UC Berkeley

This dissertation develops a unified framework for using large language models (LLMs) as simulators of human users—both as synthetic participants in behavioral and social-scientific research and as synthetic interlocutors for the training and evaluation of AI assistant systems. The central position of this work is that user simulation is a fundamentally distinct modeling problem from assistant alignment, and that pretrained base models, which we frame as Human–Language Mixture Models (HLMMs), are the appropriate substrate for it. Trained via cross-entropy on web-scale corpora authored by hundreds of millions of distinct individuals, pretrained models internalize a “mixture of voices” that reflects the diversity of human linguistic, social, and pragmatic behavior. Instruction-tuned assistant language models (ALMs), by contrast, are further optimized under objectives that systematically suppress the breadth, ambiguity, and diverse behaviors that real users exhibit. A central claim of this dissertation is that much of the “failure” of LLM-based user simulation reported in the recent literature is in fact a failure of taking the wrong approach of instructing an ALM to role-play a user, rather than a limitation of language models in general.Building on this argument, the dissertation introduces a unifying mechanism for binding pretrained LLMs to coherent virtual personas. We propose conditioning language models with unconstrained and natural life narratives, which we refer to as backstories, generated by base models themselves. Backstories both explicitly and implicitly encodes diverse details about its author, including demographic information, psychological context, and human beliefs, values, and perspectives. When conditioned with backstories, language models are bound to a naturalistic persona of a particular individual, and throughout this work we show how backstory conditioning improves realism, representativeness, consistency, and behavioral fidelity in simulated populations than existing methods that rely on prescriptive persona design.Chapter 2 formalizes the distinction between HLMMs and ALMs and provides experimental evidence across diverse conversational corpora that pretrained base models better estimate the conditional distribution of human utterances, better reflect pragmatic dialogue structure, and better preserve lexical and semantic diversity than their instruction-tuned counterparts. We additionally introduce Tandem Models, an inference-time procedure that pairs a base-model utterance generator with an instruction-tuned supervisor.Chapter 3 introduces the Anthology framework, which generates diverse first-person backstories through simple open-ended prompting (e.g., “Tell me about yourself”) and aligns them to target demographic distributions via maximum-weight or greedy bipartite matching. When conditioned on these backstories, base LLMs reproduce population-level opinion distributions from the Pew Research Center’s American Trends Panel with smaller distributional shifts between human and model responses and higher internal consistency than existing persona-conditioning methods, including for under-represented demographic subgroups.Chapter 4 extends the Anthology framework to model social identity and group perception. We test whether backstory-conditioned LLMs are capable of deep persona binding: responding as authentic in-group members would rather than shallow imitation of out-group stereotypes. Longer, more coherent backstories generated through a minimal Tandem modeling approach—multi-turn interview prompting pretrained models and rejection sampling by an instruction-tuned LLM-as-a-critic—yield richer and more consistent virtual personas. These backstory-conditioned LLMs reproduce partisan asymmetries in moral judgment, expectations of democratic backsliding, and exaggerated meta-perceptions observed in human data, demonstrating that narrative depth and consistency are essential for capturing identity-driven perspectives.Chapter 5 then turns to applications. It applies backstory conditioning to action prediction in partisan social-dilemma games (Dictator and Trust), introducing Temporal Grounding and Consistency Filtering as supplementary prompting strategies for situated behavioral simulation, and reproduces human partisan favoritism across replications spanning more than a decade. The same setting enables a 2×2×2 counterfactual decomposition of year, framing, and recruitment-population effects that no single human experiment can supply. The chapter then frames design problems as the natural application surface for the methodology developed in this dissertation, identifies three structural challenges that distinguish AI assistance in design from instruction-following, and positions digital hardware design as a flagship use-case where faithful user simulation could be applied to building collaborative AI agents.Together, these studies establish backstory-conditioned pretrained LLMs as a principled foundation for scalable, responsible behavioral simulation and as a faithful substrate for synthetic interlocutors in agent evaluation.