Reasoning-oriented language models inherit strong English-centric biases from pre-training, which are further reinforced by English-only post-training. This creates a dual failure mode where models fail both to reason correctly and to maintain reasoning traces in the query language. We study reasoning accuracy and language consistency under English-only post-training. SFT primarily improves language consistency with limited or inconsistent accuracy gains, while group relative policy optimization (GRPO) substantially improves accuracy at the cost of language consistency. Although translated multilingual GRPO improves both objectives, it requires translating reasoning post-training data into multiple languages and may introduce translation noise and information loss. We find that the language-consistency capability acquired by SFT and the accuracy capability acquired by GRPO are approximately orthogonal in the model’s update space. By merging an English LoRA-SFT adapter with an English GRPO checkpoint trained from the same base model, M-GOAL improves MGSM accuracy from 39.1 to 57.8 and language consistency from 17.1 to 75.2 without translated reasoning supervision. This composition outperforms sequential training, its merge coefficient exposes a tunable operating point along the accuracy–consistency frontier, and the pattern holds beyond mathematical reasoning and on benchmarks whose non-English problems are natively authored rather than translated. Our results suggest that multilingual reasoning post-training does not necessarily require multilingual post-training data.

Language models have displayed a wide array of capabilities, but the reason for their performance remains a topic of heated debate and investigation. Do these models simply recite the observed training data, or are they able to abstract away surface statistics and learn the underlying processes from which the data was generated? To investigate this question, we explore the capabilities of a GPT model in the context of Markov Decision Processes (MDPs), where the underlying transition dynamics and policies are not directly observed. The model is trained to predict the next state or action without any initial knowledge of the MDPs or the players’ policies. Despite this, we present evidence that the model develops emergent representations of the underlying parameters governing the MDPs.

In Task 9, we are required to analyze the textual intimacy of tweets in 10 languages. We fine-tune XLM-RoBERTa (XLM-R) pre-trained model to adapt to this multilingual regression task. After tentative experiments, severe class imbalance is observed in the official released dataset, which may compromise the convergence and weaken the model effect. To tackle such challenge, we take measures in two aspects. On the one hand, we implement data augmentation through machine translation to enlarge the scale of classes with fewer samples. On the other hand, we introduce focal mean square error (MSE) loss to emphasize the contributions of hard samples to total loss, thus further mitigating the impact of class imbalance on model effect. Extensive experiments demonstrate remarkable effectiveness of our strategies, and our model achieves high performance on the Pearson’s correlation coefficient (CC) almost above 0.85 on validation dataset.

Sexism is a growing online problem. It harms women who are targeted and makes online spaces inaccessible and unwelcoming. In this paper, we present our approach for Task A of SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS), which aims to perform binary sexism detection on textual content. To solve this task, we fine-tune the pre-trained model based on several popular natural language processing methods to improve the generalization ability in the face of different data. According to the experimental results, the effective combination of multiple methods enables our approach to achieve excellent performance gains.

This paper describes our proposed framework for the 10 text classification tasks of Task 1a, 2a, 2b, 3a, 4, 5, 6, 7, 8, and 9, in the Social Media Mining for Health (SMM4H) 2022. According to the pre-trained BERT-based models, various techniques, including regularized dropout, focal loss, exponential moving average, 5-fold cross-validation, ensemble prediction, and pseudo-labeling, are applied for further formulating and improving the generalization performance of our framework. In the evaluation, the proposed framework achieves the 1st place in Task 3a with a 7% higher F1-score than the median, and obtains a 4% higher averaged F1-score than the median in all participating tasks except Task 1a.