From Delegates to Trustees: How Optimizing for Long-Term Interests Shapes Bias and Alignment in LLM

Large language models (LLMs) have shown promising accuracy in predicting survey responses and policy preferences, which has increased interest in their potential to represent human interests in various domains. Most existing research has focused on behavioral cloning, effectively evaluating how well models reproduce individuals' expressed preferences. Drawing on theories of political representation, we highlight an underexplored design trade-off: whether AI systems should act as delegates, mirroring expressed preferences, or as trustees, exercising judgment about what best serves an individual's interests. This trade-off is closely related to issues of LLM sycophancy, where models can encourage behavior or validate beliefs that may be aligned with a user's short-term preferences, but is detrimental to their long-term interests. Through a series of experiments simulating votes on various policy issues in the U.S. context, we apply a temporal utility framework that weighs short and long-term interests (simulating a trustee role) and compare voting outcomes to behavior-cloning models (simulating a delegate). We find that trustee-style predictions weighted toward long-term interests produce policy decisions that align more closely with expert consensus on well-understood issues, but also show greater bias toward models' default stances on topics lacking clear agreement. These findings reveal a fundamental trade-off in designing AI systems to represent human interests. Delegate models better preserve user autonomy but may diverge from well-supported policy positions, while trustee models can promote welfare on well-understood issues yet risk paternalism and bias on subjective topics.

The Empty Chair: Using LLMs to Raise Missing Perspectives in Policy Deliberations

Deliberation is essential to well-functioning democracies, yet physical, economic, and social barriers often exclude certain groups, reducing representativeness and contributing to issues like group polarization. In this work, we explore the use of large language model (LLM) personas to introduce missing perspectives in policy deliberations. We develop and evaluate a tool that transcribes conversations in real-time and simulates input from relevant but absent stakeholders. We deploy this tool in a 19-person student citizens' assembly on campus sustainability. Participants and facilitators found that the tool sparked new discussions and surfaced valuable perspectives they had not previously considered. However, they also noted that AI-generated responses were sometimes overly general. They raised concerns about overreliance on AI for perspective-taking. Our findings highlight both the promise and potential risks of using LLMs to raise missing points of view in group deliberation settings.

On the Relationship between Truth and Political Bias in Language Models

Language model alignment research often attempts to ensure that models are not only helpful and harmless, but also truthful and unbiased. However, optimizing these objectives simultaneously can obscure how improving one aspect might impact the others. In this work, we focus on analyzing the relationship between two concepts essential in both language model alignment and political science: \textit{truthfulness} and \textit{political bias}. We train reward models on various popular truthfulness datasets and subsequently evaluate their political bias. Our findings reveal that optimizing reward models for truthfulness on these datasets tends to result in a left-leaning political bias. We also find that existing open-source reward models (i.e. those trained on standard human preference datasets) already show a similar bias and that the bias is larger for larger models. These results raise important questions about both the datasets used to represent truthfulness and what language models capture about the relationship between truth and politics.

ConGraT: Self-Supervised Contrastive Pretraining for Joint Graph and Text Embeddings

We propose ConGraT(Contrastive Graph-Text pretraining), a general, self-supervised method for jointly learning separate representations of texts and nodes in a parent (or ``supervening'') graph, where each text is associated with one of the nodes. Datasets fitting this paradigm are common, from social media (users and posts), to citation networks over articles, to link graphs over web pages. We expand on prior work by providing a general, self-supervised, joint pretraining method, one which does not depend on particular dataset structure or a specific task. Our method uses two separate encoders for graph nodes and texts, which are trained to align their representations within a common latent space. Training uses a batch-wise contrastive learning objective inspired by prior work on joint text and image encoding. As graphs are more structured objects than images, we also extend the training objective to incorporate information about node similarity and plausible next guesses in matching nodes and texts. Experiments on various datasets reveal that ConGraT outperforms strong baselines on various downstream tasks, including node and text category classification and link prediction. Code and certain datasets are available at https://github.com/wwbrannon/congrat.