if the queries used to evolve prompts are also in the evaluation set, the gains are fitted, not generalized. The abstract says prompts are built from 'the historical execution record of the same query' and the loop minimizes that query's latency, which makes the split between evolution pool and test set load-bearing. Without a dataset split or an ablation, the reported reductions might just be tuning to the benchmark. A secondary issue is that per-query latency is noisy; even with a clean split you'd want repeated runs and variance.

There's also the submission integrity issue. The manuscript text is a different paper. That's a desk-reject condition on its own, independent of the technical merits. I can't say whether SEFRQO's idea is sound; there's no body to evaluate. The Boltzina paper looks like a legitimate separate preprint, but it is not this paper.

Who should read this? Anyone tracking learned query optimizers might want to look at the abstract when a corrected version appears. As-is, this document gives no basis for a serious referee. My recommendation: desk reject this submission and tell the authors to resubmit the correct SEFRQO full text; if they do, then send it out with attention to the train/test split and the latency-variance question.

Referee Report

3 major / 3 minor

Summary. The manuscript as submitted contains an abstract for SEFRQO, a Self-Evolving Fine-Tuned RAG-based Query Optimizer. The abstract claims that SEFRQO mitigates the cold-start problem of learned query optimizers by continuously learning from execution feedback, constructing prompts from similar queries and from the historical execution record of the same query, and iteratively optimizing prompts to minimize query execution latency. It reports up to 65.05% and 93.57% reductions in query latency on the CEB and Stack workloads, respectively, compared to PostgreSQL. However, the supplied full text is an unrelated bioinformatics preprint titled 'Boltzina: Efficient and Accurate Virtual Screening via Docking-Guided Binding Prediction with Boltz-2'. It contains no description of SEFRQO's architecture, training procedure, retrieval mechanism, prompt construction, evaluation protocol, baselines, or experimental results. The central claims of the abstract are therefore not supported by any verifiable content in the submission.

Significance. If substantiated, the SEFRQO approach would be a timely contribution: using LLMs in a self-evolving retrieval-augmented loop with execution feedback could address cold-start and workload-shift limitations of learned query optimizers, and the abstract provides concrete, falsifiable performance targets. The formulation of the problem and the named workloads are the manuscript's main strengths. That said, the submission gives no methodology or evaluation to check. The reported quantitative gains cannot be assessed, and the absence of any SEFRQO content is a complete evidentiary gap.

major comments (3)

[Full text (all sections) vs Abstract] The submitted full text is not the SEFRQO paper; it is the Boltzina bioinformatics preprint about Boltz-2 and AutoDock Vina. There is no section describing SEFRQO's fine-tuning, RAG retrieval, prompt construction, self-evolution loop, or experiments. The central claim in the Abstract (up to 65.05%/93.57% latency reductions on CEB and Stack) is therefore entirely unsupported. This is a load-bearing omission that prevents any scientific evaluation of the paper's contribution.
[Abstract: 'historical execution record of the same query'] The abstract states that prompts are constructed from the historical execution record of the same query and that the self-evolving paradigm iteratively optimizes the prompt to minimize query execution latency. This creates a direct circularity risk: if the queries used to evolve prompts are also those whose latency is reported, the improvements are fitted rather than generalized. The manuscript specifies no train/evolution versus evaluation split, no temporal separation, and no exclusion of evaluation queries from the historical record. The claimed 65.05% and 93.57% reductions cannot be interpreted without this information.
[Evaluation/experimental protocol (absent)] No experimental protocol accompanies the reported numbers. The abstract names PostgreSQL and 'state-of-the-art LQOs' but gives no citation, no hardware description, no cache-state controls, no number of runs, no variance/confidence intervals, and no artifact or code link. Query latency is a noisy metric on real workloads; without repeated measurements and a precise evaluation methodology, the headline reductions are not credible evidence of improvement.

minor comments (3)

