TLDR: The paper introduces FRED, a framework for detecting and editing factual errors (hallucinations) in financial language models. It defines a domain-specific error taxonomy, creates a synthetic dataset by injecting errors into financial Q&A corpora, and fine-tunes small language models (SLMs) like Phi-4 and Qwen3 to identify and correct these inaccuracies. The fine-tuned Phi-4 model significantly outperforms larger models like OpenAI-o3 in hallucination detection and shows competitive performance in editing, offering a practical and computationally efficient solution for improving factual reliability in financial AI.
Large Language Models (LLMs) have become incredibly powerful, but they often face a significant challenge: generating ‘hallucinations.’ These are outputs that sound plausible but are factually incorrect, which is a major concern, especially in critical fields like finance where accuracy is paramount.
A new research paper introduces FRED, which stands for Financial Retrieval-Enhanced Detection and Editing of Hallucinations in Language Models. This work presents an effective method for identifying and correcting factually inaccurate content produced by these models, particularly when they are used in financial contexts.
Addressing Hallucinations in Finance
The core idea behind FRED is to make LLMs more reliable by teaching them to spot and fix their own mistakes. The researchers developed a system that works in two main stages: detection and editing. This is crucial because while existing methods like Retrieval-Augmented Generation (RAG) help ground responses in evidence, they don’t completely eliminate errors.
How FRED Works: A Step-by-Step Approach
The team behind FRED started by defining a specific set of error types common in financial data. This ‘domain-specific error taxonomy’ includes categories such as numerical miscalculations, incorrect entity references (like company names), temporal inconsistencies (wrong dates), and relation mismatches. They also identified contradictory statements and unverifiable information.
To train their models, they created a unique ‘synthetic dataset.’ They took existing financial question-answering datasets, like FinQA and TAT-QA, and intentionally inserted these tagged errors into the answers. This allowed them to create a controlled environment for supervised training, where the models could learn to recognize and correct specific types of inaccuracies.
With this specially prepared data, they fine-tuned several smaller language models (SLMs), including Phi-4, Phi-4-mini, Qwen3-4B, and Qwen3-14B. The goal was to enable these models to perform ‘span-level hallucination detection and correction,’ meaning they could pinpoint incorrect phrases and replace them with accurate, contextually grounded edits.
Impressive Results in Detection and Editing
The experiments showed remarkable improvements. In hallucination detection, the fine-tuned Phi-4 model significantly outperformed OpenAI-o3, a leading proprietary model. Specifically, Phi-4 achieved an 8% improvement in binary F1 score and a 30% gain in overall detection performance. Even the smaller Phi-4-mini model, with only 4 billion parameters, maintained competitive performance, showing only a slight drop compared to OpenAI-o3.
When it came to hallucination editing, the results were also very promising. On financial datasets, the fine-tuned Phi4-36k* model performed comparably to, and in some cases slightly better than, OpenAI-o3. This indicates that these smaller, specialized models can effectively correct errors, not just identify them. The research also highlighted that increasing the amount of training data consistently improved the models’ performance.
Also Read:
- PrismRAG: A New Approach to Enhance AI’s Factual Accuracy in Question Answering
- Beyond Relevance: How AI Models Are Learning to Pick Truly Useful Information for Better Answers
A Practical Solution for Financial AI
The FRED framework offers a practical and efficient solution for enhancing the trustworthiness of large language models in financial applications. By focusing on domain-specific errors and leveraging the power of fine-tuned small language models, it provides a way to detect and edit factual inconsistencies with high accuracy, even with limited computational resources.
This work not only addresses a critical challenge in financial text generation but also introduces a generalizable framework that could improve the reliability of LLMs across various other applications. The code and data for this research are openly available, fostering further development and adoption in the community. You can find more details about this research in the paper: FRED: Financial Retrieval-Enhanced Detection and Editing of Hallucinations in Language Models.
