Mitigating Hallucination in Financial Retrieval-Augmented Generation via Fine-Grained Knowledge Verification

TL;DR

The paper tackles hallucination in financial retrieval-augmented generation by introducing RLFKV, which decomposes responses into atomic knowledge units encoded as a quadruple $(entity, metric, value, timestamp)$ and verifies each unit against retrieved documents. It defines two rewards, $r_f$ (faithfulness) and $r_i$ (informativeness), combined as $r=(r_f+r_i)/2$, and optimizes with a PPO-based objective $\mathcal{L}$, obviating the need for annotated references. Experiments on the FDD and the newly released FDD-ANT datasets show consistent improvements in faithfulness while maintaining informativeness, demonstrating generalization beyond stock-related descriptions. Error analysis highlights remaining challenges in temporal references and numerical precision, pointing to future work on temporal alignment and rounding accuracy.

Abstract

In financial Retrieval-Augmented Generation (RAG) systems, models frequently rely on retrieved documents to generate accurate responses due to the time-sensitive nature of the financial domain. While retrieved documents help address knowledge gaps, model-generated responses still suffer from hallucinations that contradict the retrieved information. To mitigate this inconsistency, we propose a Reinforcement Learning framework enhanced with Fine-grained Knowledge Verification (RLFKV). Our method decomposes financial responses into atomic knowledge units and assesses the correctness of each unit to compute the fine-grained faithful reward. This reward offers more precise optimization signals, thereby improving alignment with the retrieved documents. Additionally, to prevent reward hacking (e.g., overly concise replies), we incorporate an informativeness reward that encourages the policy model to retain at least as many knowledge units as the base model. Experiments conducted on the public Financial Data Description (FDD) task and our newly proposed FDD-ANT dataset demonstrate consistent improvements, confirming the effectiveness of our approach.

Figures (4)

• Figure 1: An example of hallucination by the base model Qwen3-8B in financial data description.
• Figure 2: The framework of RLFKV.
• Figure 3: Performance comparison and reward analysis of fine- vs. coarse-grained rewards based on Qwen3-8B model.
• Figure 4: Error analysis on FDD-ANT dataset.