CSRP framework beats GPT-4 on Chinese text correction via RL rewards

Chinese Grammatical Error Correction (CGEC) has long struggled with over-correction and a lack of specialized linguistic knowledge. General-purpose LLMs often flag correct sentences as errors, while traditional Supervised Fine-Tuning (SFT) optimizes for likelihood rather than precision. To address this, Wei Tian, Yuhao Zhou, and Man Lan introduce CSRP: a three-stage framework that progressively builds correction expertise. First, Continual Pre-Training (CPT) on 5.9M balanced samples internalizes domain-specific linguistic priors. Second, Chain-of-Thought SFT teaches the model to explain each error before correcting it, adding transparency. Finally, Group Relative Policy Optimization (GRPO) is applied with a novel Efficiency-Aware Reward function that explicitly penalizes unnecessary edits, reducing over-correction.

The results are striking. On the NACGEC benchmark, CSRP achieves 50.99 F0.5 and 57.17 precision—substantially outperforming all previous methods. On the CSCD spelling correction dataset, it reaches 59.61 F1, surpassing GPT‑4 by a significant 5.20 points. Ablation studies reveal that the RL alignment stage contributes an 8% relative gain over the SFT baseline, and this improvement is orthogonal to the gains from large-scale CPT. This confirms that explicit optimization for edit efficiency is critical for high-quality grammatical correction. The work has been accepted at ACL 2026, and the code is open-sourced.