Text simplification supports second language (L2) learning by providing comprehensible input, consistent with the Input Hypothesis. However, constructing personalized parallel corpora is costly, while existing large language model (LLM)-based readability control methods rely on pre-labeled sentence corpora and primarily target English. We propose Re-RIGHT, a unified reinforcement learning framework for adaptive multilingual text simplification without parallel corpus supervision. We first show that prompting-based lexical simplification at target proficiency levels (CEFR, JLPT, TOPIK, and HSK) performs poorly at easier levels and for non-English languages, even with state-of-the-art LLMs such as GPT-5.2 and Gemini 2.5. To address this, we collect 43K vocabulary-level data across four languages (English, Japanese, Korean, and Chinese) and train a compact 4B policy model using Re-RIGHT, which integrates three reward modules: vocabulary coverage, semantic preservation, and coherence. Compared to the stronger LLM baselines, Re-RIGHT achieves higher lexical coverage at target proficiency levels while maintaining original meaning and fluency.
Figures & tables
Figure 1: We propose Re-RIGHT , a unified multilingual GRPO training framework for text simplification at learners’ proficiency levels (§ 6 ) without relying on parallel training corpora, overcoming the limitation of the prompting-based approach (§ 4 ) that fails to meet the target vocabulary coverage. (1) At the preparation phase, we collect “Featured Articles” in Wikipedia as a training seed dataset, while constructing 43,786 vocabulary level data from official proficiency standards across four languages. (2) We then train a 4B policy model with three reward modules: vocabulary coverage, semantic preservation, and coherence. (3) The trained policy model enables adaptive multilingual text simplification with higher vocabulary coverage than GPT-5.2 and Gemini 2.5, generating level-appropriate, semantically preserved, and coherent outputs.
Figure 2: Vocabulary coverage scores (percentage of content words generated at or below the target level) across models and target level variants for each language. The results show that as the target vocabulary level becomes easier, almost all models, including GPT-5.2, Gemini 2.5, and Claude 4.5, fail to provide high vocabulary coverage via the prompting-based approach. The black line box means the average score across models, while the gray dotted line indicates the baseline score of the reference texts. The color depth in each bar plot also represents average vocabulary coverage of each proficiency level from model-generated results for that language (i.e., the lightest color stands for the easiest level), and the gray parts on the top indicate unknown words beyond all levels.
Text Simplification Test
You are a careful rewrite assistant.
Rewrite the <TEXT> in {language} so that every word, except proper nouns or proper adjectives, is at or below the {level} vocabulary level.
Replace or simplify any other words above {level} level with easier alternatives while preserving the original meaning and coherence.
Do not skip, shorten, or omit any part of the text. Keep sentence count and structure.
Output only the fully converted text with no explanations, instructions, or extra words.
<TEXT>
Table 1: Text simplification prompt used in § 4 and § 6 . We provide LLMs with the original text along with the target language and proficiency level, requesting text simplification that preserves the original meaning.
Figure 3: Entailment scoring process for the semantic preservation reward. We introduce a two-phase approach with (1) greedy pairing using BERTScore and (2) entailment scoring via an NLI predictor.
Lang.
Method
Vocabulary Coverage
Semantic Pres.
Coherence
Total (Std.)
Easy (Std.)
Total
Easy
Total
Easy
EN
Reference
58.3 (24.2)
28.9 (12.6)
–
–
–
–
Base (Qwen3-4B)
72.6 (19.2)
50.4 (14.4)
75.3
72.3
86.9
83.7
FUDGE (Qwen3-4B)
74.2 (18.4)
53.4 (14.4)
75.0
71.8
86.7
83.1
GPT-5.2
71.0 (16.8)
52.4 (12.2)
76.1
70.4
84.6
75.8
Gemini 2.5
77.7 (16.3)
59.1 (13.5)
72.0
69.2
82.8
78.7
Table 2: Evaluation results across four languages, reporting averaged vocabulary coverage, semantic preservation, and coherence scores for overall proficiency levels (Total) and easier-level subsets (Easy: average over the two lowest levels for each language). Re-RIGHT outperforms baselines in vocabulary coverage while preserving meaning and fluency, only using a 4B policy model. In particular, Re-RIGHT improves text simplification performance at easier levels, while existing baselines (including GPT-5.2 and Gemini 2.5) fail to achieve high vocabulary coverage.
Lang.