[Header/front matter] The full text is marked 'arXiv:2508.17555v1 [q-bio.BM]', which conflicts with the submitted identifier arXiv:2508.17556 (cs.DB). This appears to be a file/submission mismatch.
[References] The abstract refers to 'state-of-the-art LQOs', CEB, and Stack without citations or definitions. If the correct SEFRQO manuscript is supplied, it should identify these baselines and workloads precisely.
[Unrelated content] The Boltzina preprint is unrelated to SEFRQO. If this was a submission error, the correct manuscript should be uploaded; as presented, the document contains no SEFRQO content beyond the abstract.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity established: same-query historical context is a normal adaptive feature, and no evidence shows evaluation queries overlap with prompt-evolution data.

full rationale

The abstract describes a self-evolving RAG-based query optimizer that constructs prompts from similar queries and the same query's historical execution record, then iteratively optimizes the prompt to minimize latency. This is a feedback-driven adaptive system, not a derivation that reduces to its inputs. To establish circularity under the paper's own reasoning, one would need to show that the evaluation queries are the very queries whose historical records were used to evolve prompts (fitted_input_called_prediction) or that the prompt optimization target is identical to the reported metric by construction (self_definitional). The abstract provides no dataset split, no methodology, and no equations; it merely states that historical records of the same query are used. Absent any explicit statement that the test workloads are contained in the evolution pool, this remains a potential concern, not a demonstrable circularity. Moreover, the supplied full text is an unrelated Boltzina preprint, so the actual SEFRQO evaluation methodology is unavailable for inspection. Under the rule that circularity must be exhibited with a specific reduction and not speculated, no circular step can be identified. No self-citations or imported uniqueness theorems are present. Thus the evidence supports a score of 0 (no significant circularity).

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

The abstract names three mechanisms the result rests on: fine-tuning (SFT and RFT) producing valid and efficient hints, RAG retrieval of similar queries and the same query's execution history improving generations, and latency as the reward driving self-evolution. No free parameters can be enumerated because the body, which would list retrieval counts, evolution rounds, and reward weights, is absent (the full text is an unrelated paper). No invented physical entities apply. These axioms are domain assumptions asserted by the abstract, unverified in the available material.

assumptions (3)

domain assumption Supervised and reinforcement fine-tuning of the LLM yields syntactically correct and performance-efficient query hints at inference time.
This is the paper's stated preparation method (abstract); its success is assumed in the claimed results, and no fine-tuning data or evaluation is present in the available material.
domain assumption Retrieving similar queries and the same query's historical execution record improves in-context hint generation.
Central mechanism of the RAG framework; the abstract asserts it, with no ablations available to confirm.
domain assumption Execution latency is a stable, informative reward signal for prompt evolution.
The self-evolving loop minimizes latency; this assumes latency measurements are reproducible across runs and not dominated by system noise, which the abstract does not address.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SEFRQO: A Self-Evolving Fine-Tuned RAG-Based Query Optimizer." pith.science (2026). https://pith.science/paper/UG45BKGA

@misc{pith2026250817556,
  author       = {Pith},
  title        = {Pith review of: SEFRQO: A Self-Evolving Fine-Tuned RAG-Based Query Optimizer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UG45BKGA}},
  note         = {Machine review of arXiv:2508.17556}
}

read the original abstract

Query optimization is a crucial problem in database systems that has been studied for decades. Learned query optimizers (LQOs) can improve performance over time by incorporating feedback; however, they suffer from cold-start issues and often require retraining when workloads shift or schemas change. Recent LLM-based query optimizers leverage pre-trained and fine-tuned LLMs to mitigate these challenges. Nevertheless, they neglect LLMs' in-context learning and execution records as feedback for continuous evolution. In this paper, we present SEFRQO, a Self-Evolving Fine-tuned RAG-based Query Optimizer. SEFRQO mitigates the cold-start problem of LQOs by continuously learning from execution feedback via a Retrieval-Augmented Generation (RAG) framework. We employ both supervised fine-tuning and reinforcement fine-tuning to prepare the LLM to produce syntactically correct and performance-efficient query hints. Moreover, SEFRQO leverages the LLM's in-context learning capabilities by dynamically constructing prompts with references to similar queries and the historical execution record of the same query. This self-evolving paradigm iteratively optimizes the prompt to minimize query execution latency. Evaluations show that SEFRQO outperforms state-of-the-art LQOs, achieving up to 65.05% and 93.57% reductions in query latency on the CEB and Stack workloads, respectively, compared to PostgreSQL.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 43 canonical work pages

