Automated Essay Scoring (AES) in low-resource, non-English educational settings faces a compounded challenge: existing corpora record only a final holistic score, discarding the rich intermediate reasoning that human examiners rely on. We argue this annotation gap may contribute to Rubric Drift—the observed tendency of LLMs to apply scoring norms misaligned with the target curriculum, potentially reflecting Anglocentric defaults. To address this, we introduce MPEAB-2430, a multilingual AES benchmark spanning Malay, Hindi, and Arabic curricula, annotated with Pedagogical Chain-of-Thought (P-CoT) trajectories that capture sub-dimensional analyses, explicit error attribution, and curriculum-conformant feedback, derived entirely from independent, language-specific expert practice. Leveraging this rich but modest-scale supervision, we propose RL-CRAF++, a sample-efficient pipeline combining CoT-SFT with Constrained GRPO. Crucially, our RL formulation relies entirely on human-annotation-derived rewards, completely bypassing the potential biases of external LLM judges. Evaluated independently per language, a 4B parameter RL-CRAF++ model consistently surpasses zero-shot frontier models (including GPT-5 and DeepSeek-V3.2) on in-domain test sets. Ablation analysis reveals that P-CoT supervision accounts for approximately 78.1% of the alignment gain over standard SFT, providing strong evidence that, in this culturally situated evaluation setting, process-level structural supervision contributes substantially more than RL-stage calibration alone. Our code is available at https://github.com/TW-NLP/MPEAB

Chinese elementary-school writing is rubric-driven. We study it as a teacher-facing instructional support task: given a classroom prompt and requirements, generate a teacher-reviewable example matching the target genre, rubric, length, and age-appropriate register. Directly prompted frontier LLMs often mis-route genres or adopt an adult register.

We present **RA-MOPD**, a four-stage pipeline that post-trains a 4B model into a unified essay writer across four genres and eight rubric dimensions. Its core stage augments on-policy reverse-KL distillation with *Rubric-Weighted KL*, which up-weights rubric-critical tokens via reward-model attribution, and *Cross-Genre Mixing*, a parameter-free expert blend guided by an empirical transfer prior.

On a new 6,000-essay benchmark with eight reward models (QWK = 0.864), RA-MOPD reaches 4.21/5.00 and 97.8% routing—within 0.09 of GPT-5.3 3-shot (p < 0.01). No statistically significant difference is detected relative to a 4×-cost four-expert logit-averaging baseline (Δ = +0.03, p = 0.18).

A GPT-4.1 judge, BERTScore, a 1,200-essay teacher study (62.8% preference), and register analysis provide converging evidence beyond RM metrics. GPT-5.5 holistic evaluation further supports the component-level gains, which remain stable on Qwen3-1.7B without model-specific retuning.

Code and data: [https://github.com/TW-NLP/MOPD](https://github.com/TW-NLP/MOPD)

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 F_0.5 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.

With the scale of Large Language Models(LLMs) and the size of the training data continuing to expand, the computational costs required for training or tuning have significantly increased as well. In this work we propose an efficient and effective Large-Scale Data Compression (LSDC) method to substantially reduce the size of training data and thus enhance the training efficiency without compromising the performance of LLMs through a bifurcated quantization strategy. Specifically, our method first segments the dataset into multiple clusters, significantly reducing the time and memory requirements for data compression. Then, during the second phase of coreset selection, the diversity of samples is ensured by maximizing the submodular gain in order to avoid performance degradation. The comparative experiments showed that the performance of LLMs fine-tuned on a 20% compressed subset of the Alpaca dataset using LSDC outperformed those on the full dataset. Moreover,on a domain-specific instruction dataset of millions of samples, the LLMs fine-tuned on a 10% compressed dataset using LSDC outperformed those on the entire dataset, which dramatically enhances the domain-adaption capabilities of LLMs. This provides a promising potential of LSDC in training bigger LLMs from scratch and supervised fine-tuning as well.