UtilityMax Prompting: A Formal Framework for Multi-Objective Large Language Model Optimization
This addresses the problem of imprecise prompt specification for LLM users in multi-objective scenarios, offering an incremental improvement over natural language baselines.
The paper tackles the ambiguity of natural language prompts in multi-objective LLM tasks by introducing UtilityMax Prompting, a formal framework using mathematical language and influence diagrams to optimize answers based on expected utility, resulting in consistent improvements in precision and NDCG on the MovieLens 1M dataset across three models.
The success of a Large Language Model (LLM) task depends heavily on its prompt. Most use-cases specify prompts using natural language, which is inherently ambiguous when multiple objectives must be simultaneously satisfied. In this paper we introduce UtilityMax Prompting, a framework that specifies tasks using formal mathematical language. We reconstruct the task as an influence diagram in which the LLM's answer is the sole decision variable. A utility function is defined over the conditional probability distributions within the diagram, and the LLM is instructed to find the answer that maximises expected utility. This constrains the LLM to reason explicitly about each component of the objective, directing its output toward a precise optimization target rather than a subjective natural language interpretation. We validate our approach on the MovieLens 1M dataset across three frontier models (Claude Sonnet 4.6, GPT-5.4, and Gemini 2.5 Pro), demonstrating consistent improvements in precision and Normalized Discounted Cumulative Gain (NDCG) over natural language baselines in a multi-objective movie recommendation task.