Level
Text
EN
Reference
Pigeon photography is an aerial photography technique invented in 1907 by the German apothecary Julius Neubronner, who also used pigeons to deliver medications. …
CEFR A2
Pigeon photography is a way of taking photos from the air. It was made in 1907 by a man named Julius Neubronner from Germany. He also used pigeons to carry medicine. …
CEFR B2
Pigeon photography is a way of taking photos from the air that was first created in 1907 by a German man named Julius Neubronner. He also used pigeons to carry medicine. …
Table 3: Examples of proficiency-controlled simplification results across languages. Each block shows the reference and two simplified outputs aligned to different target levels. Without pre-labeled parallel corpora, Re-RIGHT enables the trained 4B policy model to generate level-appropriate sentences, preserving information and coherence.
Figure 4: Average evaluation scores across languages on out-of-domain data (PGV), compared to the Wikipedia test data. We show that Re-RIGHT endows the model with general simplification abilities from complex seed texts of an arbitrary domain. The black dashed line represents the performance of the 4B base model.
La.
Method
Voc.
Sem.
Coh.
EN
Base (Qw3 4B)
72.6
75.3
86.9
Base (Ge3 4B)
74.5
69.0
84.7
Re-RIGHT (Qw3 4B)
81.6
80.8
82.9
Re-RIGHT (Ge3 4B)
81.5
79.9
80.8
JA
Base (Qw3 4B)
53.9
65.4
88.7
Base (Ge3 4B)
50.2
61.6
90.0
Table 4: Performances of Re-RIGHT on Gemma 3 4B, compared to the baselines with Qwen3 4B. We show that Re-RIGHT can be generalized to other models.
Appendix figures & tables27 assets
Supplementary material from the paper’s appendix.
Appendix
Lang.
Criterion
Level
# Lemma
EN
CEFR
A1, A2, B1, B2, C1, C2
9413
JA
JLPT
N5, N4, N3, N2, N1
14146
KO
TOPIK
Level 1, 2, 3, 4, 5, 6
9357
ZH
HSK 3.0
Level 1, 2, 3, 4, 5, 6, 7-9
10870
Total
43786
Appendix
Table 5: Overall summary for our vocabulary level data. We collect proficiency level information of lemmas across four languages (English, Japanese, Korean, and Chinese) based on authoritative examinational standards. Vocabulary coverage module utilizes this data to calculate the proficiency level adequacy of an input.
Dataset
Preprocessing
Wikipedia
• Document Filtering & Chunking – Extract ‘Featured articles’ documents – Detect paragraphs by double newline characters – Remove reference paragraphs or paragraphs with <20 tokens – Ensure each chunk is <=512 tokens – Maintain paragraph boundaries • Sampling – Duplicate each chunk by number of target levels – Uniformly sample from each language
PGV
• Document Filtering & Chunking – Remove paragraphs with key ‘crawlinfo’ – Remove paragraphs with key ‘type’ and value ‘title’, ‘caption’, ‘contributor’, ‘notes’, or ‘tweet-info’ – Concatenate all paragraphs with blank in between – Remove documents with <300 tokens – Truncate to retain <=512 tokens only • Sampling – Duplicate each chunk by number of target levels – Uniformly sample from each language
Appendix
Table 6: Checklist for dataset preprocessing.
Setting
Voc.
Sem.
Coh.
Vanilla
59.2
69.5
88.7
Full framework
76.0
80.6
83.1
1:1 entailment
74.0
73.3
83.7
No Sem.
84.0
52.7
83.1
No Coh.
88.6
79.4
54.6
Appendix
Table 7: Evaluation results from the ablation study.
Table 8: An example comparison between generations from the base 4B model with the prompting method and the trained 4B model with Re-RIGHT.
Reference
Judith Arlene Resnik (April 5, 1949 – January 28, 1986) was an American electrical engineer, software engineer, biomedical engineer, pilot and NASA astronaut who died in the Space Shuttle Challenger disaster.
CEFR A1
Judith Arlene Resnik (April 5, 1949 – January 28, 1986) was an American person who worked in electricity, software, medicine, flew planes , and also worked for NASA. She died in the Space Shuttle Challenger problem.
Appendix
Table 9: An example failure case: grammatical error.
Reference
어린 프리드리히는 1840년대 이래 독일에서 맹위를 떨치던 자유주의 세력 중심의 혼탁한 정국을 겪었다.
( Translation: Young Friedrich experienced the turbulent political climate dominated by liberal forces that had been rampant in Germany since the 1840s. )
TOPIK1
어린 프리드리히는 1840년대부터 독일에서 자유주의 세력이 강하게 나타나는 혼란스러운 정치 상황을 겪었다.
TOPIK6
어린 프리드리히는 1840년대부터 독일에서 자유주의 세력이 강하게 나타나는 혼란스러운 정치 상황을 겪었다.
Appendix
Table 10: An example failure case: proficiency level misalignment (identical outputs).
