Personalized Decision Modeling: Utility Optimization or Textualized-Symbolic Reasoning

TL;DR

ATHENA addresses the gap between population-optimal policies and individual behavior in high-stakes decisions by combining group-level symbolic utility discovery with individual-level semantic adaptation using LLMs. It first identifies robust symbolic utility functions for demographic groups and then personalizes decisions by adapting textual templates guided by these utilities, via a TextGrad-based process. On travel mode choice (Swissmetro) and vaccine uptake tasks, ATHENA consistently outperforms traditional utility-based models, standard machine learning methods, and pure LLM approaches, with a minimum 6.5% improvement in F1, and ablations show both stages are necessary. This work delivers a scalable, interpretable framework that marries symbolic reasoning with semantic adaptation to model nuanced human decisions.

Abstract

Decision-making models for individuals, particularly in high-stakes scenarios like vaccine uptake, often diverge from population optimal predictions. This gap arises from the uniqueness of the individual decision-making process, shaped by numerical attributes (e.g., cost, time) and linguistic influences (e.g., personal preferences and constraints). Developing upon Utility Theory and leveraging the textual-reasoning capabilities of Large Language Models (LLMs), this paper proposes an Adaptive Textual-symbolic Human-centric Reasoning framework (ATHENA) to address the optimal information integration. ATHENA uniquely integrates two stages: First, it discovers robust, group-level symbolic utility functions via LLM-augmented symbolic discovery; Second, it implements individual-level semantic adaptation, creating personalized semantic templates guided by the optimal utility to model personalized choices. Validated on real-world travel mode and vaccine choice tasks, ATHENA consistently outperforms utility-based, machine learning, and other LLM-based models, lifting F1 score by at least 6.5% over the strongest cutting-edge models. Further, ablation studies confirm that both stages of ATHENA are critical and complementary, as removing either clearly degrades overall predictive performance. By organically integrating symbolic utility modeling and semantic adaptation, ATHENA provides a new scheme for modeling human-centric decisions. The project page can be found at https://yibozh.github.io/Athena.

Figures (6)

• Figure 1: Overview of the proposed Athena framework.Group-level symbolic utility discovery: Symbolic & semantic constraints library feed an LLM-driven symbolic-optimization engine that iteratively proposes candidate utility functions, scores them with loss $\mathcal{L}_{g}$, and prunes the search via analysis, crossover, and mutation. Red rings in the contour maps illustrate how the feasible solution space shrinks across iterations until the optimal formula $f^{*}_{g}$ is selected. Individual-level semantic adaptation: The optimal group utility $f^{*}_{g}$ seeds a personalized template space. For each individual $i$, TextGrad computes textual gradients of an individual loss and updates the template $\mathcal{P}^{t}_{i}$ into a more personalized decision rule $\mathcal{P}^{t+1}_{i}$. Finally, the optimal $\mathcal{P}^{*}_{i}$ is used to predict personal decisions.
• Figure 2: Athena pipeline applied to a travel–mode choice example. Here we use Swissmetro as an example to illustrate Athena framework. The Initialization panel encodes conceptual constraints, a mixed semantic–numerical feature space, and a symbolic library of operations. In Group‑level symbolic optimization, an LLM samples, scores, and prunes candidate utility expressions for each alternative to produce compact formulas $\{f_g^*\}$ that best explain group behavior. In Individual semantic adaptation, each $f_g^*$ seeds a group‑specific prompt template $\mathcal{P}_i^0$, which is refined to a personalized template via TextGrad to capture individual heterogeneity ($\mathcal{P}_i^0 \rightarrow \mathcal{P}_i^*$).
• Figure 3: Accuracy trajectories of symbolic regression. As shown here, the accuracy keeps growing in 30 iterations for all four groups in the vaccine dataset. Each orange dot is the average accuracy at a given iteration; the red dashed curve is a fit showing the overall upward trend and convergence.
• Figure 4: Aggregated fragment importance extracted from the learned symbolic utilities. For each task, we plot the top-ranked atomic fragments $\varphi_m$ that appear in the three best group-level utility formulas and weight them by fragment importance score \ref{['eq:frag_score']}. Values shown here are the normalized scores in $[0,1]$.
• Figure 7: Athena yields improvements on the classes that matter most yet were previously hard to distinguish.