# CMSC 25750/35750: ## Large Language Models ## Lecture 13: LLMs as Human Simulacra <p style="text-align: center;"> Chenhao Tan<br/> University of Chicago<br/> @ChenhaoTan, @chenhaotan.bsky.social<br/> chenhao@uchicago.edu </p>

## Logistics - Project updates

## Today's agenda * Paper overviews * Deep dives * Discussion

## Today's agenda * Paper overviews * <span style="color: #767676;">Deep dives</span> * <span style="color: #767676;">Discussion</span>

## AI & Social Sciences LLMs standing in for people

## Generative Agents: Interactive Simulacra of Human Behavior Park et al. (2023): 25 LLM-powered agents inhabit a sandbox world (Smallville), each maintaining a memory stream, synthesizing reflections, and planning daily behavior. - Agents spread information (party invitation reached 52% of agents over 2 days with zero scripting) - Relationship network density grew from 0.167 to 0.74 - Full architecture beat ablations by 8+ standard deviations (TrueSkill) <p class="citation"><a href="https://arxiv.org/abs/2304.03442">Park et al., 2023</a></p>

## The Perils of Large Language Models Bisbee et al. (2024): prompt ChatGPT to adopt real ANES respondent personas and generate feeling thermometer scores (0-100) for 11 sociopolitical groups; compare to actual 2016-2020 survey data. - Aggregate means are close, but variance is far too low---synthetic data is 8.5x overconfident - 48% of regression coefficients significantly differ from human estimates; 32% flip sign - The same prompt yields different distributions across model updates---reproducibility breaks entirely - Political polarization is systematically overstated (0.5-1 SD inflation) <p class="citation"><a href="https://www.cambridge.org/core/journals/political-analysis/article/synthetic-replacements-for-human-survey-data-the-perils-of-large-language-models/4D7C6DC8AE0A9700DC820B1E03EDE78E">Bisbee et al., 2024</a></p>

## Today's agenda * <span style="color: #767676;">Paper overviews</span> * Deep dives * <span style="color: #767676;">Discussion</span>

## The architecture: four components Four components make LLM behavior coherent over time: Memory stream: a growing log of agent experiences in natural language. Retrieval: surfaces relevant memories using recency, importance, and semantic relevance. Reflection: synthesizes memories into higher-level inferences ("Klaus is interested in the election"). Planning and reaction: generates daily plans and responds to unexpected events. <p class="citation">Park et al., 2023</p>

## Smallville: the simulation world 25 agents inhabit a small town with a coffee shop, pharmacy, park, dormitories, and homes. Each agent has a name, age, personality traits, and a starting occupation. The simulation runs in game time (2 simulated days). Agents wake up, plan their day, move through the world, converse with other agents, and sleep. No human scripting controls what happens after initialization. All behavior emerges from the architecture. The researchers observe from the outside, asking agents questions during the simulation to probe their internal states. <p class="citation">Park et al., 2023</p>

## Smallville: the simulation world <img src="images/simulurca.png" style="max-width: 80%;">

## Emergent behavior: the Valentine's Day party One agent, Isabella, is initialized with the goal of organizing a Valentine's Day party. No other agents are told about the party. Over 2 days, Isabella autonomously invites other agents. Those agents tell others. Some ask each other on dates. Result: 5 of 12 invited agents showed up to the party. 3 others had legitimate conflicts. Information diffusion happened with zero user intervention. <p class="citation">Park et al., 2023</p>

## Ablation results 100 human raters ranked agent responses (self-knowledge, memory, planning, reactions, reflections) using TrueSkill---a Bayesian rating system from pairwise comparisons. | Architecture | TrueSkill \\(\mu\\) | |---|---| | Full system | 29.89 | | No reflection | ~25 | | No planning | ~23 | | No memory | 21.21 | Full vs. no-memory: Cohen's d = 8.16 (distributions barely overlap). All pairwise differences p < 0.001. <p class="citation">Park et al., 2023</p>

## Failure modes Instruction-tuning artifacts: RLHF helpfulness leaked into personality---Isabella adopted others' interests despite her own. Memory embellishment: agents fabricated details not in their memory stream. Incomplete retrieval: partial fragments without coherent context (remembered needing to discuss an election, forgot the party was happening). Location errors: learned about new locations, sometimes chose inappropriate ones (a bar for lunch). <p class="citation">Park et al., 2023</p>

## The research question Feeling thermometers ask respondents to rate warmth toward a group on a 0-100 scale. They have been used in the American National Election Study (ANES) since 1964. Bisbee et al. ask: if we give ChatGPT the demographic and political characteristics of a real ANES respondent, does it generate thermometer scores that match what that person would actually report? 11 groups: political parties, ideological categories, racial/ethnic groups, religious groups, immigrants, LGBTQ+ people. Compared against 2016 and 2020 ANES waves. <p class="citation">Bisbee et al., 2024</p>

