LLM-based fact-checking¶
Large language models have emerged as a powerful approach for automated claim verification, replacing or augmenting traditional supervised learning pipelines. LLMs can: (1) decompose complex claims into verifiable sub-claims via structured prompting; (2) retrieve and rank evidence using retrieval-augmented generation (RAG); (3) perform natural language inference to determine whether evidence entails, contradicts, or is neutral to a claim; (4) explain their reasoning in natural language; (5) handle evidence conflicts via factual consensus scoring with quality constraints.
Key capabilities include structured output generation (via constrained decoding), chain-of-thought reasoning (improving accuracy by 10–20% over single-step classification), and integration with external knowledge bases via retrieval-augmented prompts. Challenges include hallucination, handling conflicting evidence, biases from training data, and computational cost at scale.
Key papers¶
- Zhang et al. (2026) — CoVer: Conflict-Aware Claim Verification — Uses LLMs with structured constraints for evidence quality scoring and factual adjudication; achieves 86–88.5% accuracy on conflict resolution tasks, demonstrating effectiveness of LLM-based consensus scoring for handling contradictory social media evidence
- Quelle & Bovet (2023) — The Perils & Promises of Fact-checking with Large Language Models — Comprehensive evaluation of GPT-3.5/4 for claim verification; introduces ReAct agent framework combining LLM reasoning with iterative retrieval; shows 10–20pp accuracy gains from context retrieval but identifies training-data biases limiting non-English performance
Related topics¶
- Claim Verification — the core fact-checking task
- Fact-checking and corrections — manual and automated verification approaches
- Natural Language Inference — reasoning over evidence to classify veracity
- Evidence Based Fact Verification — grounding verdicts in retrieved evidence