via Selective Multilingual and Multitask Specialization UMUSP at SemEval-2026 Task 9: Mitigating Cross-Lingual Interference via Selective Multilingual and Multitask Specialization This paper proposes a selective multilingual and multitask fine-tuning strategy for online po- larization detection that improves cross-lingual stability over fully joint training. Covering all three subtasks — polarization detection (POLARDETECT), polarization type classifi- cation (POLARTYPE), and rhetorical mani- festation identification (POLARMANIFEST) — across all 22 languages of the shared task, the approach introduces controlled specialization, where languages and subtasks are grouped em- pirically and separate specialist models are fine-tuned for each subset. Restricting parame- ter sharing substantially improves performance even without ensemble averaging, whereas en- sembling jointly trained models fails to mit- igate instability. The final specialist ensem- ble improves Task 3 macro-F1 from 0.3330 to 0.4920 and reduces cross-lingual dispersion (CV: 0.613 → 0.321). Under the official rank- ing framework, the system ranks 7th among 16 submissions with complete multilingual and multitask coverage and remains within 5% of the best system in 37.70% of evaluation condi- tions.