💻 computer science

FinGround: Detecting and Grounding Financial Hallucinations via Atomic Claim Verification

FinGround is a three-stage "verify-then-ground" pipeline designed to reduce financial hallucinations in LLMs by decomposing answers into atomic claims, applying specialized verification strategies (including arithmetic re-calculation), and providing precise citations, ultimately achieving a 78% reduction in errors compared to GPT-4o.

Original authors: Dongxin Guo, Jikun Wu, Siu Ming Yiu

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Dongxin Guo, Jikun Wu, Siu Ming Yiu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a high-powered financial assistant to read through thousands of pages of complex company reports. This assistant is incredibly fast and sounds very confident, but they have a dangerous habit: they occasionally make things up.

They might tell you, "The company’s profit grew by 15%," when it actually grew by 5%. Or they might say, "According to Rule 10b-5..." when that rule doesn't even exist. In the world of finance, a "confident lie" like this isn't just a mistake; it can lead to massive lawsuits, regulatory fines, and ruined careers.

This paper introduces FinGround, a specialized "Truth-Checker" system designed to sit behind the AI and make sure it never tells a lie.

Here is how it works, explained through three simple metaphors:

1. The "Atomic" Microscope (Breaking it down)

Most AI detectors look at a whole paragraph and ask, "Does this sound true?" That’s like a teacher glancing at an essay and saying, "This looks okay." It’s too vague.

FinGround works like a scientist with a high-powered microscope. Instead of looking at the whole paragraph, it breaks every sentence down into "Atomic Claims"—the smallest possible pieces of truth.

Instead of checking: "Apple's revenue was great in Q4,"
It checks:
1. The entity (Apple)
2. The time (Q4)
3. The specific number (The revenue amount)

By breaking the sentence into these tiny "atoms," the system can check each one individually. It’s the difference between checking if a whole meal is good and checking every single ingredient to make sure nothing is rotten.

2. The "Math Teacher" Strategy (Type-aware verification)

Standard AI detectors treat all information the same. But in finance, a "date" is different from a "name," and a "name" is different from a "calculation."

FinGround uses a Six-Type Taxonomy. Think of this like a specialized detective team:

The Historian checks if the dates are right.
The Librarian checks if the regulatory rules are real.
The Accountant is the most important. If the AI says, "The profit margin is 20%," the Accountant doesn't just look for the number "20%." It actually pulls out a calculator, finds the original numbers in the company's tables, and re-does the math to see if it actually equals 20%.

Most AI systems fail here because they "read" numbers like words; FinGround "calculates" them like a human professional.

3. The "Fact-Checker" Editor (Grounded Regeneration)

If the system finds a lie, it doesn't just scream "ERROR!" and stop working. It acts like a professional editor.

If the AI assistant writes a sentence that is wrong, FinGround grabs a red pen, crosses out the lie, and rewrites it using the actual facts found in the original documents. Most importantly, it adds "Digital Footnotes." It doesn't just say, "The profit was $5B," it says, "The profit was $5B [See Page 12, Table 3, Row 2]." This allows a human to verify the truth in seconds.

Why does this matter? (The Bottom Line)

The researchers didn't just build this to be "smart"; they built it to be cheap and fast.

They took a massive, expensive "Brain" (GPT-4o) and "distilled" its wisdom into a much smaller, leaner "Brain" (an 8B model). This makes the system:

18 times faster per claim.
15 times cheaper to run.
Highly reliable, catching the kind of math errors that usually trip up even the smartest AI.

In short: FinGround turns a "confident storyteller" into a "rigorous auditor," making AI safe enough for the high-stakes world of global finance.

Technical Summary: FinGround: Detecting and Grounding Financial Hallucinations via Atomic Claim Verification

1. Problem Statement

Large Language Models (LLMs) used in financial services face critical reliability issues, including the fabrication of metrics, invention of regulatory citations, and errors in derived computational quantities (e.g., ratios). These "hallucinations" are particularly dangerous in finance, where accuracy is a regulatory requirement (e.g., the EU AI Act).

