GreenMind: A Next-Generation Vietnamese Large Language Model for Structured and Logical Reasoning

2025-04-23

AI Summary

Vietnamese large language models suffer from pervasive code-mixing and factual inconsistency in structured logical reasoning—particularly in multi-step intermediate reasoning tasks. Method: We introduce GreenMind-Medium-14B-R1, built upon three key innovations: (1) the first Vietnamese synthetic Chain-of-Thought (CoT) dataset; (2) a dual reward mechanism comprising a character-level language bias detector for stylistic consistency and a Sentence-BERT–based factual consistency reward; and (3) reinforcement fine-tuning via Group Relative Policy Optimization (GRPO). Results: On the VLSP 2023 Vietnamese benchmark, GreenMind achieves significant improvements in linguistic coherence and reasoning accuracy. It also outperforms few-shot prompting baselines on the multilingual SeaExam evaluation suite, demonstrating strong cross-lingual generalization and robustness in logical reasoning. These results validate the efficacy of our dual-reward GRPO framework for enhancing both fidelity and fluency in Vietnamese LLMs.

Technology Category

• Natural Language Processing: Language Grounding & Multi-modal NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Language and Vision

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Chain-of-Thought (CoT) is a robust approach for tackling LLM tasks that require intermediate reasoning steps prior to generating a final answer. In this paper, we present GreenMind-Medium-14B-R1, the Vietnamese reasoning model inspired by the finetuning strategy based on Group Relative Policy Optimization. We also leverage a high-quality Vietnamese synthesized reasoning dataset and design two reward functions to tackle the main limitations of this technique: (i) language mixing, where we explicitly detect the presence of biased language characters during the process of sampling tokens, and (ii) we leverage Sentence Transformer-based models to ensure that the generated reasoning content maintains factual correctness and does not distort the final output. Experimental results on the Vietnamese dataset from the VLSP 2023 Challenge demonstrate that our model outperforms prior works and enhances linguistic consistency in its responses. Furthermore, we extend our evaluation to SeaExam-a multilingual multiple-choice dataset, showing the effectiveness of our reasoning method compared to few-shot prompting techniques.

Problem

• Develop Vietnamese LLM for structured reasoning tasks
• Address language mixing bias in token sampling
• Ensure factual correctness in generated reasoning content

Innovation

• Group Relative Policy Optimization finetuning strategy
• Vietnamese synthesized reasoning dataset utilization
• Sentence Transformer-based factual correctness enforcement