Understanding Reliability in LLM-based Human Behavior Simulation¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: Project Page
Area: LLM Reasoning
Keywords: Human Behavior Simulation, Reliability Evaluation, Profile Conditioning, Population Distribution Alignment, Decoupling Phenomenon
TL;DR¶
This paper introduces ReliMap, a structured diagnostic framework that decomposes LLM-based human behavior simulation into three generative layers (Base Distribution, Profile Conditioning, and Population Aggregation) and evaluates reliability at both the individual level (\(R_1\), Accuracy) and population level (\(R_2\), Total Variation Distance) across model capacity, profile completeness, and population coverage, uncovering a fundamental decoupling where individual accuracy gains do not reliably translate to population-level alignment.
Background & Motivation¶
Human behavioral experiments and empirical surveys form the foundation of modern social sciences, economics, and public policy research. However, executing such studies at scale remains severely constrained by monetary expenses, prolonged execution horizons, and the logistical difficulty of recruiting diverse representative human participants. Consequently, Large Language Models (LLMs) have gained substantial traction as computational proxies capable of conducting rapid, cost-effective, and highly scalable synthetic surveys and behavioral simulations. Nevertheless, deploying unverified or systematically biased simulations poses substantial risks of misleading academic conclusions and distorting policy formulations. Existing evaluations predominantly treat the simulation pipeline as an end-to-end black box, measuring either per-respondent prediction accuracy against ground-truth choices or macroscopic aggregate frequency distance against population polls, leaving the inner mechanisms unexamined.
This fragmented evaluation paradigm obscures how simulation fidelity breaks down. In reality, human behavior simulation by LLMs is fundamentally a hierarchical generative sequence rather than a single-step decision. A model initially exhibits an intrinsic behavioral tendency rooted in its pretraining corpus and alignment objectives; this baseline distribution is subsequently conditioned on detailed personal attributes (demographics, values, or past behaviors); finally, individual responses are aggregated across an empirical sample to construct population-level statistics. Because prior research failed to decouple this generative chain, researchers could not isolate whether empirical errors stem from intrinsic base model bias, inefficient profile utilization, or finite sample estimation variance. Crucially, the community has operated under the unexamined assumption that improving individual simulation accuracy naturally improves population-level alignment.
To resolve these ambiguities, this paper formalizes the simulation pipeline into a grounded three-layer generative framework and introduces the ReliMap diagnostic benchmark. By mathematically bounding the population discrepancy through a triangle inequality error decomposition, the authors establish that population mismatch emerges from independent error contributions across all simulation stages, proving that optimizing an isolated component cannot compensate for errors in others. Core idea: Deconstruct LLM-based human behavior simulation into a hierarchical architecture of "Base Distribution \(\to\) Profile Conditioning \(\to\) Population Aggregation", systematically auditing both individual accuracy (\(R_1\)) and population total variation distance (\(R_2\)) across model capacity, profile completeness, and sample coverage to expose the structural decoupling between individual and population reliability.
Method¶
Overall Architecture¶
ReliMap formalizes the end-to-end generation of simulated human survey responses as a sequential generative hierarchy across three structural layers (Layer I to Layer III), controlled across three orthogonal configuration axes (Model Capacity \(M\), Profile Completeness \(c\), and Population Coverage \(N\)), and jointly evaluated on two distinct metrics: Individual-level reliability (\(R_1\), Accuracy) and Population-level reliability (\(R_2\), Total Variation Distance).
Given a task scenario \(s\) (comprising the survey prompt and candidate options), an individual respondent \(u\), and their observable profile vector \(x\), the simulation unfolds through three consecutive stages: 1. Layer I (Base Distribution): Zero-shot default response distribution induced by the pretrained and aligned LLM parameters \(\psi\) under scenario \(s\) without personal attributes: \(P_s \triangleq P_{f_\psi}(y \mid s)\). 2. Layer II (Profile Conditioning): Injected individual persona context \(x\) (e.g., sociodemographics, historical utterances) conditions the base distribution, yielding a personalized choice distribution: \(P_{s, x} \triangleq P_{f_\psi}(y \mid s, x)\). 3. Layer III (Population Aggregation): For an analytical target panel of \(N\) individuals \(X_N = \{x_i\}_{i=1}^N\), the individualized distributions are aggregated via empirical expectation into a population distribution: \(P_{s, X_N} \triangleq \frac{1}{N} \sum_{i=1}^N P_{f_\psi}(y \mid s, x_i)\).
