Imagine trying to simulate a town hall meeting using Large Language Models (LLMs). You assign one agent the persona of a frustrated 70-year-old retiree, another a progressive college student, and a third a cautious local business owner. You expect a vibrant, messy debate full of conflicting values, idiosyncratic biases, and unpredictable human friction.
Instead, you get a remarkably polite, highly rational consensus. The agents agree on almost everything, sounding less like distinct individuals and more like a single, committee-approved average.
This phenomenon is known as behavioral homogenization or the average persona bias. Across a rapidly growing body of literature, researchers are discovering that despite their massive training sets and surface-level fluency, LLMs struggle to maintain the variance, noise, and specific cultural perspectives of real human populations. When prompted to act like diverse individuals, they consistently collapse back toward a homogenized statistical mean.
The Statistical Gravity of Alignment
Why do models collapse to the mean? It is a direct consequence of how they are trained and aligned.
At their core, base models are trained on next-token prediction over massive web corpora. When safety and alignment tuning—such as Reinforcement Learning from Human Feedback (RLHF)—are layered on top, this variance is squeezed even further. RLHF explicitly penalizes controversial, extreme, or highly unusual outputs, steering the model toward a safe, polite, and homogenized baseline.
As Run Peng et al. explain in LLM-Based Social Simulations Require a Boundary, training objectives that reward high-probability responses naturally create a powerful central tendency. This systematic reduction in behavioral variance means that the model defaults to an "average persona" that lacks the behavioral diversity found in human populations. This rigidity is reinforced by standard alignment techniques, which Nigel Collier et al. note in iNews: A Multimodal Dataset for Modeling Personalized Affective Responses to News can actively suppress a model's ability to represent diverse, subjective human perspectives.
This manifests as a "mode-seeking" behavior. In The Persona Fidelity Gap: Behaviorally Grounded LLM Personas Still Compress Real-User Preference Diversity, Rishav Kumar et al. demonstrate that even when LLM personas are explicitly grounded in real-world user histories, they produce regularized, safe personas that cluster tightly together. The authors find that these models capture less than 10% of the preference diversity found in actual human populations, completely missing the "tails" of the distribution.
The Failure of Sociodemographic Steering
If the default model is an average, can we simply steer it using explicit system prompts like "Act as a 70-year-old"? The evidence suggests that simple identity-based prompting acts as a superficial mask rather than a deep cognitive shift.
When instructed to adopt a specific demographic identity, models often resort to surface-level token manipulation. For instance, Jinsook Lee et al. found in Poor Alignment and Steerability of Large Language Models: Evidence Using 30,000 College Admissions Essays that instead of adopting nuanced cultural voices, models mechanically insert explicit demographic markers verbatim, producing rigid and formulaic prose.
More fundamentally, steering fails to alter the model's underlying value system. In The Generation Gap: Exploring Age Bias in the Value Systems of Large Language Models, Siyang Liu et al. demonstrated that prompting a model to adopt an older persona fails to shift its ethical stances; its reasoning remains anchored to its default, youthful training distribution.
[Standard Prompting] ──> Mimics surface tokens & clichés ──> Retains default training bias
[True Human Persona] ──> Deep cultural/value shifts ──> High variance & nuance
This limitation is quantified by Shibani Santurkar et al. in Whose Opinions Do Language Models Reflect?, where the authors show that demographic steering only shifts model opinions by a constant factor, failing to mirror the complex, nuanced distribution of real public opinion. In fact, Huaman Sun et al. reveal in Sociodemographic Prompting is Not Yet an Effective Approach for Simulating Subjective Judgments with LLMs that forcing these identities can paradoxically increase prediction errors and refusal rates, because the models lack the underlying internal representations to faithfully emulate these diverse viewpoints.
The Rationality Bias and the Erasure of Human Noise
Human behavior is defined not just by what we think, but by our inconsistencies, cognitive biases, and emotional noise. LLMs, by contrast, suffer from an intense rationality and helpfulness bias.
In multi-agent strategic environments, this bias destroys behavioral diversity. In ALYMPICS: LLM Agents Meet Game Theory, Shaoguang Mao et al. observed that LLM agents frequently default to a baseline of rational, utility-maximizing behavior, completely ignoring the irrational or suboptimal cognitive personas they were assigned. Over extended interactions, these distinct personas erode, collapsing back into a standardized, helpful baseline.
This lack of "idiosyncratic noise" is a recurring theme. Marios Papachristou evaluated multi-LLM social networks in Network Formation and Dynamics Among Multi-LLMs and found that while synthetic agents can replicate macro-level trends (like "the rich get richer"), they exhibit an unnaturally high degree of internal correlation. They lack the unpredictable, messy noise characteristic of individual humans. This is echoed by Olufunke Oluyemi Sarumi in The Impact of Annotator Personas on LLM Behavior Across the Perspectivism Spectrum, who notes that LLM annotators default to prototypical viewpoints, yielding an unnaturally high inter-annotator agreement that erases real-world human disagreement.
Furthermore, there is a deep disconnect between how a persona speaks and how it acts. Rafal Kocielnik et al. explored this in Rethinking Psychometric Evaluation of LLMs: When and Why Self-Reports Predict Behavior, uncovering a massive "self-report vs. behavior" gap. A model might successfully fill out a personality survey mimicking a specific trait, but its actual functional behavior in deployment fails to reflect that identity unless the survey context is actively maintained in its immediate context window.
Why This Limits Social and Collective Simulations
If LLMs primarily simulate the reasoning of an aggregate average user—as documented by Zekai Shao 0001 et al. in Do Language Model Agents Align with Humans in Rating Visualizations? An Empirical Study—they become fundamentally limited tools for fields like sociology, economics, and collective cognitive science.
Many of the most important social phenomena—such as political tipping points, cultural revolutions, market crashes, or the spread of radical ideas—depend entirely on outliers, rebels, and heterogeneous agents. As Ilia Sucholutsky et al. point out in Using LLMs to Advance the Cognitive Science of Collectives, the tendency of LLMs to gravitate toward polite norms and statistical averages suppresses the individual variance necessary to study true collective dynamics.
When variance is compressed, population simulations become overly idealized, stable, and predictable. As Qirui Mi et al. argue in MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework, standard high-capacity models produce uniform responses that lack the diversity found in human groups, limiting our ability to simulate complex, heterogeneous social dynamics accurately.
To build truly pluralistic AI or reliable social simulations, researchers must move beyond surface-level prompting and address the root cause: the mathematical homogenization built into current alignment and training paradigms.
Key Foundational Work
Whose Opinions Do Language Models Reflect? (2023) — contributed 2 concept elements to this direction.
Network Formation and Dynamics Among Multi-LLMs (2024) — contributed 2 concept elements to this direction.
The Generation Gap: Exploring Age Bias in the Value Systems of Large Language Models (2024) — contributed 1 concept elements to this direction.