Paper Reading Journals #10
Personalized Decision Modeling: Utility Optimization or Textualized-Symbolic Reasoning
Basic Info
| Item | Content |
|---|---|
| Title | Personalized Decision Modeling: Utility Optimization or Textualized-Symbolic Reasoning |
| Journal | arXiv |
| Author | Zhao, Y., Zhao, Y., Du, H., & Yang, H. F. |
| Year Published | 2025 |
| Link | https://doi.org/10.48550/arXiv.2511.02194 |
| Geography of Study | N/A (no specific) |
| Area of Study | Decision Modeling, LLM, Symbolic-LLM Integration |
Description
(Including data and methods)
Data: Two real-world dataset use for evaluation and validation:
- Swissmetro dataset (Link) - Transportation choice study originally performed by researchers from ETH Zurich.
- COVID-19 Vaccination Choice dataset (Link) - Covid 19 vaccination choice research lead by CUNY.
Model:
- Adaptive Textual-symbolic Human-centric Reasoning (ATHENA): A two-staged personalized human decision making modeling framework powered by LLM
- LLM used: GPT-4o-mini, gemini-2.0-flash
Main Contribution
(Innovative method or Application)
Innovation Type: Application Innovation
This model consists of two stages. In Stage 1, it simulates decision-making at the population level and derives a symbolic utility expression that represents group-level behavioral patterns.
In the second stage, which introduces the large language model (LLM) component, the framework is enhanced with personalization capabilities.Specifically, the LLM builds upon the previously identified group-level symbolic utility expression to incorporate additional individual-level variables—such as semantic attributes and textual descriptions—thereby generating a semantic template for each individual. Subsequently, this template is refined through Textual Gradient Optimization, producing the final individual-level semantic adaptation function.
Data Source
Two real-world dataset use for evaluation and validation:
- Swissmetro dataset (Link) - Transportation choice study originally performed by researchers from ETH Zurich.
- COVID-19 Vaccination Choice dataset (Link) - Covid 19 vaccination choice research lead by CUNY.
Favorite Thing
This article provide a potential way to integrate LLM in our research framework. By modifying the LLM’s semantic features and textual descriptions, we can make it more capable of handling healthcare decision-making, thereby generating a healthcare decision model that compensates for unquantifiable choice parameters. This approach can serve as an alternative to traditional health economics methods such as Markov Chain and Agent-Based Modeling (ABM), forming a Synthetic Healthcare System Modeling Framework that maintains interpretability even at the individual level.