@inproceedings{yun-etal-2025-ultrabench,
title = "{ULTRABENCH}: Benchmarking {LLM}s under Extreme Fine-grained Text Generation",
author = "Yun, Longfei and
Peng, Letian and
Shang, Jingbo",
editor = "Christodoulopoulos, Christos and
Chakraborty, Tanmoy and
Rose, Carolyn and
Peng, Violet",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
month = nov,
year = "2025",
address = "Suzhou, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-emnlp.835/",
pages = "15438--15453",
ISBN = "979-8-89176-335-7",
abstract = "Fine-grained control is essential for precise and customizable text generation, yet existing benchmarks evaluate models on only a few attributes, typically fewer than five. We introduce UltraBench, a new benchmark for extremely fine-grained controllable generation (EFCG), which evaluates large language models (LLMs) under dense, multi-attribute constraints. Each sample includes approximately 70 attributes, combining LLM-extracted soft constraints (e.g., style and tone) with programmatically enforced hard constraints (e.g., word count). Using UltraBench, we conduct a comprehensive evaluation of state-of-the-art LLMs and prompting strategies. Models such as GPT-4.1 and Qwen3-8B perform well on individual constraints, achieving instruction-level accuracy above 70{\%}, but consistently fail to satisfy all constraints simultaneously. To understand this limitation, we analyze model behavior across different dimensions. First, we observe a clear position bias: models tend to prioritize constraints presented later in the prompt while neglecting those that appear earlier. Second, we find that structural and formatting-related constraints are significantly more difficult to satisfy than content- or style-based ones, suggesting that current models struggle to coordinate global structure with token-level control. Finally, our error analysis reveals distinct failure modes: GPT-4.1 often attempts to follow constraints but falls short in precision, whereas LLaMA frequently omits constraints, particularly in multi-turn settings. These findings highlight fundamental limitations in EFCG and underscore the need for new methods that support dense, instruction-sensitive generation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yun-etal-2025-ultrabench">
<titleInfo>
<title>ULTRABENCH: Benchmarking LLMs under Extreme Fine-grained Text Generation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Longfei</namePart>
<namePart type="family">Yun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Letian</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingbo</namePart>
<namePart type="family">Shang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2025</title>
</titleInfo>
<name type="personal">
<namePart type="given">Christos</namePart>
<namePart type="family">Christodoulopoulos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tanmoy</namePart>
<namePart type="family">Chakraborty</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Carolyn</namePart>
<namePart type="family">Rose</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Violet</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Suzhou, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-335-7</identifier>
</relatedItem>
<abstract>Fine-grained control is essential for precise and customizable text generation, yet existing benchmarks evaluate models on only a few attributes, typically fewer than five. We introduce UltraBench, a new benchmark for extremely fine-grained controllable generation (EFCG), which evaluates large language models (LLMs) under dense, multi-attribute constraints. Each sample includes approximately 70 attributes, combining LLM-extracted soft constraints (e.g., style and tone) with programmatically enforced hard constraints (e.g., word count). Using UltraBench, we conduct a comprehensive evaluation of state-of-the-art LLMs and prompting strategies. Models such as GPT-4.1 and Qwen3-8B perform well on individual constraints, achieving instruction-level accuracy above 70%, but consistently fail to satisfy all constraints simultaneously. To understand this limitation, we analyze model behavior across different dimensions. First, we observe a clear position bias: models tend to prioritize constraints presented later in the prompt while neglecting those that appear earlier. Second, we find that structural and formatting-related constraints are significantly more difficult to satisfy than content- or style-based ones, suggesting that current models struggle to coordinate global structure with token-level control. Finally, our error analysis reveals distinct failure modes: GPT-4.1 often attempts to follow constraints but falls short in precision, whereas LLaMA frequently omits constraints, particularly in multi-turn settings. These findings highlight fundamental limitations in EFCG and underscore the need for new methods that support dense, instruction-sensitive generation.</abstract>
<identifier type="citekey">yun-etal-2025-ultrabench</identifier>
<location>
<url>https://aclanthology.org/2025.findings-emnlp.835/</url>
</location>
<part>
<date>2025-11</date>
<extent unit="page">
<start>15438</start>
<end>15453</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ULTRABENCH: Benchmarking LLMs under Extreme Fine-grained Text Generation
%A Yun, Longfei
%A Peng, Letian
%A Shang, Jingbo
%Y Christodoulopoulos, Christos
%Y Chakraborty, Tanmoy
%Y Rose, Carolyn
%Y Peng, Violet
%S Findings of the Association for Computational Linguistics: EMNLP 2025
%D 2025
%8 November
%I Association for Computational Linguistics
%C Suzhou, China
%@ 979-8-89176-335-7
%F yun-etal-2025-ultrabench
%X Fine-grained control is essential for precise and customizable text generation, yet existing benchmarks evaluate models on only a few attributes, typically fewer than five. We introduce UltraBench, a new benchmark for extremely fine-grained controllable generation (EFCG), which evaluates large language models (LLMs) under dense, multi-attribute constraints. Each sample includes approximately 70 attributes, combining LLM-extracted soft constraints (e.g., style and tone) with programmatically enforced hard constraints (e.g., word count). Using UltraBench, we conduct a comprehensive evaluation of state-of-the-art LLMs and prompting strategies. Models such as GPT-4.1 and Qwen3-8B perform well on individual constraints, achieving instruction-level accuracy above 70%, but consistently fail to satisfy all constraints simultaneously. To understand this limitation, we analyze model behavior across different dimensions. First, we observe a clear position bias: models tend to prioritize constraints presented later in the prompt while neglecting those that appear earlier. Second, we find that structural and formatting-related constraints are significantly more difficult to satisfy than content- or style-based ones, suggesting that current models struggle to coordinate global structure with token-level control. Finally, our error analysis reveals distinct failure modes: GPT-4.1 often attempts to follow constraints but falls short in precision, whereas LLaMA frequently omits constraints, particularly in multi-turn settings. These findings highlight fundamental limitations in EFCG and underscore the need for new methods that support dense, instruction-sensitive generation.
%U https://aclanthology.org/2025.findings-emnlp.835/
%P 15438-15453
Markdown (Informal)
[ULTRABENCH: Benchmarking LLMs under Extreme Fine-grained Text Generation](https://aclanthology.org/2025.findings-emnlp.835/) (Yun et al., Findings 2025)
ACL