The POLAR dataset contains various social media texts that might be polarized (conflict-inducing or dangerously divisive). The task at hand is to identify whether any of the following types of polarization are present: political, racial/ethnic, religious, gender/sexual, and other types across 22 languages. In this paper, we propose a system of fine-tuned language-specific small language models and compare our approach with state-of-the-art large language models on the POLAR dataset. By fine-tuning models for each language, we demonstrate that fine-tuned small encoder-only models consistently outperform large language models, especially for low-resource languages. Our system performs well on this task for most low-resource languages, notably taking the top spot on the leaderboard in Burmese (mya), appearing within the top 10 for 12 languages, and within the top 20 for all remaining languages.

Many benchmarks show that large language models can answer direct questions about culture. We study a different question: do they also change how they speak when culture is only implied by the situation? We introduce a triad-diagnostic framework; applying it to four deployed endpoints to evaluate 60 culturally grounded conversational scenarios. We examine responses across five languages in three conditions: a neutral baseline (Prompt A), an explicit cultural instruction (Prompt B), and implicit situational cueing (Prompt C). We use an LLM-judge ensemble with explicit rubrics to score responses on 12 pragmatic features covering deference to authority, individual-versus-group framing, and uncertainty management. We define Pragmatic Context Sensitivity (PCS) as the fraction of the judge-scored Prompt A→B feature shift that reappears under Prompt A→C. Across four deployed LLM endpoints and five languages (English, German, Hindi, Nepali, and Urdu), the primary stable-only PCS mean is 0.196 (SD = 0.113), indicating that the tested models recover only about one-fifth of the explicitly prompted rubric-aligned shift. Transfer is strongest for authority-related cues (0.299) and weakest for individual-versus-group framing (0.120). We argue that multilingual cultural pragmatics is an explicit-versus-implicit deployment problem, not only a factual knowledge problem.