TLDR: The research paper “FAITH: A Framework for Assessing Intrinsic Tabular Hallucinations in Finance” introduces a new, scalable framework to evaluate Large Language Model (LLM) hallucinations in financial applications. It addresses the critical challenge of LLMs generating inconsistent or incorrect information from tabular financial data. The framework uses a context-aware masked span prediction task over S&P 500 annual reports, categorizing reasoning into four complexity levels. Experiments show that while top LLMs perform well on simple tasks, their accuracy significantly degrades with increasing reasoning complexity, particularly in multi-step calculations, highlighting the ongoing challenge of ensuring factual integrity in financial AI.
Large Language Models (LLMs) are transforming many industries, and finance is no exception. From automating information extraction to powering client-facing chatbots, LLMs offer significant potential. However, a critical challenge remains: hallucination. In the financial sector, where even minor numerical errors can have severe consequences for decision-making and regulatory compliance, ensuring accuracy is paramount.
A new research paper, titled “FAITH: A Framework for Assessing Intrinsic Tabular Hallucinations in Finance,” introduces a rigorous and scalable framework designed to evaluate these intrinsic hallucinations in LLMs specifically within financial contexts. Intrinsic hallucinations occur when an LLM’s generated output contradicts the provided input context, a particularly dangerous flaw when dealing with sensitive financial data.
The authors, Mengao Zhang, Jiayu Fu, Tanya Warrier, Yuwen Wang, Tianhui Tan, and Ke-Wei Huang, highlight that existing hallucination benchmarks often fall short because they don’t account for the unique requirements of financial applications. These applications frequently rely on context-dependent, numerical, and often proprietary tabular data.
The FAITH Framework: A New Approach to Evaluation
The core of this study is the development of the FAITH framework. It conceptualizes hallucination evaluation as a context-aware masked span prediction task. In simple terms, numerical values are masked (hidden) within real-world financial documents, and the LLM is prompted to recover them. This approach ensures that the evaluation directly measures the model’s ability to accurately extract and calculate information from provided financial tables and text.
To ensure the reliability of this evaluation, the framework incorporates strict masking criteria:
- Uniqueness: Each masked span must have only one correct answer.
- Consistency: The true answer must be consistent with the source document, preventing evaluation based on incorrect source data.
- Answerability: The masked span must be inferable from the provided context, ensuring the task is solvable by an LLM.
The researchers used S&P 500 annual reports (10-K filings) as their data source, leveraging their regulatory rigor and consistency to minimize contradictions. They also employed LLMs themselves to annotate the answerability of spans, validating this automated process against human annotations.
Robust Evaluation and Reasoning Scenarios
The evaluation protocol is designed to be robust, handling nuances like formatting variations (e.g., “$1,230 million” vs. “USD 1.23 billion”). It uses a precision-relaxed matching system that normalizes numerical values and checks for unit consistency.
To provide a fine-grained analysis of model performance, the study categorizes financial reasoning into four distinct scenarios:
- Direct Lookup: The answer is directly found in a table.
- Comparative Calculation: Requires simple calculation across different time periods or categories (e.g., year-over-year growth).
- Bivariate Calculation: Involves a simple calculation between two distinct metrics (e.g., a financial ratio).
- Multivariate Calculation: Requires multi-step reasoning over three or more metrics or a sequence of arithmetic operations.
This classification allows for a detailed understanding of where LLMs struggle most.
Key Findings: Complexity and Hallucination Risk
The experiments, conducted on a range of state-of-the-art LLMs, revealed a clear performance hierarchy. Top-tier proprietary models like Claude-Sonnet-4 and Gemini-2.5-Pro showed high overall accuracy, but even they exhibited a 4-8% error rate, which is significant in finance. Smaller open-source models performed considerably worse, often proving unsuitable for tasks requiring financial fidelity.
A crucial finding was the direct relationship between reasoning complexity and hallucination risk. While models performed adequately on direct lookups, accuracy systematically decreased as tasks required more complex calculations and logical inference. The most significant performance drop occurred in the Multivariate Calculation scenario, where many models scored near 0.0%, indicating a fundamental breakdown in their ability to reason under complex contexts and leading to output fabrication.
The study also identified common error patterns, with “scale error” being particularly prevalent (e.g., reporting “$150” instead of “$150 million”). A detailed case study illustrated how only the top-performing models could accurately infer latent variables by synthesizing information across both tabular and textual contexts, while others failed to integrate tabular data effectively.
Also Read:
- Unpacking AI’s Factual Accuracy: A Deep Dive into Language Model Fact-Checking
- The Hidden Flaw: How Large Language Models Handle Bad Code Instructions
Towards More Trustworthy Financial AI
This research provides an essential foundation for evaluating hallucinations in finance. It highlights that while leading LLMs are improving, fundamental challenges in ensuring factual integrity, especially regarding numerical scale in complex reasoning, remain a primary barrier to their safe deployment in accuracy-critical financial workflows. The authors plan to expand the dataset to include more document types and complex scenarios in future work, emphasizing that robust, domain-specific evaluation is crucial for building more reliable and trustworthy LLMs for financial applications.
You can read the full research paper here: FAITH: A Framework for Assessing Intrinsic Tabular Hallucinations in Finance.
