CEFR-oriented Russian Text Simplification Using Reinforcement Learning

Kristina Mustakova; Alexey Sorokin

cs.CLOct 19, 2026

CEFR-oriented Russian Text Simplification Using Reinforcement Learning

Authors: Kristina Mustakova, Alexey Sorokin

Organizations: MSU Center of Artificial Intelligence · Yandex

Abstract

We present a study on controlled automatic simplification of Russian texts using preference-based optimization algorithms with a focus on adapting texts to the A1, A2 and B1 proficiency levels. We fine-tune YandexGPT using Rejection Sampling and Direct Preference Optimization (DPO) methods to improve control over the complexity of the generated text. We compare several training configurations that differ in dataset composition, text type and the strategy used to construct preference pairs for DPO. For automatic evaluation we train a regression model based on ruRoBERTa-large to predict continuous complexity scores corresponding to CEFR levels. We also consider semantic similarity, lexical coverage and quantitative readability metrics. Human evaluation assesses simplicity, fluency and adequacy. Our best model achieves 87.15% target level accuracy, an MAE of 0.35 CEFR levels and a target interval MAE of 0.11. Manual assessment shows that 75% of the evaluated simplifications match the target morphological and syntactic level, and fluency and adequacy scores are especially high for A2 and B1. Although A1 remains the most challenging level, the model significantly reduces the proportion of vocabulary and grammatical constructions beyond the target level while generally preserving meaning. Our model is freely available on HuggingFace.

Figures & tables

Panel of Figure 1 — Figure 1: Continuous CEFR scale.

	MSE	RMSE	MAE	$R^{2}$	Explained variance
validation	0.39	0.62	0.45	0.84	0.87
test	0.51	0.71	0.50	0.73	0.83

Table 1: Text complexity regression metrics on the validation (at the best checkpoint) and test sets.

Dataset	A1	A2	B1	Total
v0	3424	3425	3391	10240
v1 , v2 , v3	3384	3384	3384	10152
nplus1	34207	34207	34205	102619
nplus1_2	13143	23366	29555	66064
merged	16527	23366	29555	72832
merged_2	6768	6768	6768	20304

Table 2: Size of preference datasets.

	A1		A2		B1
	$\uparrow$	$\downarrow$	$\uparrow$	$\downarrow$	$\uparrow$	$\downarrow$
v0	98.83	1.17	50.83	49.17	20.73	79.27
v1	100.0	0.00	51.77	48.23	0.00	100.0
v2	100.0	0.00	16.08	83.92	0.00	100.0
v3	100.0	0.00	51.77	48.23	30.38	69.62
nplus1	100.0	0.00	93.71	6.28	50.00	50.00

Table 3: Direction of rejected responses compared with chosen responses across target levels and preference datasets (%).

\uparrow

and

\downarrow

denote that the rejected response is more complex or simpler than the corresponding chosen response, respectively.

	test_n+1						test_ruadapt
	MAE	MAE i	total	A1	A2	B1	MAE	MAE i	total	A1	A2	B1
base	1.07	0.66	44.24	8.51	50.53	73.67	0.78	0.38	60.28	57.18	83.78	39.89
v0	1.02	0.60	44.77	16.22	47.07	71.01	0.69	0.32	65.34	62.77	85.90	47.34
v1	1.07	0.65	41.31	14.36	46.01	63.56	0.71	0.32	66.58	64.89	83.78	51.06
v2	1.07	0.66	40.34	15.16	39.10	66.76	0.71	0.32	65.69	61.97	82.71	52.39
v3	1.01	0.60	44.68	14.10	50.27	69.68	0.68	0.32	66.22	62.23	84.84	51.60

Table 4: Rejection Sampling: target level accuracy.

	test_n+1						test_ruadapt
	MAE	MAE i	total	A1	A2	B1	MAE	MAE i	total	A1	A2	B1
base	1.07	0.66	44.24	8.51	50.53	73.67	0.78	0.38	60.28	57.18	83.78	39.89
v0	1.08	0.66	43.17	8.78	46.81	73.93	0.77	0.38	59.75	56.65	83.51	39.10
v1	0.93	0.53	51.60	73.14	71.81	9.84	0.53	0.22	76.33	93.35	85.37	50.27
v2	1.14	0.69	37.23	77.39	22.07	12.23	0.55	0.25	73.14	93.88	73.40	52.13
v3	0.67	0.30	66.40	73.40	66.76	59.04	0.49	0.17	79.26	93.62	86.97	57.18

Table 5: Direct Preference Optimization: target level accuracy.

Subset	A1	A2	B1
test_ruadapt	0.82	0.90	0.93
test_n+1	0.70	0.83	0.90

Table 6: Semantic similarity across levels for the two test subsets.