[1]

Many-shot in-context learning

Rishabh Agarwal, Avi Singh, Lei Zhang, Bernd Bohnet, Luis Rosias, Stephanie Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, et al. Many-shot in-context learning. Advances in Neural Information Processing Systems , 37:76930--76966, 2024

work page 2024
[2]

The unreasonable effectiveness of LLMs for query optimization

Peter Akioyamen, Zixuan Yi, and Ryan Marcus. The unreasonable effectiveness of LLMs for query optimization. In Machine Learning for Systems Workshop at the 38th Conference on Neural Information Processing Systems (NeurIPS 2024) , 2024

work page 2024
[3]

What learning algorithm is in-context learning? investigations with linear models

Ekin Aky \"u rek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learning algorithm is in-context learning? investigations with linear models. arXiv preprint arXiv:2211.15661 , 2022

arXiv 2022
[4]

https://www.tpc.org/tpcds/

TPC-DS Benchmark. https://www.tpc.org/tpcds/

work page
[5]

A Robust and Explainable Query Optimization Cost Model Based on Bidirectional Graph Neural Networks

Baoming Chang. A Robust and Explainable Query Optimization Cost Model Based on Bidirectional Graph Neural Networks . PhD thesis, University of Ottawa, 2024

work page 2024
[6]

Loger: A learned optimizer towards generating efficient and robust query execution plans

Tianyi Chen, Jun Gao, Hedui Chen, and Yaofeng Tu. Loger: A learned optimizer towards generating efficient and robust query execution plans. Proceedings of the VLDB Endowment , 16(7):1777--1789, 2023

work page 2023
[7]

Murag: Multimodal retrieval-augmented generator for open question answering over images and text

Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William W Cohen. Murag: Multimodal retrieval-augmented generator for open question answering over images and text. arXiv preprint arXiv:2210.02928 , 2022

arXiv 2022
[8]

Leon: A new framework for ml-aided query optimization

Xu Chen, Haitian Chen, Zibo Liang, Shuncheng Liu, Jinghong Wang, Kai Zeng, Han Su, and Kai Zheng. Leon: A new framework for ml-aided query optimization. Proceedings of the VLDB Endowment , 16(9):2261--2273, 2023

work page 2023

Show all 65 references

[9]

Cocomic: Code completion by jointly modeling in-file and cross-file context

Yangruibo Ding, Zijian Wang, Wasi Uddin Ahmad, Murali Krishna Ramanathan, Ramesh Nallapati, Parminder Bhatia, Dan Roth, and Bing Xiang. Cocomic: Code completion by jointly modeling in-file and cross-file context. arXiv preprint arXiv:2212.10007 , 2022

2022 arXiv
[10]

Kepler: robust learning for parametric query optimization

Lyric Doshi, Vincent Zhuang, Gaurav Jain, Ryan Marcus, Haoyu Huang, Deniz Altinb \"u ken, Eugene Brevdo, and Campbell Fraser. Kepler: robust learning for parametric query optimization. Proceedings of the ACM on Management of Data , 1(1):1--25, 2023

2023
[11]

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

Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. Text-to-sql empowered by large language models: A benchmark evaluation. Proc. VLDB Endow. , 17(5):1132–1145, January 2024

2024
[12]

Demonstrating -tune: Exploiting large language models for workload-adaptive database system tuning

