SEA-Guard: Culturally Grounded Multilingual Safeguard for Southeast Asia

Culturally aware safeguards are crucial for AI alignment in real-world settings, where safety extends beyond common sense and encompasses diverse local values, norms, and region-specific regulations. However, building large-scale, culturally grounded datasets is challenging due to limited resources and a scarcity of native annotators. Consequently, many safeguard models rely on machine translation of English datasets, often missing regional and cultural nuances. We present a novel agentic data-generation framework to scalably create authentic, region-specific safety datasets for Southeast Asia (SEA). On this foundation, we introduce the SEA-Guard family, the first multilingual safeguard models grounded in SEA cultural contexts. Evaluated across multiple benchmarks and cultural variants, SEA-Guard consistently outperforms existing safeguards at detecting regionally sensitive or harmful content while maintaining strong general safety performance.

💡 Research Summary

The paper introduces SEA‑Guard, the first multilingual safety‑guard model explicitly grounded in Southeast Asian (SEA) cultural contexts. Recognizing that existing safeguards are largely built on English data and machine‑translated corpora, the authors argue that such approaches miss region‑specific norms, traditions, and regulatory nuances, leading to poor performance on culturally sensitive content. To address this gap, they devise an agentic data‑generation pipeline that automatically creates a large, culturally nuanced safety dataset covering eight SEA languages (Burmese, English, Tagalog, Indonesian, Malay, Tamil, Thai, Vietnamese) and 53 cultural categories (food, festivals, politics, etc.).

The pipeline consists of four stages. First, a requirement formulation step defines four metadata dimensions—cultural topic, country, prompt type, and label type—and prioritizes under‑represented combinations to ensure balanced coverage. A “guideline agent” then produces detailed, step‑by‑step generation instructions (including sensitivity levels, prohibited actions, and validation criteria). Second, prompts and responses are generated by combining these guidelines with persona information (age, gender, residence) and the target language, allowing the system to capture subtle intra‑regional differences (e.g., the timing of Songkran in Thailand vs. Myanmar). Prompt paraphrasing mitigates keyword bias, and four strong LLMs (Gemma‑SEA‑LION‑v4‑27B, Llama3.1‑70B‑IT, Gemma‑3‑27B‑IT, GPT‑OSS‑20B‑IT) produce diverse responses.

Third, the authors introduce Monte Carlo Reasoning Ensemble (MCRE) for labeling and verification. For each input, N = 10 stochastic chain‑of‑thought reasoning passes are performed, yielding a distribution over five ordinal safety classes (Safe, Safe‑Sensitive, Sensitive, Sensitive‑Harmful, Harmful). By mapping these to a continuous harmfulness score and applying fixed thresholds, a final three‑way label (Safe, Sensitive, Harmful) is derived. This ensemble approach captures uncertainty and reduces over‑confidence that single‑pass zero‑shot annotators exhibit on culturally nuanced cases. Additional zero‑shot classifiers (culture, topic, usage) built with MCRE further validate alignment with the original requirements. A bag‑of‑words shortcut detector removes near‑duplicate, superficial samples, trimming the dataset from 1 M to 870 k per language. Human verification by 32 native annotators confirms that roughly 80 % of the data are high‑quality.

Using the curated dataset, the authors fine‑tune three model sizes—SEA‑Guard‑4B, ‑8B, and ‑12B—based on Qwen‑SEA‑LION‑v4‑VL and Gemma‑3, which have demonstrated strong SEA‑HELM performance on language and cultural understanding.

Evaluation spans three axes. (1) A SEA‑specific safety benchmark assesses multilingual consistency (RQ1) and cultural knowledge (RQ2); SEA‑Guard outperforms prior safeguards (ShieldGemma, LlamaGuard, PolyGuard, etc.) by a sizable margin. (2) A generic multilingual safety benchmark tests generalization (RQ3); despite not being trained on generic safety data, SEA‑Guard remains competitive. (3) Zero‑shot vision‑language safety tasks probe domain transfer; SEA‑Guard improves baseline performance in six of seven domains, especially on images depicting regional customs. Additional analyses show robustness to under‑ and over‑defensiveness and resistance to adversarial prompt attacks.

The contributions are: (i) a scalable, agentic framework for culturally grounded safety data creation, (ii) the MCRE method for robust, uncertainty‑aware labeling, (iii) the SEA‑Guard family of multilingual safeguards, and (iv) extensive empirical validation across text and vision modalities. Limitations include reliance on the biases of the LLMs used for generation, the computational cost of MCRE (making it unsuitable for real‑time labeling), and the focus on eight languages, leaving many SEA languages unaddressed. Future work should explore lightweight inference‑time labeling, expand coverage to additional low‑resource languages, and integrate human‑in‑the‑loop feedback to further mitigate systematic biases.