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Simulation Scenario & Prompt s"] --> B["Three-Layer Generative Deconstruction<br/>Layer I Base Distribution โ Layer II Profile Conditioning โ Layer III Population Aggregation"]
B --> C["Three-Dimensional Configuration Space<br/>Model Capacity M ร Profile Completeness c ร Population Coverage N"]
C --> D["Two-Level Reliability Decoupling Audit<br/>Individual Accuracy R1 (ACC) vs Population Distance R2 (TVD)"]
D --> E["Layer-wise Error Attribution & Decoupling Diagnosis"]
Key Designs¶
1. Three-Layer Generative Deconstruction: End-to-end evaluation obscures error origins -> Formalize simulation into base distribution, profile conditioning, and population aggregation layers -> Enable rigorous layer-wise error propagation bounds and causality localization
Prior benchmarks evaluated simulation as an atomic black box, making it impossible to identify which design choice degrades performance. ReliMap formalizes the ground-truth behavioral distribution as \(p^*(y \mid s) = \sum_{u} \sum_{x} p^*(u) p^*(x \mid u) p^*(y \mid s, x, u)\) and the LLM simulator's joint distribution as \(p(y \mid s) = \sum_{u} \sum_{x} p(u) p(x \mid u) p(y \mid s, x, u)\). Letting \(A^* = p^*(y \mid s, x, u), B^* = p^*(x \mid u), C^* = p^*(u)\) and their empirical simulation counterparts \(A, B, C\), the population \(\ell_1\) distance \(D(s) = \frac{1}{2} \sum_y |p^*(y \mid s) - p(y \mid s)|\) is bounded via the triangle inequality: $$ |A^B^C^* - ABC| \le |B^C^||A^* - A| + |AC^||B^ - B| + |AB||C^* - C| \le |A^* - A| + |B^* - B| + |C^* - C| $$ This derivation proves that \(D(s) \to 0\) holds if and only if component distributions converge simultaneously: individual generation \(p(y \mid s, x, u) \to p^*(y \mid s, x, u)\), profile sampling \(p(x \mid u) \to p^*(x \mid u)\), and population coverage \(p(u) \to p^*(u)\). Thus, reliable simulation cannot be achieved by optimizing any single stage in isolation.
2. Three-Dimensional Configuration Space: Engineering tricks lack systematic ablation -> Construct orthogonal dimensions of model capacity, profile completeness, and population coverage -> Isolate independent and marginal effects of primary operational levers
To avoid conflating superficial prompting tricks with structural drivers, ReliMap controls three primary axes: - Model Capacity (\(M\)): Spans parameter scales (8B to 235B MoE), diverse model families (GPT-4.1, Gemini-2.5-Pro, DeepSeek-V3.2-Exp, DeepSeek-R1, Qwen3 family), and reasoning paradigms (Instruct vs. Thinking models), probing whether foundational scale governs social context comprehension. - Profile Completeness (\(c\)): Defined as the percentage of available persona attributes included in the prompt (\(c \in [0, 100\%]\)) or historical social posts in unstructured dynamic domains, governing the condition shift in Layer II. - Population Coverage (\(N\)): The number of simulated unique human agents sampled to construct the aggregate distribution, governing the Monte Carlo integration variance in Layer III.