Victor Giannakouris and Immanuel Trummer. Demonstrating -tune: Exploiting large language models for workload-adaptive database system tuning. In Companion of the 2024 International Conference on Management of Data , pages 508--511, 2024

2024
[13]

Supervised contrastive learning for pre-trained language model fine-tuning

Beliz Gunel, Jingfei Du, Alexis Conneau, and Ves Stoyanov. Supervised contrastive learning for pre-trained language model fine-tuning. arXiv preprint arXiv:2011.01403 , 2020

2011 arXiv
[14]

Retrieval augmented language model pre-training

Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pre-training. In International conference on machine learning , pages 3929--3938. PMLR, 2020

2020
[15]

Transformer in transformer

Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in neural information processing systems , 34:15908--15919, 2021

2021
[16]

Next-generation database interfaces: A survey of llm-based text-to-sql, 2025

Zijin Hong, Zheng Yuan, Qinggang Zhang, Hao Chen, Junnan Dong, Feiran Huang, and Xiao Huang. Next-generation database interfaces: A survey of llm-based text-to-sql, 2025

2025
[17]

Llmtune: Accelerate database knob tuning with large language models

Xinmei Huang, Haoyang Li, Jing Zhang, Xinxin Zhao, Zhiming Yao, Yiyan Li, Zhuohao Yu, Tieying Zhang, Hong Chen, and Cuiping Li. Llmtune: Accelerate database knob tuning with large language models. arXiv preprint arXiv:2404.11581 , 2024

2024 arXiv
[18]

A survey on large language models for code generation, 2024

Juyong Jiang, Fan Wang, Jiasi Shen, Sungju Kim, and Sunghun Kim. A survey on large language models for code generation, 2024

2024
[19]

Ehsan Kamalloo, Nouha Dziri, Charles L. A. Clarke, and Davood Rafiei. Evaluating open-domain question answering in the era of large language models, 2023

2023
[20]

On information and sufficiency

Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics , 22(1):79--86, 1951

1951
[21]

Gptuner: A manual-reading database tuning system via gpt-guided bayesian optimization

Jiale Lao, Yibo Wang, Yufei Li, Jianping Wang, Yunjia Zhang, Zhiyuan Cheng, Wanghu Chen, Mingjie Tang, and Jianguo Wang. Gptuner: A manual-reading database tuning system via gpt-guided bayesian optimization. arXiv preprint arXiv:2311.03157 , 2023

2023 arXiv
[22]

How good are query optimizers, really? Proceedings of the VLDB Endowment , 9(3):204--215, 2015

Viktor Leis, Andrey Gubichev, Atanas Mirchev, Peter Boncz, Alfons Kemper, and Thomas Neumann. How good are query optimizers, really? Proceedings of the VLDB Endowment , 9(3):204--215, 2015

2015
[23]

Fvqa 2.0: Introducing adversarial samples into fact-based visual question answering

Weizhe Lin, Zhilin Wang, and Bill Byrne. Fvqa 2.0: Introducing adversarial samples into fact-based visual question answering. arXiv preprint arXiv:2303.10699 , 2023

2023 arXiv
[24]

Serag: Self-evolving rag system for query optimization

Hanwen Liu, Qihan Zhang, Ryan Marcus, and Ibrahim Sabek. Serag: Self-evolving rag system for query optimization. 2025

2025
[25]

Reacc: A retrieval-augmented code completion framework

Shuai Lu, Nan Duan, Hojae Han, Daya Guo, Seung-won Hwang, and Alexey Svyatkovskiy. Reacc: A retrieval-augmented code completion framework. arXiv preprint arXiv:2203.07722 , 2022

2022 arXiv
[26]

Rafe: ranking feedback improves query rewriting for rag

Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Rafe: ranking feedback improves query rewriting for rag. arXiv preprint arXiv:2405.14431 , 2024

2024 arXiv
[27]

