Multilingual Fact-Checking at Scale: Fine-Tuned Compact Models vs LLMs
About
We present a multilingual fact-checking system deployed at Factiverse, designed for high-throughput and low-latency operation across diverse languages. The system follows a modular pipeline with three stages: claim detection, evidence retrieval and re-ranking, and veracity prediction. We fine-tune XLM-RoBERTa-Large for claim detection, mmBERT-base for three-label stance classification (Supports/Refutes/Mixed), and a SetFit-based multilingual re-ranker for claim--evidence matching. We compare these components against strong LLM baselines, including GPT-5.2, Claude Opus~4.6, and Qwen3-8b. Experiments on production data spanning 114 languages for claim detection and 28 languages for veracity prediction show that task-specific fine-tuning provides strong and stable multilingual performance, while the fine-tuned retrieval model remains competitive with modern proprietary embeddings. Same-hardware latency measurements further show large efficiency gains for encoder-based components, supporting their use in production deployments with tight cost and privacy constraints. Overall, compact fine-tuned, self-hosted models remain a practical and effective foundation for multilingual fact-checking at scale. Code and data used for this study are available at https://github.com/factiverse/factcheck-editor.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Evidence Retrieval | Factiverse Evaluation Dataset | F1 Score82.16 | 7 | |
| Re-ranking | Multilingual Fact-Checking (test) | Median Latency (ms)4.1 | 6 | |
| Claim Detection | Multilingual Fact-Checking (114 languages) (test) | Median Latency (ms)1.3 | 4 | |
| Stance Detection | Multilingual Fact-Checking 28 languages (test) | Median Latency (ms)3.7 | 4 | |
| Claim Detection | Factiverse Multilingual Evaluation Set Aggregate Mean (test) | Macro F172.77 | 4 | |
| Veracity Prediction | Factiverse Multilingual Evaluation Set Aggregate Mean (test) | Macro F162.05 | 4 |