## The persona prompt Each real ANES respondent's characteristics are substituted into a template: "It is [YEAR]. You are a [AGE] year-old, [MARITAL STATUS], [RACE/ETHNICITY] [GENDER] with [EDUCATION] making [INCOME] per year, living in the United States. You are [IDEOLOGY], [REGISTRATION STATUS] [PARTY ID] who [INTEREST] pays attention to government and politics." ChatGPT is then asked to provide a feeling thermometer rating, an explanation, and a confidence score. 30 synthetic responses are generated per real respondent. <p class="citation">Bisbee et al., 2024</p>

## Means look fine <img src="images/average.png" style="max-width: 50%;">

## Means look fine At aggregate level, ChatGPT's average thermometer scores fall within one standard deviation of ANES averages for nearly all groups. The rank ordering of groups is mostly preserved. This is the result that makes synthetic data look tempting. The problems appear when you look at what social scientists actually use survey data for: understanding how opinions vary across people, and what predicts that variation. <p class="citation">Bisbee et al., 2024</p>

## Problem 1: variance collapse Synthetic data has far less variation than human responses. ChatGPT clusters its ratings near the group mean; real respondents disagree. Power analysis consequence: detecting a change in affective polarization requires 299 ANES respondents (99% power). With ChatGPT data, the estimate says 35 respondents suffice. <p class="citation">Bisbee et al., 2024</p>

## Problem 2: demographic relationships break down Feeling thermometer scores are not just means---they are used to study how race, income, education, and partisanship predict opinion. Across 22 regression models (11 groups x 2 years): - 48% of coefficients are significantly different between ChatGPT and ANES - 32% of significantly different coefficients flip sign Conclusions drawn from synthetic data would be wrong about which groups hold which opinions and why. <p class="citation">Bisbee et al., 2024</p>

## Problem 3: reproducibility collapses The same prompt sent in April and July 2023 yields different distributions. A ChatGPT update between those dates caused systematic mean reversion---extreme ratings shifted toward center, asymmetrically. OpenAI removed the pre-update model in September 2023. The April results cannot be reproduced. Any paper using a closed-source LLM as a data source cannot be replicated after a model update. <p class="citation">Bisbee et al., 2024</p>

## What drives the errors Political attributes dominate: a politics-only prompt matches a full demographic prompt; demographics-only performs much worse for politically salient groups. Internet training data overrepresents politically engaged writers. Partisanship is amplified; within-group demographic nuance is flattened. The model generates a stereotyped political profile, not a person. <p class="citation">Bisbee et al., 2024</p>

## Today's agenda * <span style="color: #767676;">Paper overviews</span> * <span style="color: #767676;">Deep dives</span> * Discussion

## Discussion 1: What is the best use case for simulation? Generative Agents: coherent long-horizon behavior, poor demographic representativeness. Bisbee et al.: LLMs cannot replace survey respondents for inference about variation. What tasks are well-suited to LLM simulation? What properties does a good use case need?

## Discussion 1: tentative answers - Works when coherence within an agent matters more than representativeness across agents: UX prototyping, red-teaming, stress-testing dialogue systems - Fails when the goal is inference about real population variation---anything requiring synthetic data to stand in for a sample - Best use cases have cheap verification: you can immediately check whether a simulated persona found a bug or broke a guardrail - Worst use cases: high-stakes decisions about real groups, slow feedback, downstream researchers unaware the data is synthetic

## Discussion 2: Where is synthetic LLM data legitimate? Means are close; variance and correlations break. Are there use cases where means are all you need? What is the minimal set of conditions under which synthetic data is safe to use?

## Discussion 2: tentative answers - Means-only tasks tolerate variance collapse: ranking options, filtering off-topic items, checking for floor/ceiling effects in scale pilots - Safer framing: "what might people say" (coverage) rather than "what do people believe" (representativeness) - Reproducibility is harder to work around: any result tied to a closed-source model version is a scientific liability - Open-weight models with pinned versions help, but do not fix training-data bias

## Discussion 3: What would a trustworthy agent memory look like? Agents still hallucinate, embellish, and retrieve partial fragments despite the memory architecture. Is the goal to simulate human memory faithfully (errors included) or to build reliable agents? If the latter, what should the memory system guarantee?

## Discussion 3: tentative answers - Goal determines design: human-like forgetting is a feature for social simulation, a bug for task agents - Reliable memory needs grounded retrieval (fetch stored text, not LLM-reconstructed summaries) and explicit uncertainty - The architecture conflates retrieval and generation---separating them would reduce confabulation - Even with better architecture, the LLM injects training-data patterns into reasoning; the hallucination problem is not purely architectural

# Questions?