Fit-rag: black-box rag with factual information and token reduction

Yuren Mao, Xuemei Dong, Wenyi Xu, Yunjun Gao, Bin Wei, and Ying Zhang. Fit-rag: black-box rag with factual information and token reduction. arXiv preprint arXiv:2403.14374 , 2024

2024 arXiv
[28]

Bao: Making learned query optimization practical

Ryan Marcus, Parimarjan Negi, Hongzi Mao, Nesime Tatbul, Mohammad Alizadeh, and Tim Kraska. Bao: Making learned query optimization practical. In Proceedings of the 2021 International Conference on Management of Data , pages 1275--1288, 2021

2021
[29]

Neo: A learned query optimizer

Ryan Marcus, Parimarjan Negi, Hongzi Mao, Chi Zhang, Mohammad Alizadeh, Tim Kraska, Olga Papaemmanouil, and Nesime Tatbul. Neo: A learned query optimizer. Proc. VLDB Endow. , 12(11):1705–1718, jul 2019

2019
[30]

Lemo: A cache-enhanced learned optimizer for concurrent queries

Songsong Mo, Yile Chen, Hao Wang, Gao Cong, and Zhifeng Bao. Lemo: A cache-enhanced learned optimizer for concurrent queries. Proceedings of the ACM on Management of Data , 1(4):1--26, 2023

2023
[31]

Mysql 8.4

MySQL. Mysql 8.4. https://dev.mysql.com/downloads/mysql/, 2025

2025
[32]

Webgpt: Browser-assisted question-answering with human feedback

Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 , 2021

2021 arXiv
[33]

A comprehensive overview of large language models

Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 , 2023

2023 arXiv
[34]

Flow-loss: Learning cardinality estimates that matter

Parimarjan Negi, Ryan Marcus, Andreas Kipf, Hongzi Mao, Nesime Tatbul, Tim Kraska, and Mohammad Alizadeh. Flow-loss: Learning cardinality estimates that matter. Proc. VLDB Endow. , 14(11):2019–2032, jul 2021

2019
[35]

GPT-o3‑mini Model Documentation , 2025

OpenAI . GPT-o3‑mini Model Documentation , 2025. Accessed: 2025-07-30

2025
[36]

pg\_hint\_plan

OSSC-DB. pg\_hint\_plan. https://github.com/ossc-db/pg_hint_plan
[37]

Smaug: Fixing failure modes of preference optimisation with dpo-positive, 2024

Arka Pal, Deep Karkhanis, Samuel Dooley, Manley Roberts, Siddartha Naidu, and Colin White. Smaug: Fixing failure modes of preference optimisation with dpo-positive, 2024

2024
[38]

https://www.postgresql.org/

PostgreSQL. https://www.postgresql.org/
[39]

https://www.postgresql.org/

PostgreSQL Global Development Group . https://www.postgresql.org/
[40]

Direct preference optimization: Your language model is secretly a reward model

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems , 36:53728--53741, 2023

2023
[41]

Maximizing rag efficiency: A comparative analysis of rag methods

Tolga S akar and Hakan Emekci. Maximizing rag efficiency: A comparative analysis of rag methods. Natural Language Processing , 31(1):1--25, 2025

2025
[42]

Picard: Parsing incrementally for constrained auto-regressive decoding from language models

Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. Picard: Parsing incrementally for constrained auto-regressive decoding from language models. arXiv preprint arXiv:2109.05093 , 2021

2021 arXiv
[43]

Proximal policy optimization algorithms

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 , 2017

2017 arXiv
[44]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024

2024
[45]

Improving retrieval-augmented text-to-sql with ast-based ranking and schema pruning

Zhili Shen, Pavlos Vougiouklis, Chenxin Diao, Kaustubh Vyas, Yuanyi Ji, and Jeff Z Pan. Improving retrieval-augmented text-to-sql with ast-based ranking and schema pruning. arXiv preprint arXiv:2407.03227 , 2024

2024 arXiv
[46]

