Researchers from DS@GT ARC have developed a system for the CheckThat! 2026 competition focused on verifying numerical claims in English and Arabic. Their approach involves ranking reasoning traces generated by large language models (LLMs) and using grouped reward modeling. They explored two methods: fine-tuning an LLM with LoRA for trace scoring and using a lightweight TF-IDF reward model. Results indicated that the LLM-based approach generally outperformed the TF-IDF model, particularly in recall, while the latter showed strength in identifying conflicting claims. For Arabic, a language-specific model, AraBERT, proved more effective than a general multilingual model. AI
IMPACT This research advances methods for LLM-based claim verification, potentially improving the accuracy and reliability of automated fact-checking systems.
RANK_REASON The cluster contains a research paper detailing a system for numerical claim verification using LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
