CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards
Authors: Wei Tian, Yuhao Zhou, Man Lan
Organizations: School of Computer Science and Technology, East China Normal University · Shanghai Institute of Artificial Intelligence for Education, East China Normal University
Large Language Model (LLM) based Chinese Grammatical Error Correction (CGEC) systems face two critical challenges: general-purpose models lack specialized linguistic priors for subtle grammatical distinctions, and Supervised Fine-Tuning (SFT) with Maximum Likelihood Estimation fails to optimize for precision-focused metrics, leading to systematic over-correction. We propose CSRP, a three-stage framework that progressively builds correction capability through Continual Pre-training (CPT) on 5.9M balanced samples to internalize domain knowledge, Chain-of-Thought SFT with explicit error reasoning for diagnostic transparency, and Group Relative Policy Optimization with a novel Efficiency-Aware Reward that explicitly penalizes unnecessary edits. On the NACGEC benchmark, CSRP achieves state-of-the-art performance with 50.99 F0.5 and 57.17 precision, substantially outperforming previous best results while effectively mitigating the over-correction bias inherent in MLE-trained models. Our method also advances CSCD spelling correction to 59.61 F1, surpassing GPT-4 by 5.20 points. Comprehensive ablation studies demonstrate that the RL alignment stage contributes a 8% relative gain over the SFT baseline, and that this gain is orthogonal to the contribution of large-scale CPT, validating that explicit optimization for edit efficiency is essential for high-quality grammatical error correction. Our code is available at https://github.com/TW-NLP/ChineseErrorCorrector.
Figure 2: Overview of the proposed CSRP (CPT-SFT-RL) framework.
Figure 3: CPT Data Processing Process.
Figure 4: Cot Processing Process.
Phase
Source
Count
CPT
Wiki-zh, CCI2
4.7M
Lang8, HSK
1.2M
SFT
CSCD-train, NACGEC
37.5K
FCGEC, CGED
55K
WK271
251K
(80% split)
269K
Table 1: Data statistics for the three-stage training curriculum and evaluation sets. The SFT and RL phases use an 80:20 split of the correction data (total 336K samples after filtering). Note that the actual sample count (336K) is lower than the sum of individual subsets, as a portion of sentences containing sensitive content were rejected by the cloud-based teacher model’s safety policy during CoT distillation.
Model
P
R
F0.5
BART Lewis et al. (2020)
34.67
41.88
35.91
HW-CGEC Su et al. (2023)
50.95
32.29
45.26
ScholarGEC (14B) Kong et al. (2025)
45.08
59.33
47.35
CEC3 (4B) Tian and others (2025)
54.20
34.75
48.74
CSRP (4B)
57.17
35.60
50.99
Table 2: Main results on the NACGEC benchmark.
Model
F1
BERT Tan et al. (2020)
25.49
SoftMask Zhang et al. (2020)
44.48
SMBERT Li et al. (2021)
44.67
MDCSpell+ARM Liu et al. (2024)
48.93
PGT (BERT) Wei et al. (2024)
48.57
GPT-4 Achiam et al. (2023)
54.41
Table 3: Chinese spelling check performance on the CSCD dataset.
Configuration
NACGEC
NACGEC
CSCD
P
R
F0.5
F1
SFT only (merged data)
42.13
34.02
40.21
49.71
SFT + GRPO (w/o CPT)
50.54
33.75
45.97
52.96
CPT + SFT (no CoT)
44.90
35.50
42.64
52.01
CPT + SFT
48.73
35.80
45.45
56.28
CPT + SFT (w/ RL data)
52.20
36.00
47.21
57.92
Table 4: Progressive ablation study. “SFT only (merged data)” combines all data without CPT. “SFT + GRPO (w/o CPT)” applies GRPO directly to the base model to isolate the RL contribution. “CPT + SFT (no CoT)” excludes chain-of-thought reasoning. “CPT + SFT (w/ RL data)” uses all data for SFT as a controlled baseline.
NACGEC
Configuration
P
R
F0.5
CPT + SFT
48.73
35.80
45.45
CPT + SFT + GRPO
57.17
35.60
50.99
Change
+8.44
-0.20
+5.54
Table 5: Impact of GRPO on precision-recall trade-off. The RL stage significantly improves precision while maintaining comparable recall, demonstrating effective learning of conservative editing behavior.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Train
Test
Type
CSCD-NS
30,000
5,000
CSC
WK271+SIGHAN15
251,835
1,100
CSC
CGED
20,449
-
CGEC
NACGEC
7,568
5,869
CGEC
FCGEC
35,355
-
CGEC
Subtotal
345,207
11,969
-
Appendix
Table 6: Data composition for Stage II (SFT) and Stage III (RL).
