Research Areas

Aligned Information Evaluation

We design scoring mechanisms to evaluate information quality—like peer reviews or LLM outputs—when no ground truth exists. Our work balances two goals: accuracy and strategic robustness (preventing agents from gaming the system). Using the peer-prediction paradigm, we've developed provably robust mechanisms that score text based on mutual information between reports, successfully differentiating quality levels while resisting manipulation. We've also created GRE-bench, a strategically robust benchmark for evaluating LLMs on peer review. A unifying insight: rewarding agents based on mutual information between their reports makes truth-telling a dominant strategy, providing theoretical foundations for techniques like co-training and enabling applications from peer grading to crowdsourced data quality.

Recommendation and Temptation

We analyze recommendation systems from a social good perspective, recognizing that users have dual selves: seeking both long-term enrichment and instant gratification. Traditional systems based on revealed preferences generate recommendations users click on but don't provide lasting satisfaction. We've developed a user model accounting for this dual-self behavior and an optimal recommendation strategy that maximizes enrichment while navigating a key tension: showing options is necessary for learning preferences but may harm users by exposing them to tempting, unenriching content. Current work includes human-subjects studies and exploring agents that actively filter content on users' behalf.

Making Decisions with Diverse Preferences and Information

We address collective decision-making when agents have different information and objectives. When deliberation isn't feasible due to strategic considerations, scale, or sensitivity (the "3 s's"), our mechanisms aim to reach decisions agents would make with complete information sharing. We've shown that majority voting's strong equilibria select the fully-informed choice, and designed the Wisdom of the Crowd Voting mechanism using ML to achieve this with simpler strategies. Current work examines Twitter/X's Community Notes, providing theoretical foundations for its biased matrix factorization approach and clarifying how algorithmic choices shape outcomes when crowdsourcing misinformation context.

AI and Social Decision Making

These research areas form a cohesive framework for LLM development and deployment. Our decision-making work addresses pluralistic alignment—how to align AI with diverse human values when people disagree about good outputs. Our aligned information evaluation mechanisms tackle a core bottleneck: evaluating subjective model outputs where human raters disagree or models exploit evaluation metrics. Our recommendation research raises fundamental questions about who decides what is "good"—the user or the platform. Together, these approaches provide theoretical foundations and practical tools for building LLMs that navigate human values, strategic behavior, and conflicting objectives.