@inproceedings{rajendhran-etal-2026-polaris,
title = "{POLARIS}: Guiding Small Models to Write Long Stories",
author = "Rajendhran, Rishanth and
Russell, Jenna and
Iyyer, Mohit and
Wieting, John Frederick",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.627/",
pages = "14392--14428",
ISBN = "979-8-89176-459-0",
abstract = "Small open-weight models struggle at long-form creative writing: their generated stories either fall far short of the requested length, or their quality significantly degrades as length increases, especially when compared to frontier models. We present **POLARIS** (**P**olicy **O**ptimization with **L**LM-as-a-judge rewards and **A**nchored-**R**eference **I**njection for **S**torywriting), **a lower-compute GRPO recipe** with two key ingredients: a frontier LLM judge with a **structured *Story Quality* rubric** as the online reward, and **human-reference injection (HRI)**, where a teacher-forced human-written story serves as a high-reward anchor within each GRPO group. By applying our training recipe to Qwen3.5-9B, using a dataset of $\sim$1.4K prompt{--}story pairs derived from 100 short-story anthologies and 4 A100 GPUs, we obtain **POLARIS-9B**. Across five benchmarks spanning in-distribution and out-of-distribution prompts and rubrics, POLARIS-9B is **competitive with much larger open-weight models** while following length instructions more closely. A blinded human evaluation confirms that POLARIS-9B is preferred to the base Qwen3.5-9B and on par with Qwen3.5-27B. Despite training only on stories up to 4k words, POLARIS-9B preserves quality on prompts requesting stories up to $3\times$ the training length, a regime where most open-weight models degrade substantially in quality, length adherence, or both. More broadly, our results suggest that **length generalization is a meaningful stress test for creative-writing models** and a useful lens for distinguishing otherwise close models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="rajendhran-etal-2026-polaris">
<titleInfo>
<title>POLARIS: Guiding Small Models to Write Long Stories</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rishanth</namePart>
<namePart type="family">Rajendhran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jenna</namePart>
<namePart type="family">Russell</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohit</namePart>
<namePart type="family">Iyyer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">John</namePart>
<namePart type="given">Frederick</namePart>
<namePart type="family">Wieting</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Small open-weight models struggle at long-form creative writing: their generated stories either fall far short of the requested length, or their quality significantly degrades as length increases, especially when compared to frontier models. We present **POLARIS** (**P**olicy **O**ptimization with **L**LM-as-a-judge rewards and **A**nchored-**R**eference **I**njection for **S**torywriting), **a lower-compute GRPO recipe** with two key ingredients: a frontier LLM judge with a **structured *Story Quality* rubric** as the online reward, and **human-reference injection (HRI)**, where a teacher-forced human-written story serves as a high-reward anchor within each GRPO group. By applying our training recipe to Qwen3.5-9B, using a dataset of \sim1.4K prompt–story pairs derived from 100 short-story anthologies and 4 A100 GPUs, we obtain **POLARIS-9B**. Across five benchmarks spanning in-distribution and out-of-distribution prompts and rubrics, POLARIS-9B is **competitive with much larger open-weight models** while following length instructions more closely. A blinded human evaluation confirms that POLARIS-9B is preferred to the base Qwen3.5-9B and on par with Qwen3.5-27B. Despite training only on stories up to 4k words, POLARIS-9B preserves quality on prompts requesting stories up to 3\times the training length, a regime where most open-weight models degrade substantially in quality, length adherence, or both. More broadly, our results suggest that **length generalization is a meaningful stress test for creative-writing models** and a useful lens for distinguishing otherwise close models.</abstract>
<identifier type="citekey">rajendhran-etal-2026-polaris</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.627/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>14392</start>
<end>14428</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T POLARIS: Guiding Small Models to Write Long Stories
%A Rajendhran, Rishanth
%A Russell, Jenna
%A Iyyer, Mohit
%A Wieting, John Frederick
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F rajendhran-etal-2026-polaris
%X Small open-weight models struggle at long-form creative writing: their generated stories either fall far short of the requested length, or their quality significantly degrades as length increases, especially when compared to frontier models. We present **POLARIS** (**P**olicy **O**ptimization with **L**LM-as-a-judge rewards and **A**nchored-**R**eference **I**njection for **S**torywriting), **a lower-compute GRPO recipe** with two key ingredients: a frontier LLM judge with a **structured *Story Quality* rubric** as the online reward, and **human-reference injection (HRI)**, where a teacher-forced human-written story serves as a high-reward anchor within each GRPO group. By applying our training recipe to Qwen3.5-9B, using a dataset of \sim1.4K prompt–story pairs derived from 100 short-story anthologies and 4 A100 GPUs, we obtain **POLARIS-9B**. Across five benchmarks spanning in-distribution and out-of-distribution prompts and rubrics, POLARIS-9B is **competitive with much larger open-weight models** while following length instructions more closely. A blinded human evaluation confirms that POLARIS-9B is preferred to the base Qwen3.5-9B and on par with Qwen3.5-27B. Despite training only on stories up to 4k words, POLARIS-9B preserves quality on prompts requesting stories up to 3\times the training length, a regime where most open-weight models degrade substantially in quality, length adherence, or both. More broadly, our results suggest that **length generalization is a meaningful stress test for creative-writing models** and a useful lens for distinguishing otherwise close models.
%U https://aclanthology.org/2026.emnlp-main.627/
%P 14392-14428
Markdown (Informal)
[POLARIS: Guiding Small Models to Write Long Stories](https://aclanthology.org/2026.emnlp-main.627/) (Rajendhran et al., EMNLP 2026)
ACL
- Rishanth Rajendhran, Jenna Russell, Mohit Iyyer, and John Frederick Wieting. 2026. POLARIS: Guiding Small Models to Write Long Stories. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 14392–14428, Budapest, Hungary. Association for Computational Linguistics.