@inproceedings{hu-etal-2026-optimizing-reasoning,
title = "Optimizing Reasoning {LLM}s for Automated Feature Engineering via Multi-Stage Direct Preference Optimization",
author = "Hu, Mingzhi and
Pan, Menghai and
Chen, Yuzhong and
Wang, Junpeng and
Dou, Yingtong and
Rakesh, Vineeth and
Le, Hoang Anh Duy and
Saini, Uday Singh and
Dai, Xin and
Fan, Xiran and
Liu, Yuqing and
Cai, Yiwei",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.233/",
pages = "4599--4618",
ISBN = "979-8-89176-444-6",
abstract = "Feature engineering is central to tabular machine learning, but useful transformations must be both statistically predictive and semantically valid with respect to feature names, task descriptions, and observed data patterns. Although large language models can exploit such context, prompt-based feature engineering remains unreliable: generated transformations may be redundant, schema-invalid, non-executable, or weakly aligned with downstream performance. We propose \textsc{MSDPO}, a two-stage Direct Preference Optimization framework for automated feature engineering. Proposal-DPO aligns the model toward useful, diverse, and non-redundant feature proposals, while Function-DPO aligns it toward executable feature-construction code using execution and validation feedback. After alignment, the frozen model directly generates transformation functions for unseen tabular tasks. Across 20 classification datasets, \textsc{MSDPO} achieves the best overall mean rank among strong classical and LLM-based feature-engineering baselines. Under a budget-matched comparison inside the same feature-engineering harness, MSDPO also attains a better mean rank than Claude Sonnet 4.6, despite using an 8B open-weight backbone. Moreover, preferences constructed from classification signals transfer to four regression tasks. Ablation results show that the two DPO stages are complementary, improving both feature design and executable implementation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hu-etal-2026-optimizing-reasoning">
<titleInfo>
<title>Optimizing Reasoning LLMs for Automated Feature Engineering via Multi-Stage Direct Preference Optimization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mingzhi</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Menghai</namePart>
<namePart type="family">Pan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuzhong</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junpeng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yingtong</namePart>
<namePart type="family">Dou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vineeth</namePart>
<namePart type="family">Rakesh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hoang</namePart>
<namePart type="given">Anh</namePart>
<namePart type="given">Duy</namePart>
<namePart type="family">Le</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Uday</namePart>
<namePart type="given">Singh</namePart>
<namePart type="family">Saini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">Dai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiran</namePart>
<namePart type="family">Fan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuqing</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yiwei</namePart>
<namePart type="family">Cai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Feature engineering is central to tabular machine learning, but useful transformations must be both statistically predictive and semantically valid with respect to feature names, task descriptions, and observed data patterns. Although large language models can exploit such context, prompt-based feature engineering remains unreliable: generated transformations may be redundant, schema-invalid, non-executable, or weakly aligned with downstream performance. We propose MSDPO, a two-stage Direct Preference Optimization framework for automated feature engineering. Proposal-DPO aligns the model toward useful, diverse, and non-redundant feature proposals, while Function-DPO aligns it toward executable feature-construction code using execution and validation feedback. After alignment, the frozen model directly generates transformation functions for unseen tabular tasks. Across 20 classification datasets, MSDPO achieves the best overall mean rank among strong classical and LLM-based feature-engineering baselines. Under a budget-matched comparison inside the same feature-engineering harness, MSDPO also attains a better mean rank than Claude Sonnet 4.6, despite using an 8B open-weight backbone. Moreover, preferences constructed from classification signals transfer to four regression tasks. Ablation results show that the two DPO stages are complementary, improving both feature design and executable implementation.</abstract>
<identifier type="citekey">hu-etal-2026-optimizing-reasoning</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.233/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>4599</start>
<end>4618</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Optimizing Reasoning LLMs for Automated Feature Engineering via Multi-Stage Direct Preference Optimization
%A Hu, Mingzhi
%A Pan, Menghai
%A Chen, Yuzhong
%A Wang, Junpeng
%A Dou, Yingtong
%A Rakesh, Vineeth
%A Le, Hoang Anh Duy
%A Saini, Uday Singh
%A Dai, Xin
%A Fan, Xiran
%A Liu, Yuqing
%A Cai, Yiwei
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F hu-etal-2026-optimizing-reasoning
%X Feature engineering is central to tabular machine learning, but useful transformations must be both statistically predictive and semantically valid with respect to feature names, task descriptions, and observed data patterns. Although large language models can exploit such context, prompt-based feature engineering remains unreliable: generated transformations may be redundant, schema-invalid, non-executable, or weakly aligned with downstream performance. We propose MSDPO, a two-stage Direct Preference Optimization framework for automated feature engineering. Proposal-DPO aligns the model toward useful, diverse, and non-redundant feature proposals, while Function-DPO aligns it toward executable feature-construction code using execution and validation feedback. After alignment, the frozen model directly generates transformation functions for unseen tabular tasks. Across 20 classification datasets, MSDPO achieves the best overall mean rank among strong classical and LLM-based feature-engineering baselines. Under a budget-matched comparison inside the same feature-engineering harness, MSDPO also attains a better mean rank than Claude Sonnet 4.6, despite using an 8B open-weight backbone. Moreover, preferences constructed from classification signals transfer to four regression tasks. Ablation results show that the two DPO stages are complementary, improving both feature design and executable implementation.
%U https://aclanthology.org/2026.findings-aacl.233/
%P 4599-4618
Markdown (Informal)
[Optimizing Reasoning LLMs for Automated Feature Engineering via Multi-Stage Direct Preference Optimization](https://aclanthology.org/2026.findings-aacl.233/) (Hu et al., Findings 2026)
ACL
- Mingzhi Hu, Menghai Pan, Yuzhong Chen, Junpeng Wang, Yingtong Dou, Vineeth Rakesh, Hoang Anh Duy Le, Uday Singh Saini, Xin Dai, Xiran Fan, Yuqing Liu, and Yiwei Cai. 2026. Optimizing Reasoning LLMs for Automated Feature Engineering via Multi-Stage Direct Preference Optimization. In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026, pages 4599–4618, Hengqin, China. Association for Computational Linguistics.