[CS] Mourad Heddaya Candidacy Exam/Jan 13, 2026

[via cs]

This is an announcement of Mourad Heddaya's Candidacy Exam.
===============================================
Candidate: Mourad Heddaya

Date: Tuesday, January 13, 2026

Time:  1:30 pm CST

Remote Location: https://urldefense.com/v3/__https://uchicago.zoom.us/j/94455125892?pwd=lzfqx6EB7DJSxwYttakaPhwv0TJVN0.1__;!!BpyFHLRN4TMTrA!9e_oUaaMViAx5LqRBcuPc04PY7mum58M-8DGwe_ozHnRMGvOwUEKI97iKREmjr8_GW-8OJRSGlVjKSuG2q5qPA$

Location: DSI 105

Title: Modeling Language That Shapes Decisions

Abstract: Language shapes how people understand information and make consequential decisions. Studying these effects at scale, and building AI that supports good decisions, requires getting language models to interpret and generate language in ways that capture what matters for each task and domain. This work addresses both. To study language effects at scale, it develops a framework for extracting causal micro-narratives and applies it to a century of U.S. newspaper coverage (4.2 million sentences). Narrative framing predicts inflation expectations: social and political narratives predict dispersion 1.8 times more strongly than realized inflation, and lower-income households show 4.6 times greater sensitivity. In negotiation, this work uses language models to classify bargaining strategies and identify features associated with outcomes, showing that natural language enables cooperation, producing faster agreements with less variance than numeric exchange alone. This work also evaluates language models' ability to serve as productive tools in consequential domains. In negotiation, LLMs acting as agents anchor at extreme positions and ignore contextual cues, a pattern invariant to model improvements. In legal summarization, LLM-based evaluators fail to predict what experts value. To support voters in deliberative democracy, an interactive AI chatbot drives engagement and more critical evaluation than static summaries. The result is conceptual and practical methods for getting language models to adopt appropriate perspectives, both for studying how language shapes decisions and for building AI systems that support them.

Advisors: Chenhao Tan

Committee Members: Chenhao Tan, Ari Holtzman, and Alexander Zentefis