Risk-MCTS: Table-Reward Enhanced LLM with Monte Carlo Tree Search for Interpretable Financial Risk Detection

ACL ARR 2025 May Submission1714 Authors

Abstract: Financial risk detection is an important yet challenging task. Exsiting machine learning or deep learning-based approaches have primarily treated it as a binary classification task. Although these approaches already achieved good model performance, they still fail to capture complex risk patterns as well as to provide interpretable steps for financial risk detection. To address aforementioned research limitations, we propose this Risk-MCTS, a novel framework integrating large language model with monte-carlo tree search method, which leverages both cell data and headers in financial tables for step-by-step risk inference. To better understanding financial tabular data, we dedicately design a table reward model which quantitatively evaluates table content during the analytical process, thereby enhancing the detection of salient financial content. Extensive experiments demonstrate that prooposed Risk-MCTS achieves the SOTA model performance on real world datasets with respect to a numder of evaluation criteria.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: financial/business NLP

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 1714