FinGround Detects Financial Hallucinations in LLMs via Atomic Claim Verification

Researchers have introduced FinGround, a three-phase approach designed to identify and validate financial inaccuracies in large language models. As the EU AI Act's deadline for high-risk enforcement approaches in August 2026, financial AI systems frequently generate false metrics, fabricate citations, and miscalculate derived figures, leading to potential regulatory issues. Current hallucination detection methods fail to distinguish between claims, resulting in a 43% oversight of computational mistakes that need arithmetic verification against structured data. In Stage 1, FinGround conducts finance-aware hybrid retrieval of text and tables. Stage 2 breaks down responses into individual claims categorized by a six-type financial taxonomy, verified through type-specific strategies, including formula reconstruction. Finally, Stage 3 revises unsupported claims with citations at the paragraph and table-cell levels, aiming to separate verification value from retrieval quality.