Data Source
Raw Count
After Dedup & Filter
Wiki-zh-25
1,442,291
1,417,899
Wiki-zh-23
254,547
0
CCI2
4,021,572
3,306,931
Lang8+HSK
1,568,885
1,176,870
Total
7,287,295
5,901,700
General:Correction = 8:2 (4,707,480:1,176,870)
Appendix
Table 7: Data processing statistics for the CPT phase.
Metric
Score
Semantic Agreement
95.2%
Average Pairwise κ
0.81
Appendix
Table 8: Human evaluation metrics on 1,000 sampled CoT rationales. Semantic Agreement denotes the proportion of samples reaching a majority consensus ( ≥ 2/3 labeled as Faithful). The Average Pairwise Cohen’s κ of 0.81 indicates almost perfect inter-annotator agreement.
Grammatical error correction using large language models often suffers from the over-correction issue. To mitigate this, we propose a training-free inference method that performs edit-level majority voting over multiple candidates generated by a single model, without requiring model modifications or additional training. Across nine benchmarks covering English, Czech, German, Ukrainian, Korean, Hindi, and Romanian, the proposed method outperforms both greedy and MBR decoding in most cases. Moreover, it yields stable correction quality regardless of the instruction prompts used. We release two repository supporting GEC datasets loading and LLM inference.
Chinese Semantic Error Correction (CSEC) targets semantic errors in Chinese text, which are typically more subtle and complex than spelling and grammatical errors but remain relatively underexplored. Existing LLM-based approaches face two recurring obstacles in this task: over-correction, and unclear interaction between Chain-of-Thought (CoT) reasoning and self-consistency decoding, such that the benefits brought by CoT cannot be reliably transferred to final corrections. We propose Vote-guided Advantage Allocation for CSEC (VAA-CSEC), a multi-stage framework that combines CoT distillation, Supervised Fine-Tuning (SFT), Reinforcement Learning (RL) and self-consistency decoding. During RL, we design a task-specific reward function that directly aligned with the minimal-editing principle of CSEC. We further introduce Group-Level Relative Policy Optimization (GLPO), which reallocates GRPO advantages according to the margin between individual rollout rewards and the vote-aggregated group reward, aligning the RL training objective with the self-consistency objective used at inference time. Experiments on CSED-C and NaSGEC-Exam show that VAA-CSEC outperforms all LLM-based baselines on CSED-C with an F0.5 of 47.72%, achieves the highest recall of 42.15% among all methods, and establishes a new state of the art of 41.55% F0.5 on NaSGEC-Exam.
Yitong Han, Nankai Lin, Juan Luo +3
School of Information Science and Technology, Guangdong University of Foreign Studies · Guangdong Engineering Research Center of Data Security Governance and Privacy Computing · College of Computer Science and Technology, National University of Defense Technology
Automated assistants for Grammatical Error Correction are now embedded in educational platforms serving millions of learners, yet three critical gaps remain in this domain: (1) latest-generation Large Language Models (LLMs) lack comprehensive evaluation on grammar correction tasks; (2) whether combining these LLMs improves correction quality is unexplored; and (3) the extent to which reference-based metrics underestimate GEC system performance has not been adequately quantified. In this study, first, we evaluate latest-generation LLMs on edit precision, fluency preservation, and meaning retention, showing fine-tuned GPT-4o achieves state-of-the-art performance across all three dimensions. Second, through grammatical error type analysis we demonstrate that individual LLMs exhibit highly similar error correction patterns (ρ=0.947). Third, we show that reference-based metrics underestimate GEC performance with 73.76% of GPT-4o corrections different from gold standards being equally valid or even superior. These GEC evaluation findings equip educators with guidance for selecting GEC assistants that enhance rather than constrain student linguistic development. We make our data, code, and models publicly available.
Adnan Labib, Qiao Wang, Yixuan Huang +1
King’s College London, London, United Kingdom · Hosei University, Tokyo, Japan · Waseda University, Tokyo, Japan +1