In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations
arXiv:2602.15456v1 Announce Type: new Abstract: Agents based on Large Language Models (LLMs) are increasingly being deployed as interfaces to information on online platforms. These agents filter, prioritize, and synthesize information retrieved from the platforms' back-end databases or via web search. In these scenarios, LLM agents govern the information users receive, by drawing users' attention to particular instances of retrieved information at the expense of others. While much prior work has focused on biases in the information LLMs themselves generate, less attention has been paid to the factors that influence what information LLMs select and present to users. We hypothesize that when information is attributed to specific sources (e.g., particular publishers, journals, or platforms), current LLMs exhibit systematic latent source preferences- that is, they prioritize information from some sources over others. Through controlled experiments on twelve LLMs from six model providers,
arXiv:2602.15456v1 Announce Type: new Abstract: Agents based on Large Language Models (LLMs) are increasingly being deployed as interfaces to information on online platforms. These agents filter, prioritize, and synthesize information retrieved from the platforms' back-end databases or via web search. In these scenarios, LLM agents govern the information users receive, by drawing users' attention to particular instances of retrieved information at the expense of others. While much prior work has focused on biases in the information LLMs themselves generate, less attention has been paid to the factors that influence what information LLMs select and present to users. We hypothesize that when information is attributed to specific sources (e.g., particular publishers, journals, or platforms), current LLMs exhibit systematic latent source preferences- that is, they prioritize information from some sources over others. Through controlled experiments on twelve LLMs from six model providers, spanning both synthetic and real-world tasks, we find that several models consistently exhibit strong and predictable source preferences. These preferences are sensitive to contextual framing, can outweigh the influence of content itself, and persist despite explicit prompting to avoid them. They also help explain phenomena such as the observed left-leaning skew in news recommendations in prior work. Our findings advocate for deeper investigation into the origins of these preferences, as well as for mechanisms that provide users with transparency and control over the biases guiding LLM-powered agents.
Executive Summary
This article sheds light on the hitherto unexplored area of latent source preferences in Large Language Models (LLMs). The authors conducted a comprehensive study on twelve LLMs from six model providers, revealing systematic biases in information selection and presentation. The study's findings suggest that LLMs prioritize information from certain sources over others, with these preferences being sensitive to contextual framing and persisting despite explicit prompting. The research contributes significantly to our understanding of LLMs' decision-making processes and highlights the need for mechanisms providing users with transparency and control over biases. The study's implications are far-reaching, with potential consequences for the dissemination of information, the accuracy of news recommendations, and the trustworthiness of LLM-powered agents.
Key Points
- ▸ LLMs exhibit systematic latent source preferences in selecting and presenting information to users.
- ▸ These preferences are sensitive to contextual framing and can outweigh the influence of content itself.
- ▸ Explicit prompting has little effect on mitigating these biases, which persist despite attempts to avoid them.
Merits
Strength
The study's rigorous methodology, involving controlled experiments on multiple LLMs from different providers, lends credibility to its findings and provides a comprehensive understanding of the issue.
Demerits
Limitation
The study primarily focuses on LLMs from six model providers, which may not be representative of the broader LLM landscape, potentially limiting the generalizability of its findings.
Expert Commentary
This study marks a significant advancement in our understanding of the complex decision-making processes underlying LLMs. However, its findings also underscore the need for more nuanced and context-dependent approaches to mitigating biases. Furthermore, the study's emphasis on the importance of transparency and user control highlights the need for more stringent standards for LLM development and deployment. Ultimately, this research serves as a critical reminder of the importance of critically evaluating and addressing the limitations and biases inherent in AI systems.
Recommendations
- ✓ Future research should investigate the origins and mechanisms driving latent source preferences in LLMs, with a focus on developing more effective mitigation strategies.
- ✓ LLM developers and deployers should prioritize the implementation of transparency and control mechanisms, ensuring that users receive accurate and unbiased information.
