Unlocking LLM Safeguards for Low-Resource Languages via Reasoning and Alignment with Minimal Training Data

TL;DR

This work tackles the problem of safeguarding LLMs in low-resource languages, where classifier-based methods struggle with interpretability and cross-lingual gaps. It introduces ConsistentGuard, a three-stage framework combining SFT-based cold-start distillation, GRPO-based reasoning training with length and diversity rewards, and CAO cross-lingual alignment to harmonize multilingual reasoning. Key contributions include a novel reasoning-based multilingual safeguard, a cross-lingual alignment algorithm (CAO) to reduce language-induced gaps, and a multilingual benchmark extension evaluated on six languages with only 1,000 seed samples, outperforming larger models trained on far more data. The approach yields strong generalization and interpretability, and the authors release code and extended benchmarks to support future research in multilingual safety.

Abstract

Recent advances in LLMs have enhanced AI capabilities, but also increased the risk posed by malicious requests, highlighting the need for effective LLM safeguards to detect such queries. Existing approaches largely rely on classifier-based methods that lack interpretability and perform poorly on low-resource languages. To address these limitations, we propose ConsistentGuard, a novel reasoning-based multilingual safeguard, which enhances explainability via reasoning and boosts knowledge transfer between languages through alignment. With only 1,000 training samples, our method demonstrates superior performance on three datasets across six languages, outperforming larger models trained with significantly more data, and exhibits strong interpretability and generalization ability. We also contribute a multilingual benchmark extension and release our codes to support future research.

Figures (6)

• Figure 1: The general framework of the proposed ConsistentGuard. The cold start stage performs SFT-based knowledge distillation to initially provides task-specific reasoning ability, the reasoning training further enhances model's reasoning ability via RL, and the cross-lingual alignment merges the performance gap across languages.
• Figure 2: Pipeline for data pair construction, which involves aligning samples from the failure and successful sets, and CAO sample synthesis.
• Figure 3: Performances across ablation models. None of the models have undergone cross-lingual alignment.
• Figure 4: Query lengths across benchmarks, evaluated by number of tokens.
• Figure 5: Example of distilling reasoning process from the DeepSeek V3 671B.