Table 11: Examples of non-parallel “Featured Article” data chunks, excerpted from different articles.
Table 12: Examples of Parallel Global Voices data chunks, excerpted from different articles.
Lang.
Level
Original
(Avg. ± Std.)
EN
A1
20.2 ± 7.5
A2
37.5 ± 10.5
B1
58.0 ± 10.4
B2
73.5 ± 9.0
C1
77.2 ± 8.5
Appendix
Table 13: Vocabulary coverage score baseline statistics for Reference Texts in Wikipedia Featured Article Dataset.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
39.8 ± 9.7
A2
60.9 ± 9.8
B1
75.0 ± 10.0
B2
84.8 ± 8.2
C1
85.5 ± 7.2
Appendix
Table 14: Vocabulary coverage score statistics for Qwen3-4B-Instruct-2507.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
29.0 ± 10.9
A2
47.4 ± 12.0
B1
65.5 ± 11.6
B2
78.7 ± 10.1
C1
79.8 ± 8.9
Appendix
Table 15: Vocabulary coverage score statistics for Qwen3-4B.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
35.3 ± 11.1
A2
54.9 ± 11.3
B1
72.9 ± 11.3
B2
83.6 ± 9.1
C1
83.2 ± 8.3
Appendix
Table 16: Vocabulary coverage score statistics for Qwen3-8B.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
33.9 ± 11.0
A2
54.6 ± 11.8
B1
69.5 ± 11.2
B2
80.4 ± 9.8
C1
78.9 ± 8.8
Appendix
Table 17: Vocabulary coverage score statistics for Qwen3-14B.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
38.3 ± 9.6
A2
59.6 ± 9.4
B1
73.9 ± 10.2
B2
83.2 ± 8.9
C1
84.1 ± 8.0
Appendix
Table 18: Vocabulary coverage score statistics for Qwen3-32B.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
48.3 ± 10.4
A2
65.6 ± 9.1
B1
76.9 ± 9.1
B2
83.0 ± 7.6
C1
84.4 ± 6.8
Appendix
Table 19: Vocabulary coverage score statistics for gemma-3-4b-it.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
57.0 ± 10.3
A2
73.8 ± 8.7
B1
82.5 ± 8.3
B2
87.1 ± 7.3
C1
87.2 ± 6.7
Appendix
Table 20: Vocabulary coverage score statistics for gemma-3-12b-it.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
53.4 ± 10.6
A2
72.4 ± 9.0
B1
82.5 ± 8.4
B2
87.8 ± 6.7
C1
87.3 ± 6.2
Appendix
Table 21: Vocabulary coverage score statistics for gemma-3-27b-it.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
45.1 ± 9.8
A2
59.7 ± 9.7
B1
72.5 ± 9.6
B2
81.2 ± 8.5
C1
81.5 ± 7.7
Appendix
Table 22: Vocabulary coverage score statistics for gpt-5.2.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
49.6 ± 10.2
A2
68.7 ± 8.9
B1
81.7 ± 7.3
B2
87.6 ± 7.2
C1
87.9 ± 6.5
Appendix
Table 23: Vocabulary coverage score statistics for gemini-2.5-flash.
Lang.
Level
Simplified
(Avg. ± Std.)
EN
A1
39.3 ± 13.0
A2
52.6 ± 12.0
B1
70.2 ± 10.1
B2
79.8 ± 8.5
C1
81.7 ± 7.5
Appendix
Table 24: Vocabulary coverage score statistics for claude-sonnet-4-5-20250929.
Lang.
Level
Voc.
Sem.
Coh.
EN
A1
57.7
78.2
76.4
A2
76.0
78.1
77.7
B1
84.6
81.8
82.1
B2
89.8
81.9
84.4
C1
89.3
81.1
87.8
C2
91.2
83.8
88.6
Appendix
Table 25: Evaluation results of Re-RIGHT on Qwen3-4B-Instruct-2507 with Wikipedia test data.
Lang.
Level
Voc.
Sem.
Coh.
EN
A1
60.3
75.8
76.4
A2
76.4
67.9
78.8
B1
85.1
71.5
82.4
B2
91.4
75.0
84.6
C1
92.2
80.1
87.8
C2
93.3
78.8
88.8
Appendix
Table 26: Evaluation results of Re-RIGHT on Qwen3-4B-Instruct-2507 with PGV test data.
Lang.
Level
Voc.
Sem.
Coh.
EN
A1
57.7
76.4
76.2
A2
74.4
78.4
77.6
B1
84.7
81.8
79.7
B2
89.2
80.7
82.3
C1
90.1
80.1
84.0
C2
92.0
81.9
84.6
Appendix
Table 27: Evaluation results of Re-RIGHT on Qwen3-4B-Instruct-2507 with Wikipedia test data.
