LLM-Derived Preference Judgments Are Not Self-Consistent
Researchers have found that large language models (LLMs) are not consistent in their preference judgments. When asked how much someone would be willing to pay for an item, LLMs often provide conflicting answers. This is a problem because many AI systems rely on these judgments to make decisions. The study used six different LLMs and tested them with scenarios involving flights, apartments, and hotels. The results show that the models' responses are not self-consistent, meanin
Researchers have found that large language models (LLMs) are not consistent in their preference judgments. When asked how much someone would be willing to pay for an item, LLMs often provide conflicting answers. This is a problem because many AI systems rely on these judgments to make decisions. The study used six different LLMs and tested them with scenarios involving flights, apartments, and hotels. The results show that the models' responses are not self-consistent, meaning they cannot be accurately summarized by a single utility function. This has implications for how AI systems interpret human preferences.
---
Why it matters: This matters because many AI systems rely on LLM-derived preference judgments to make decisions, such as choosing actions based on estimated utility. The inconsistency in these judgments can lead to suboptimal or even incorrect choices.
Source: https://arxiv.org/abs/2608.17644
This article was originally published at: https://arxiv.org/abs/2608.17644