Existing hallucination detection methods fail in the financial domain for three specific reasons:

Uniform Treatment: They treat all claims the same, failing to apply specialized logic (like arithmetic re-computation) to numerical or computational claims.
Lack of Structural Awareness: They struggle to verify claims against structured data within tables, leading to "dangling citations" or misaligned cell attributions.
Reasoning vs. Retrieval: Current systems often conflate retrieval quality with verification capability, making it difficult to isolate whether a system is failing because it can't find the data or because it can't verify the logic.

2. Methodology: The FINGROUND Pipeline

The authors propose a three-stage "verify-then-ground" pipeline designed to handle the hybrid nature of financial documents (text and tables).

Stage 1: Finance-Aware Hybrid Retrieval:
The system uses a RoBERTa-based classifier to route queries into three complexity tiers (Simple, Moderate, Complex). It employs a hybrid retrieval strategy combining BM25, dense retrieval (E5-large), and a structure-aware table extraction method that uses a similarity function considering both cell content and column headers. This preserves the row-column relationships necessary for precise attribution.
Stage 2: Atomic Financial Claim Verification:
Instead of verifying whole paragraphs, the system decomposes answers into atomic claims. These claims are classified into a specialized six-type taxonomy: Numerical, Temporal, Entity-Attribute, Comparative, Regulatory, and Computational.
- Type-Routed Verification: Each type uses a specific strategy. Notably, Computational claims undergo "formula reconstruction," where the system identifies the implied formula, retrieves operands from tables, and re-computes the value with a $\pm0.5\%$ tolerance.
- Verdict: A distilled 8B model classifies claims as Supported, Contradicted, or Unverifiable.
Stage 3: Grounded Regeneration:
For claims marked as contradicted or unverifiable, the system performs targeted rewriting. It attaches granular citations at the paragraph or specific table-cell level (e.g., [Table 2, Row: Gross Margin, Col: Q4 2023]). If an answer requires too many repairs ( $\ge3$ ), the system triggers a full re-generation to prevent error compounding.

3. Key Contributions

Domain-Specific Taxonomy: A validated six-type financial claim taxonomy that enables specialized verification (especially for arithmetic).
Retrieval-Equalized Evaluation: A new methodological standard where all baselines receive identical retrieval, allowing researchers to isolate the true value of the verification architecture from the retrieval engine.
Efficient Distillation: The authors distilled the capabilities of GPT-4o into an 8B Llama-3 model, achieving $91.4\%$ F1 detection accuracy while reducing per-claim latency by $18\times$ and cost by $15.7\times$ .
Structural Attribution: A mechanism for precise table-cell-level grounding, addressing the "dangling citation" problem common in standard RAG.

4. Key Results

Hallucination Reduction: FINGROUND achieved a 78% relative reduction in hallucination rates compared to GPT-4o + Chain-of-Thought (CoT) across major benchmarks (FinQA, TAT-QA, FinanceBench).
Verification Value: Under retrieval-equalized conditions, FINGROUND reduced hallucination rates by 68–76% over the strongest baselines, proving the architecture's independent value.
Computational Accuracy: The formula reconstruction strategy provided a massive boost, specifically increasing F1 by 18.9 points for computational claims compared to generic detectors.
Production Viability: The distilled 8B model enables a deployment cost of approximately $0.003 per query, making it economically feasible for high-volume financial environments.
Human Alignment: A pilot study with 24 financial analysts showed high practical acceptability (mean 4.1/5) and stable "alert fatigue" levels.

5. Significance

FINGROUND represents a shift from "general-purpose" hallucination detection to "domain-aware" verification. By treating financial data as a structured, multi-modal problem (text + math + tables), it provides a production-ready path toward meeting the strict accuracy and interpretability mandates of upcoming regulations like the EU AI Act. Its ability to provide cell-level citations transforms the AI from a "black box" into a verifiable tool for professional analysts.