Lang.
Level
Voc.
Sem.
Coh.
EN
A1
58.7
76.1
77.2
A2
75.6
67.6
78.0
B1
85.8
67.7
79.9
B2
91.0
74.4
82.4
C1
92.4
75.4
84.4
C2
92.9
71.2
85.0
Appendix
Table 28: Evaluation results of Re-RIGHT on Qwen3-4B-Instruct-2507 with Wikipedia test data.
Level
Simplified Text
Reference
Pigeon photography is an aerial photography technique invented in 1907 by the German apothecary Julius Neubronner, who also used pigeons to deliver medications. A homing pigeon was fitted with an aluminium breast harness to which a lightweight time-delayed miniature camera could be attached. Neubronner’s German patent application was initially rejected, but was granted in December 1908 after he produced authenticated photographs taken by his pigeons. …
CEFR A1
Pigeon photography is a way to take pictures from the air. It was made in 1907 by a man named Julius Neubronner from Germany. He also used pigeons to carry medicine. A kind of pigeon that can find its way home was given a small metal thing on its chest. On that chest, a small camera that did not weigh much and worked after a short time could be put. Neubronner wrote a plan in Germany, but first it was not accepted. In December 1908, it was allowed after he showed real pictures taken by his pigeons. …
CEFR A2
Pigeon photography is a way of taking photos from the air. It was made in 1907 by a man named Julius Neubronner from Germany. He also used pigeons to carry medicine. A kind of pigeon that can find its way home was given a small metal thing on its chest. On that chest, a small camera that did not weigh much and worked after a short time could be put. Neubronner wrote a patent in Germany, but first it was not accepted. In December 1908, it was allowed after he showed real photos taken by his pigeons. …
CEFR B1
Pigeon photography is a way of taking photos from the air that was first made in 1907 by a German man named Julius Neubronner. He also used pigeons to carry medicine. A type of pigeon that can find its way home was given a small metal piece on its chest, where a small camera that worked slowly and weighed very little could be put. Neubronner wrote a patent in Germany, but it was not accepted at first. In December 1908, it was finally allowed after he showed real photos taken by his pigeons. …
CEFR B2
Pigeon photography is a way of taking photos from the air that was first created in 1907 by a German man named Julius Neubronner. He also used pigeons to carry medicine. A type of pigeon that can find its way home was given a small metal piece on its chest, where a very light camera — which worked after a short delay — could be placed. Neubronner’s idea for a German patent was first turned down, but in December 1908, it was finally accepted after he showed real photos taken by his pigeons. …
CEFR C1
Pigeon photography is a type of aerial photography that was first created in 1907 by a German man named Julius Neubronner, who also used pigeons to carry medicine. A pigeon capable of finding its way home was given an aluminium device on its chest, where a small, light camera with a delay feature could be placed. Neubronner’s patent in Germany was first turned down, but later approved in December 1908 after he showed real photos taken by his pigeons. …
Appendix
Table 29: Re-RIGHT generation examples by proficiency levels in English. The trained policy model is Qwen3-4B-Instruct-2507.
Level
Simplified Text
Reference
… 집회와 결사의 자유, 언론의 자유 등의 자유권의 보장과 독일 의회의 수립, 그리고 헌법의 제정을 요구했다. 비록 독일에서의 혁명은 뚜렷한 족적을 남기지는 않았지만, 프리드리히가 어릴 때 목도한 이 자유주의는 훗날 그의 일생에 걸쳐 큰 영향력을 발휘하게 된다.
( Translation: … They demanded the guarantee of civil liberties such as freedom of assembly and association, freedom of the press, the establishment of a German parliament, and the enactment of a constitution. Although the revolution in Germany did not leave a distinct mark, the liberalism Friedrich witnessed in his youth would exert a profound influence throughout his life.”)
TOPIK1
… 사람들은 모임과 함께 행동할 수 있고, 언론도 자유로이 쓸 수 있도록 하는 등의 자유를 받고, 독일의 회의를 만들고, 새로운 법을 만드는 것을 요구했다. 독일에서 일어난 혁명은 특히 큰 변화를 만들지는 않았지만, 프리드리히가 어릴 때 보았던 자유주의는 그 후 그의 삶 전체에서 큰 도움이 되었다.
