Abstract: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.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2603.11583 [cs.CL] |
| (or arXiv:2603.11583v4 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2603.11583 arXiv-issued DOI via DataCite |
From: Ofir Marom [view email]
[v1]
Thu, 12 Mar 2026 06:17:09 UTC (11 KB)
[v2]
Thu, 26 Mar 2026 15:32:11 UTC (12 KB)
[v3]
Sat, 4 Apr 2026 10:20:35 UTC (13 KB)
[v4]
Sat, 23 May 2026 15:14:02 UTC (15 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。