LLM Agents Are Hypersensitive to Nudges

TL;DR

The work reveals that LLM-based agents are hypersensitive to choice-architectures (nudges) in sequential decision tasks, diverging from human patterns despite superficial similarities. By adapting a resource-rational, meta-level decision task and testing across multiple models and prompting conditions, the study shows canonical nudges (default options, suggestions, information highlighting) and optimal nudges can markedly influence LLM choices and information acquisition, with only partial alignment to human behavior. Simple prompt strategies (zero-shot CoT) can shift distributions, while few-shot prompting with human data improves alignment, yet hypersensitivity persists. The findings highlight the need for behavioral testing and robust design frameworks when deploying LLM agents in complex environments, and suggest that human-guided nudges can, in some cases, improve both human and model performance via an optimal nudging approach.

Abstract

LLMs are being set loose in complex, real-world environments involving sequential decision-making and tool use. Often, this involves making choices on behalf of human users. However, not much is known about the distribution of such choices, and how susceptible they are to different choice architectures. We perform a case study with a few such LLM models on a multi-attribute tabular decision-making problem, under canonical nudges such as the default option, suggestions, and information highlighting, as well as additional prompting strategies. We show that, despite superficial similarities to human choice distributions, such models differ in subtle but important ways. First, they show much higher susceptibility to the nudges. Second, they diverge in points earned, being affected by factors like the idiosyncrasy of available prizes. Third, they diverge in information acquisition strategies: e.g. incurring substantial cost to reveal too much information, or selecting without revealing any. Moreover, we show that simple prompt strategies like zero-shot chain of thought (CoT) can shift the choice distribution, and few-shot prompting with human data can induce greater alignment. Yet, none of these methods resolve the sensitivity of these models to nudges. Finally, we show how optimal nudges optimized with a human resource-rational model can similarly increase LLM performance for some models. All these findings suggest that behavioral tests are needed before deploying models as agents or assistants acting on behalf of users in complex environments.

Figures (28)

• Figure 1: (Left) A screenshot from the original game setup callaway2023optimal displaying the number of prizes with their points, and the hidden cells for each basket. (Right) Our reconstruction of the game with just text and minor rephrasing, indicating hidden cells with a question mark.
• Figure 2: An example of the default option nudge with an option to accept or decline the default basket. If declined, the game continues as in the control condition.
• Figure 3: An example of an early suggested alternative revealing the highest value in the basket at no extra cost.
• Figure 4: An example of information highlighting where revealing cells for the nudged prize C costs one point, while it costs three points in the other prizes.
• Figure 5: An example of a derived optimal nudge from the resource rational model in callaway2023optimal. A total of six boxes are initially revealed.