TOPIK2
… 사람들이 모여 이야기할 수 있고, 언론도 자유롭게 쓸 수 있도록 하는 등의 자유를 받고, 독일의 회의를 만들고, 새로운 법을 만드는 것을 요구했다. 독일에서 일어난 혁명은 특히 큰 변화를 가져오지는 않았지만, 프리드리히가 어릴 때 보았던 자유주의의 생각은 그 후 그의 삶 전체에서 큰 영향을 주게 되었다.
TOPIK3
… 회의나 함께 행동할 수 있는 자유, 언론에서 자유롭게 말할 수 있는 권리 같은 중요한 권리들을 받고, 독일의 의회를 만들고, 새로운 헌법도 만들 것을 요구했다. 독일 내에서 일어난 혁명은 특히 큰 변화를 가져오지는 않았지만, 프리드리히가 어릴 때 보았던 자유주의의 생각은 그 후 그의 삶 전체에서 매우 큰 영향을 주게 되었다.
TOPIK4
… 회의나 결사, 언론 등 자유를 보장하고, 독일의 의회를 만들고, 헌법을 만들어야 한다고 요구했다. 독일 내에서 일어난 혁명은 특히 큰 변화를 가져오지는 않았지만, 프리드리히가 어릴 때 보았던 자유주의의 모습은 그 후 그의 삶 전체에서 매우 중요한 영향을 주게 되었다.
Appendix
Table 30: Re-RIGHT generation examples by proficiency levels in a non-English language (Korean). The trained policy model is Qwen3-4B-Instruct-2507.
You are evaluating {language} text quality for a text simplification system.
Given [ORIGINAL_TEXT] and [SIMPLIFIED_TEXT], focus ONLY on how natural and fluent the [SIMPLIFIED_TEXT] reads as a rewrite of the [ORIGINAL_TEXT]. Rate the NATURALNESS of the [SIMPLIFIED_TEXT] as if it were written by a native speaker, strictly according to the following rules:
100 = indistinguishable from a native human-written well-edited text
80-99 = highly natural with only minor unnatural phrasing
60-79 = generally understandable but contains multiple awkward and unnatural expressions
Text simplification plays a crucial role in improving the accessibility and comprehensibility of written information for diverse audiences, including language learners and readers with limited literacy. Despite its importance, large-scale, high-quality datasets for training and evaluating text simplification models remain scarce for languages other than English. This paper reports an experimental study on the collection and processing of crowd-sourced simplification data from comparable corpora to construct a corpus suitable for both training and testing text simplification systems across multiple languages (Catalan, English, French, Italian and Spanish). We report mechanisms for sentence-level alignment from document-level data. The resulting dataset of the aligned sentence pairs is publicly available.
Kenji Hilasaca, Nouran Khallaf, Serge Sharoff
Centre for Translation, Localisation and Interpreting Studies School of Languages, Cultures and Societies University of Leeds, UK
Text simplification aims to make complex texts easier to understand while preserving their original meaning. Recent large language models can perform simplification through prompting, but it remains unclear whether adding explicit linguistic information about sentence complexity to the prompt improves their outputs. We investigate this question for multilingual sentence-level Easy-to-Read simplification in Catalan, Spanish, and Italian. Using Qwen3-8B, we compare a baseline prompt, a gold-cue prompt enriched with gold linguistic cues, and a predicted-cue prompt enriched with automatically predicted cues. We evaluate the outputs using SARI, BLEU, chrF, and BERTScore, and complement this evaluation with a manual qualitative analysis. Predicted-cue prompting obtains the best overall scores across all four metrics, although the gains over the baseline are small. Gold-cue prompting does not consistently improve over the baseline, and results vary across languages. These findings indicate that cue-based prompting can influence multilingual Easy-to-Read simplification, but its benefits are modest, metric-dependent, and language-dependent.
Although automatic text simplification (ATS) is critical for accessibility, its progress has not matched the rapid evolution of broader natural language processing techniques. This paper investigates the application of reinforcement learning (RL) to improve the quality of ATS for low-resource languages using Large Language Models (LLMs). The paper introduces a novel reward function, designed to guide LLMs toward a targeted simplification style with Group Relative Policy Optimization (GRPO), that combines the SARI metric with specific penalty components. The effectiveness of GRPO with this reward function is motivated and demonstrated by post-training IberianLLM-7B-Instruct on the ASSET dataset. After post-training on the English ASSET, the model's ATS performance improves on two curated Catalan benchmarks while also successfully suppressing previously observed negative behaviors. Cross-lingual transfer learning is explored by translating ASSET into Catalan and Spanish and post-training the model on each version, but these fail to show a significant improvement on the out-of-domain benchmark.
Arnau Ayguadé Domingo, Stefan Bott, Horacio Saggion
Universitat Pompeu Fabra · Barcelona Supercomputing Center