R-bot: An llm-based query rewrite system

Zhaoyan Sun, Xuanhe Zhou, and Guoliang Li. R-bot: An llm-based query rewrite system. arXiv preprint arXiv:2412.01661 , 2024

2024 arXiv
[47]

Vector database management systems: Fundamental concepts, use-cases, and current challenges

Toni Taipalus. Vector database management systems: Fundamental concepts, use-cases, and current challenges. Cognitive Systems Research , 85:101216, 2024

2024
[48]

Can large language models be query optimizer for relational databases? arXiv preprint arXiv:2502.05562 , 2025

Jie Tan, Kangfei Zhao, Rui Li, Jeffrey Xu Yu, Chengzhi Piao, Hong Cheng, Helen Meng, Deli Zhao, and Yu Rong. Can large language models be query optimizer for relational databases? arXiv preprint arXiv:2502.05562 , 2025

2025 arXiv
[49]

reads the manual

Immanuel Trummer. Db-bert: a database tuning tool that" reads the manual". In Proceedings of the 2022 international conference on management of data , pages 190--203, 2022

2022
[50]

Demonstrating gpt-db: Generating query-specific and customizable code for sql processing with gpt-4

Immanuel Trummer. Demonstrating gpt-db: Generating query-specific and customizable code for sql processing with gpt-4. Proceedings of the VLDB Endowment , 16(12):4098--4101, 2023

2023
[51]

Generating succinct descriptions of database schemata for cost-efficient prompting of large language models

Immanuel Trummer. Generating succinct descriptions of database schemata for cost-efficient prompting of large language models. Proc. VLDB Endow. , 17(11):3511–3523, July 2024

2024
[52]

Generating highly customizable python code for data processing with large language models

Immanuel Trummer. Generating highly customizable python code for data processing with large language models. The VLDB Journal , 34(2):21, 2025

2025
[53]

Caesura: Language models as multi-modal query planners

Matthias Urban and Carsten Binnig. Caesura: Language models as multi-modal query planners. arXiv preprint arXiv:2308.03424 , 2023

2023 arXiv
[54]

Milvus: A purpose-built vector data management system

Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xiangyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, et al. Milvus: A purpose-built vector data management system. In Proceedings of the 2021 International Conference on Management of Data , pages 2614--2...

2021
[55]

Peiyi Wang, Lei Li, Zhihong Shao, R. X. Xu, Damai Dai, Yifei Li, Deli Chen, Y. Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations, 2024

2024
[56]

The learnability of in-context learning

Noam Wies, Yoav Levine, and Amnon Shashua. The learnability of in-context learning. Advances in Neural Information Processing Systems , 36:36637--36651, 2023

2023
[57]

Fastgres: Making learned query optimizer hinting effective

Lucas Woltmann, Jerome Thiessat, Claudio Hartmann, Dirk Habich, and Wolfgang Lehner. Fastgres: Making learned query optimizer hinting effective. Proc. VLDB Endow. , 16(11):3310–3322, July 2023

2023
[58]

An explanation of in-context learning as implicit bayesian inference

Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080 , 2021

2021 arXiv
[59]

Balsa: Learning a query optimizer without expert demonstrations

Zongheng Yang, Wei-Lin Chiang, Sifei Luan, Gautam Mittal, Michael Luo, and Ion Stoica. Balsa: Learning a query optimizer without expert demonstrations. In Proceedings of the 2022 International Conference on Management of Data , pages 931--944, 2022

2022
[60]

A query optimization method utilizing large language models

Zhiming Yao, Haoyang Li, Jing Zhang, Cuiping Li, and Hong Chen. A query optimization method utilizing large language models. arXiv preprint arXiv:2503.06902 , 2025

2025 arXiv
[61]

Cost-based or learning-based? a hybrid query optimizer for query plan selection

Xiang Yu, Chengliang Chai, Guoliang Li, and Jiabin Liu. Cost-based or learning-based? a hybrid query optimizer for query plan selection. Proceedings of the VLDB Endowment , 15(13):3924--3936, 2022

