Large Language Models (LLMs) are susceptible to adversarial attacks such as jailbreaking, which can elicit harmful or unsafe behaviors. This vulnerability is exacerbated in multilingual settings, where multilingual safety-aligned data is often limited. Thus, developing a guardrail capable of detecting and filtering unsafe content across diverse languages is critical for deploying LLMs in real-world applications. In this work, we introduce a multilingual guardrail with reasoning for prompt classification. Our method consists of: (1) synthetic multilingual data generation incorporating culturally and linguistically nuanced variants, (2) supervised fine-tuning, and (3) a curriculum-based Group Relative Policy Optimization (GRPO) framework that further improves performance. Experimental results demonstrate that our multilingual guardrail, MrGuard, consistently outperforms recent baselines across both in-domain and out-of-domain languages by more than 15%. We also evaluate MrGuard’s robustness to multilingual variations, such as code-switching and low-resource language distractors in the prompt, and demonstrate that it preserves safety judgments under these challenging conditions. The multilingual reasoning capability of our guardrail enables it to generate explanations, which are particularly useful for understanding language-specific risks and ambiguities in multilingual content moderation.

Recently it has been shown that state-of-the-art NLP models are vulnerable to adversarial attacks, where the predictions of a model can be drastically altered by slight modifications to the input (such as synonym substitutions). While several defense techniques have been proposed, and adapted, to the discrete nature of text adversarial attacks, the benefits of general-purpose regularization methods such as label smoothing for language models, have not been studied. In this paper, we study the adversarial robustness provided by label smoothing strategies in foundational models for diverse NLP tasks in both in-domain and out-of-domain settings. Our experiments show that label smoothing significantly improves adversarial robustness in pre-trained models like BERT, against various popular attacks. We also analyze the relationship between prediction confidence and robustness, showing that label smoothing reduces over-confident errors on adversarial examples.

BabyBERTa, a language model trained on small-scale child-directed speech while none of the words are unmasked during training, has been shown to achieve a level of grammaticality comparable to that of RoBERTa-base, which is trained on 6,000 times more words and 15 times more parameters. Relying on this promising result, we explore in this paper the performance of BabyBERTa-based models in downstream tasks, focusing on Semantic Role Labeling (SRL) and two Extractive Question Answering tasks, with the aim of building more efficient systems that rely on less data and smaller models. We investigate the influence of these models both alone and as a starting point to larger pre-trained models, separately examining the contribution of the pre-training data, the vocabulary, and the masking policy on the downstream task performance. Our results show that BabyBERTa trained with unmasking-removal policy is a much stronger starting point for downstream tasks compared to the use of RoBERTa masking policy when 10M words are used for training and that this tendency persists, although to a lesser extent, when adding more training data.