Figure 2: Percentage of words outside RFL lexical minima.

Metric	Source	A1	A2	B1
Average sentence length, words	17.8	3.5	6.0	9.4
Average word length, chars	6.37	4.68	4.97	5.55
Average punctuation per sentence	3.87	1.15	1.56	2.15

Table 7: Quantitative metrics across levels.

	A1	A2	B1
test_ruadapt	0.1	0.0	-0.3
test_n+1	0.6	0.1	-0.1

Table 8: Mean simplicity scores by target level and test subset. Scores closer to 0 indicate better alignment with the expected syntactic and morphological material, while negative and positive deviations indicate complexity below and above the target level, respectively.

Figure 3: Confusion matrix for the manual simplicity evaluation.

		A1	A2	B1
Fluency	test_ruadapt	$3.08\pm 1.04$	$4.88\pm 0.25$	$4.87\pm 0.32$
Fluency	test_n+1	$3.30\pm 1.01$	$4.27\pm 0.60$	$4.92\pm 0.18$
Adequacy	test_ruadapt	$3.53\pm 0.91$	$4.87\pm 0.22$	$4.87\pm 0.22$
Adequacy	test_n+1	$3.53\pm 1.01$	$4.75\pm 0.42$	$4.93\pm 0.21$

Table 9: Human evaluation of fluency and adequacy across target levels. Scores are presented as mean

\pm

standard deviation, computed after averaging annotator scores for each text.

test_ruadapt examples
Original	Из-за дороговизны городской земли её застройка доходными домами была очень плотной и окна квартир, особенно дешёвых, часто выходили в тёмные дворы без деревьев, окружённые серыми стенами домов. (Due to the high cost of urban land, apartment houses were built very densely, and the windows of the flats, especially the cheaper ones, often faced dark, treeless courtyards surrounded by the grey walls of houses.)
B1	Из-за высоких цен на землю в городе дома строили близко друг к другу. Окна квартир, особенно недорогих, часто выходили на тёмные дворы, где нет деревьев и одни серые стены. (Due to high land prices in the city, houses were built close to each other. The windows of the flats, especially the cheaper ones, often faced dark courtyards with no trees and only grey walls.)
A2	В городе земля дорогая. Поэтому дома стоят близко друг к другу. Окна дешёвых квартир смотрят в тёмные дворы. Там нет деревьев, только серые стены. (Land is expensive in the city. That’s why the houses are close to each other. The windows of cheap flats look out onto dark courtyards. There are no trees there, only grey walls.)
A1	Город дорогой. Дома стоят близко. Окна смотрят в двор. Там нет деревьев. Там серые дома. (The city is expensive. The houses are close together. The windows look out onto the courtyard. There are no trees there. There are grey houses there.)
test_n+1 examples
Original	Всего исследователи сделали 500 фотографий. Главной проблемой при изучении снимков было их низкое качество и, как следствие, небольшое количество информации, которое могло было быть получено. (In total, the researchers took 500 photographs. The main problem in studying the images was their low quality and, as a result, the limited amount of information that could be obtained.)

Table 10: Randomly selected examples of simplifications generated by the model trained with merged_2 dataset.

Text Simplification

See all 20 papers

Right at My Level: A Unified Multilingual Framework for Proficiency-Aware Text Simplification
Apr 7, 2026Jinhong Jeong, Junghun Park, Youngjae YuSecond Language Acquisition Text Simplification
Yonsei University · Seoul National University
RALS: Resources and Baselines for Romanian Automatic Lexical Simplification
Jul 22, 2026Fabian Anghel, Petru Theodor Cristea, Claudiu Creanga +1Low-Resource Language Processing Text Simplification
Human Language Technologies Research Center Faculty of Mathematics and Computer Science University of Bucharest
SimpCue: Cue-Based Prompting for Multilingual Text Simplification
Aug 28, 2026Mehrzad Tareh, Horacio Saggion, Stefan BottText Simplification
TALN Group, Universitat Pompeu Fabra (UPF) Barcelona, Spain

Apr 7, 2026J/K move · Enter open · S save

Right at My Level: A Unified Multilingual Framework for Proficiency-Aware Text Simplification

Jinhong Jeong, Junghun Park, Youngjae Yu

Yonsei University · Seoul National University

Second Language Acquisition Text Simplification RL for Language Models Multilingual Language Models

Open paper AskPDF arXiv

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.

Abstract

Text Simplification

Explore similar work

Right at My Level: A Unified Multilingual Framework for Proficiency-Aware Text Simplification

RALS: Resources and Baselines for Romanian Automatic Lexical Simplification

SimpCue: Cue-Based Prompting for Multilingual Text Simplification