Hui Chen’s Research: New Ways to Interpret and Steer AI

Hui Chen, the Nomura Professor of Finance at the MIT Sloan School of Management, is developing research focused on interpreting and steering large language models (LLMs) to improve their practical application in the financial sector.

The research addresses the challenge of interpreting the internal mechanisms of AI models, a hurdle that has frequently limited the adoption of LLMs in professional investing practices.

Quantifying Prediction Uncertainty

In a working paper updated in March 2026 titled Out of the Black Box: Uncertainty Quantification for LLMs via Conditional Probabilities, Chen and co-authors Antoine Didisheim and Luciano Somoza examine how autoregressive LLMs generate text by sampling from estimated probability distributions over the next token based on prior context.

The researchers developed an entropy-based measure of prediction uncertainty, which they refer to as inner confidence. This method aims to provide a more reliable metric for the accuracy of model predictions compared to self-declared confidence, which the study found exhibits significant decoding biases and provides no comparable performance gains.

The economic relevance of this measure was tested through news classification and the formation of long-short portfolios. The results indicated that:

• LLM predictions with higher inner confidence were systematically more accurate in news classification.
• Portfolios based on high-confidence predictions achieved a Sharpe ratio approximately 20% higher than the unconditional benchmark.
• Portfolios based on low-confidence predictions yielded no excess returns.

Integrating Economic Theory with Machine Learning

Beyond uncertainty quantification, Chen is working on bridging the gap between structural economic models and machine learning. In a paper updated in January 2026 titled Teaching Economics to the Machines, authored with Yuhan Cheng, Yanchu Liu, and Ke Tang, the researchers address the limitations of both approaches.

The research notes that while structural economic models are parsimonious and interpretable, they often suffer from poor data fit and limited forecasting performance. Conversely, machine learning models offer flexibility but are prone to overfitting and weak out-of-distribution generalization.

To resolve this, the team proposed a theory-guided transfer learning framework. This framework integrates structural restrictions derived from economic theory directly into machine learning models to combine the interpretability of economic theory with the predictive power of AI.

Academic and Professional Context

Hui Chen is a research associate at the National Bureau of Economic Research and specializes in asset pricing, financial constraints, credit risk, liquidity risk, and financial machine learning.

His broader academic work includes research into the costs of debt financing, specifically how debt servicing costs can drain a firm’s precautionary cash holdings and increase expected external financing costs.

The effort to provide what has been described as a brain scan for LLMs is part of a larger effort starting in early 2024 to help investors understand the internal operations of these models to make them more viable for practical investment use.