IDEA: An Interpretable and Editable Decision-Making Framework for LLMs via Verbal-to-Numeric Calibration
In the rapidly evolving landscape of artificial intelligence, large language models (LLMs) have gained prominence for their ability to assist in decision-making processes. However, their deployment in high-stakes scenarios has been hindered by several critical issues, including miscalibrated probabilities, unfaithful explanations, and a lack of precision in incorporating expert knowledge. Addressing these challenges, a new framework named IDEA has been proposed, which stands for “Interpretable and Editable Decision-making Framework.”
Overview of IDEA
The IDEA framework aims to extract decision-making knowledge from LLMs and represent it in an interpretable parametric model. This representation focuses on semantically meaningful factors, enabling a more transparent understanding of the decision-making process. By employing a combination of innovative techniques, IDEA enhances the calibration of probabilities while facilitating effective collaboration between human experts and AI systems.
Key Features of IDEA
- Joint Learning of Verbal-to-Numerical Mappings: IDEA utilizes an Expectation-Maximization (EM) approach to learn the relationships between verbal inputs and their corresponding numerical outputs. This ensures that the model can produce more accurate and reliable predictions.
- Correlated Sampling: The framework implements correlated sampling methods that maintain dependencies between factors. This characteristic is crucial for accurately reflecting real-world scenarios where variables are interconnected.
- Direct Parameter Editing: IDEA allows for direct editing of model parameters with mathematical guarantees. This feature empowers users to refine decisions based on expert insights, further enhancing the model’s adaptability and reliability.
- Calibrated Probabilities: The framework is designed to produce calibrated probabilities, addressing one of the significant limitations of traditional LLMs. This capability is vital for ensuring that the decisions made by the AI are both trustworthy and actionable.
Performance and Results
Experiments conducted across five diverse datasets demonstrate the effectiveness of IDEA. The framework, when applied to the Qwen-3-32B model, achieved an impressive accuracy score of 78.6%. This performance significantly outstripped that of competing models, including DeepSeek R1, which scored 68.1%, and GPT-5.2, which achieved 77.9%. Notably, IDEA was able to attain perfect factor exclusion and exact calibration—levels of precision that remain elusive through traditional prompting methods alone.
Access and Implementation
For researchers and practitioners interested in exploring the capabilities of the IDEA framework, the implementation is publicly accessible on GitHub. This open-source availability encourages collaboration and further development within the community, fostering innovation in the application of LLMs in decision-making contexts.
Conclusion
As AI continues to permeate various sectors, the introduction of IDEA marks a significant step forward in enhancing the interpretability and reliability of decision-making processes powered by large language models. By addressing the critical issues of miscalibration and lack of expert integration, IDEA sets a new standard for the responsible use of AI in high-stakes environments.