2022
[62]

Reinforcement learning with tree-lstm for join order selection

Xiang Yu, Guoliang Li, Chengliang Chai, and Nan Tang. Reinforcement learning with tree-lstm for join order selection. In 2020 IEEE 36th international conference on data engineering (ICDE) , pages 1297--1308. IEEE, 2020

2020
[63]

Refsql: A retrieval-augmentation framework for text-to-sql generation

Kun Zhang, Xiexiong Lin, Yuanzhuo Wang, Xin Zhang, Fei Sun, Cen Jianhe, Hexiang Tan, Xuhui Jiang, and Huawei Shen. Refsql: A retrieval-augmentation framework for text-to-sql generation. In Findings of the Association for Computational Linguistics: EMNLP 2023 , pages 664--673, 2023

2023
[64]

Deploying a steered query optimizer in production at microsoft

Wangda Zhang, Matteo Interlandi, Paul Mineiro, Shi Qiao, Nasim Ghazanfari, Karlen Lie, Marc Friedman, Rafah Hosn, Hiren Patel, and Alekh Jindal. Deploying a steered query optimizer in production at microsoft. In Proceedings of the 2022 International Conference on Management of...

2022
[65]

Lero: A learning-to-rank query optimizer

Rong Zhu, Wei Chen, Bolin Ding, Xingguang Chen, Andreas Pfadler, Ziniu Wu, and Jingren Zhou. Lero: A learning-to-rank query optimizer. Proceedings of the VLDB Endowment , 16(6):1466--1479, 2023

2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Many-shot in-context learning

Rishabh Agarwal, Avi Singh, Lei Zhang, Bernd Bohnet, Luis Rosias, Stephanie Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, et al. Many-shot in-context learning. Advances in Neural Information Processing Systems , 37:76930--76966, 2024

work page 2024

[2] [2]

The unreasonable effectiveness of LLMs for query optimization

Peter Akioyamen, Zixuan Yi, and Ryan Marcus. The unreasonable effectiveness of LLMs for query optimization. In Machine Learning for Systems Workshop at the 38th Conference on Neural Information Processing Systems (NeurIPS 2024) , 2024

work page 2024

[3] [3]

What learning algorithm is in-context learning? investigations with linear models

Ekin Aky \"u rek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learning algorithm is in-context learning? investigations with linear models. arXiv preprint arXiv:2211.15661 , 2022

arXiv 2022

[4] [4]

https://www.tpc.org/tpcds/

TPC-DS Benchmark. https://www.tpc.org/tpcds/

work page

[5] [5]

A Robust and Explainable Query Optimization Cost Model Based on Bidirectional Graph Neural Networks

Baoming Chang. A Robust and Explainable Query Optimization Cost Model Based on Bidirectional Graph Neural Networks . PhD thesis, University of Ottawa, 2024

work page 2024

[6] [6]

Loger: A learned optimizer towards generating efficient and robust query execution plans

Tianyi Chen, Jun Gao, Hedui Chen, and Yaofeng Tu. Loger: A learned optimizer towards generating efficient and robust query execution plans. Proceedings of the VLDB Endowment , 16(7):1777--1789, 2023

work page 2023

[7] [7]

Murag: Multimodal retrieval-augmented generator for open question answering over images and text

Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William W Cohen. Murag: Multimodal retrieval-augmented generator for open question answering over images and text. arXiv preprint arXiv:2210.02928 , 2022

arXiv 2022

[8] [8]

Leon: A new framework for ml-aided query optimization

Xu Chen, Haitian Chen, Zibo Liang, Shuncheng Liu, Jinghong Wang, Kai Zeng, Han Su, and Kai Zheng. Leon: A new framework for ml-aided query optimization. Proceedings of the VLDB Endowment , 16(9):2261--2273, 2023

work page 2023