Research map / Large language models / Agents, code and math

Personalized Decision Modeling: Utility Optimization or Textualized-Symbolic Reasoning

NeurIPS 2025 · Large language models › Agents, code and math · approach: Inductive and abstract reasoning evaluation · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ATHENA, a two-stage framework that discovers group-level symbolic utility functions with LLM-augmented evolutionary search and then personalizes LLM predictions via TextGrad-optimized individual semantic templates, evaluated on Swissmetro travel mode and COVID-19 vaccine choice.

Problem

Individual human choice decisions (e.g., travel mode, vaccine uptake) diverge from population-optimal predictions; utility-based, ML, and LLM-based models fail to capture personalized utility logic that mixes numeric attributes and semantic preferences.

Approach

Combine LLM-driven symbolic regression for group-level utility functions (feedback-informed evolutionary search) with individual-level textual-gradient (TextGrad) optimization of semantic persona templates initialized from the discovered utility, then predict choices with the LLM conditioned on both.

Key results

• ATHENA with GPT-4o-mini reaches 0.8134 accuracy and 0.7655 F1 on Swissmetro, beating the strongest baseline by at least 6% accuracy and 6.5% F1.
• On Vaccine, ATHENA with GPT-4o-mini achieves 0.7345 accuracy and 0.7161 F1, versus 0.6650 accuracy for TabNet among baselines.
• Ablations show removing either stage lowers performance; semantic adaptation alone reaches 0.6044 accuracy on Swissmetro, symbolic discovery alone 0.6566.

Measured effects reported in the paper

Surprising findings

• ATHENA has higher cross-entropy than XGBoost and random forest on Swissmetro (1.0863 vs 0.7040), which the authors attribute to more conservative probabilities.
• Few-shot Gemini has the highest Swissmetro accuracy among pure prompting methods but very high cross-entropy (8.72).

Limitations

• High computational cost from textual gradient optimization, scaling with population size.
• Assumes a shared symbolic utility per demographic group; groups with high internal diversity may be poorly represented.
• All results from single runs with fixed seeds; no error bars.

Experimental setup

Datasets: Swissmetro COVID-19 vaccine international survey

Benchmarks: Swissmetro COVID-19 Vaccine choice

Metrics: Accuracy F1 Cross-Entropy AUC

Baselines: Zeroshot Zeroshot-CoT Fewshot TextGrad MNL CLogit Latent Class MNL Logistic Regression Random Forest XGBoost BERT TabNet

Builds on: Random Utility Maximization LLM-SR LASR TextGrad Multinomial Logit

Closest papers by meaning

• DeLLMa: Decision Making Under Uncertainty with Large Language Models (ICLR 2025 · similarity 0.86)
• Preference Learning of Latent Decision Utilities with a Human-like Model of Preferential Choice (NeurIPS 2024 · similarity 0.84)
• Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice (ICLR 2025 · similarity 0.84)
• Enhancing Decision Making Through the Integration of Large Language Models and Operations Research Optimization (AAAI 2025 · similarity 0.83)
• Decision Theoretic Foundations for Conformal Prediction: Optimal Uncertainty Quantification for Risk-Averse Agents (ICML 2025 · similarity 0.83)
• Decision-Making Behavior Evaluation Framework for LLMs under Uncertain Context (NeurIPS 2024 · similarity 0.83)