@inproceedings{peng-etal-2026-homogeneity,
title = "From Homogeneity to Diversity: Self-Refined {GRPO} for Controllable Creative Writing",
author = "Peng, Minna and
Tan, Wei and
Wang, Zeyu and
Hu, Yongquan and
Huang, Ruizhi and
Jie, Yang",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.972/",
pages = "20555--20571",
ISBN = "979-8-89176-462-0",
abstract = "Controllable creative writing is a text generation task that requires both high generation quality and diverse outputs. However, existing methods often improve one aspect at the expense of the other. To address this challenge, we propose SERGE, a reinforcement learning (RL) framework that introduces a diversity-guided self-refinement mechanism, where the policy itself refines low-diversity candidate texts, and employs policy shaping to amplify the gradient contributions of samples with both high generation quality and diversity. We also propose a fine-grained reward modeling framework to provide reliable quality reward signals to support RL optimization. Experiments on the short-form video copywriting generation task show that SERGE outperforms the base model and conventional diversity-enhancement strategies on both diversity metrics (DisNGram, PairBLEU, CosSim) and generation quality, effectively alleviating the ``diversity collapse'' issue in RL. Our code is available at: https://github.com/peng-mn/SERGE."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="peng-etal-2026-homogeneity">
<titleInfo>
<title>From Homogeneity to Diversity: Self-Refined GRPO for Controllable Creative Writing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Minna</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Tan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeyu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yongquan</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruizhi</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yang</namePart>
<namePart type="family">Jie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Controllable creative writing is a text generation task that requires both high generation quality and diverse outputs. However, existing methods often improve one aspect at the expense of the other. To address this challenge, we propose SERGE, a reinforcement learning (RL) framework that introduces a diversity-guided self-refinement mechanism, where the policy itself refines low-diversity candidate texts, and employs policy shaping to amplify the gradient contributions of samples with both high generation quality and diversity. We also propose a fine-grained reward modeling framework to provide reliable quality reward signals to support RL optimization. Experiments on the short-form video copywriting generation task show that SERGE outperforms the base model and conventional diversity-enhancement strategies on both diversity metrics (DisNGram, PairBLEU, CosSim) and generation quality, effectively alleviating the “diversity collapse” issue in RL. Our code is available at: https://github.com/peng-mn/SERGE.</abstract>
<identifier type="citekey">peng-etal-2026-homogeneity</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.972/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>20555</start>
<end>20571</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Homogeneity to Diversity: Self-Refined GRPO for Controllable Creative Writing
%A Peng, Minna
%A Tan, Wei
%A Wang, Zeyu
%A Hu, Yongquan
%A Huang, Ruizhi
%A Jie, Yang
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F peng-etal-2026-homogeneity
%X Controllable creative writing is a text generation task that requires both high generation quality and diverse outputs. However, existing methods often improve one aspect at the expense of the other. To address this challenge, we propose SERGE, a reinforcement learning (RL) framework that introduces a diversity-guided self-refinement mechanism, where the policy itself refines low-diversity candidate texts, and employs policy shaping to amplify the gradient contributions of samples with both high generation quality and diversity. We also propose a fine-grained reward modeling framework to provide reliable quality reward signals to support RL optimization. Experiments on the short-form video copywriting generation task show that SERGE outperforms the base model and conventional diversity-enhancement strategies on both diversity metrics (DisNGram, PairBLEU, CosSim) and generation quality, effectively alleviating the “diversity collapse” issue in RL. Our code is available at: https://github.com/peng-mn/SERGE.
%U https://aclanthology.org/2026.findings-emnlp.972/
%P 20555-20571
Markdown (Informal)
[From Homogeneity to Diversity: Self-Refined GRPO for Controllable Creative Writing](https://aclanthology.org/2026.findings-emnlp.972/) (Peng et al., Findings 2026)
ACL