What is Factiverse's new production-grade fact-checking system?

It is a modular multilingual pipeline with three stages: claim detection, evidence retrieval with re-ranking, and veracity prediction. The system uses fine-tuned XLM-RoBERTa-Large, mmBERT-base, and a SetFit-based multilingual re-ranker instead of relying on large language models.

Why do compact fine-tuned models outperform LLMs in this domain?

Across 114 languages, task-specific fine-tuned encoders delivered stronger and more consistent performance than GPT-5.2, Claude Opus 4.6, and Qwen3-8b, while offering significantly lower latency under the same hardware constraints.

What are the broader implications and next steps for this approach?

As model compression and efficient fine-tuning advance, this compact self-hosted architecture could expand to legal document review, medical information verification, and other high-stakes verticals requiring precision and data sovereignty.

多語言事實查核規模化：微調緊湊模型與大型語言模型的對比研究

本文介紹了Factiverse部署的一套面向高吞吐、低延遲的多語言事實查核系統。該系統採用模組化流水線，包含主張檢測、證據檢索與重排序、以及真實性預測三個階段。研究團隊針對特定任務微調了XLM-RoBERTa-Large用於主張檢測，mmBERT-base用於三分類立場判定（支持/反駁/混合），並基於SetFit構建了多語言重排序器以優化主張與證據的匹配。透過與GPT-5.2、Claude Opus 4.6及Qwen3-8b等強大語言模型基線進行對比，實驗涵蓋114種語言的主張檢測及28種語言的真實性預測。結果顯示，任務特定的微調模型在多語言場景下表現出更強且更穩定的效能，且在同等硬體條件下，基於編碼器的元件在延遲和效率上具有顯著優勢。這表明，在成本敏感及隱私約束嚴格的生產環境中，緊湊型的自託管微調模型依然是構建模組化多語言事實查核系統的務實且高效的基礎方案。