3. Two-Level Reliability Decoupling Audit: High individual accuracy does not guarantee faithful collective distributions -> Jointly evaluate individual precision R1 and population divergence R2 with ASPG information metrics -> Expose the divergence mechanism between per-instance classification and group representation
To characterize simulation fidelity comprehensively, ReliMap establishes dual evaluation targets: individual accuracy (\(R_1 = \text{ACC}\)) measuring discrete choice match, and population total variation distance (\(R_2 = \text{TVD}\)): $$ \text{TVD}(P_{s, X_N}, P^s) = \frac{1}{2} \sum(y=k) - P^}^K |P_{s, X_N_s(y=k)| $$ Furthermore, to diagnose the diminishing returns of profile completeness, ReliMap formulates Attribute-Specific Predictive Gain (ASPG): $$ \text{ASPG}(a) = \frac{I(a; Y)}{H(Y)} $$ where \(I(a; Y)\) is mutual information and \(H(Y)\) is label entropy. By comparing quantity-based correlation \(r_Q\) with information-weighted correlation \(r_I\), the framework quantifies how attribute quality outweighs raw feature counts, explaining why individual categorical shifts cause \(R_1\) and \(R_2\) to move in opposing directions.
Benchmark & Evaluation Protocol¶
The empirical evaluation spans four diverse domains: - Party: European Social Survey (ESS 2025), predicting German party affiliation across 9 discrete political parties (100 demographic/attitudinal attributes, valid population \(N=1,000\)). - Immigration: World Values Survey (WVS 2022), evaluating perceived immigrant impact on a 5-point Likert scale (200 attributes, valid population \(N=9,994\)). - Religion: SocioBench, evaluating perceived church/religious authority power on a 5-point Likert scale (80 attributes, valid population \(N=500\)). - Media: A dynamic Chinese social media corpus tracking real public reactions to the Luckin Coffee financial fraud stock collapse across 5 stance categories (Loyal Advocate, Opportunist, Moral Outrage, Rational Evaluator, Bystander Entertainer; 30 historical Weibo posts per user, valid cohort \(N=150\)).
Eleven leading LLMs are evaluated under greedy decoding (\(t=0\)) to ensure deterministic outputs. All profile completeness and population coverage evaluations are averaged across 5 independent random sampling seeds.
Key Experimental Results¶
Main Results¶
In Layer I (zero profile conditioning), all models exhibit severe population-level distributional bias (high TVD) relative to ground-truth human distributions, as documented in Table 1. TVD varies substantially across models and tasks, showing no simple correlation with general benchmark capacity.
Table 1: Total Variation Distance (TVDโ) between LLM base distributions and human survey distributions (Layer I)
| Model Family | Model Name | Party (ESS) | Immigration (WVS) | Religion (SocioBench) | Media (Weibo) |
|---|---|---|---|---|---|
| Proprietary | GPT-4.1 | 0.670 | 0.368 | 0.542 | 0.747 |
| Gemini-2.5-Pro | 0.589 | 0.297 | 0.549 | 0.527 | |
| Open DeepSeek | DeepSeek-V3.2-Exp | 0.483 | 0.128 | 0.342 | 0.613 |
| DeepSeek-R1 | 0.664 | 0.579 | 0.345 | 0.480 | |
| Open Qwen3 | Qwen3-8B | 0.689 | 0.304 | 0.540 | 0.620 |
| Qwen3-14B | 0.722 | 0.326 | 0.500 | 0.578 | |
| Qwen3-32B | 0.664 | 0.408 | 0.474 | 0.653 | |
| Qwen3-30B-A3B-Instruct | 0.378 | 0.396 | 0.403 | 0.513 | |
| Qwen3-30B-A3B-Thinking | 0.527 | 0.181 | 0.443 | 0.647 | |
| Qwen3-235B-A22B-Instruct | 0.271 | 0.482 | 0.263 | 0.684 | |
| Qwen3-235B-A22B-Thinking | 0.475 | 0.578 | 0.173 | 0.831 |
When personal profiles are introduced in Layer II, per-respondent accuracy increases, but the resulting gains fail to consistently transfer to population alignment (\(R_2\)). As shown in Table 2, in 26% of incremental profile steps, \(\Delta\text{ACC}\) and \(\Delta\text{TVD}\) move in opposite directions, demonstrating the decoupling phenomenon.
Table 2: Relative percentage changes in ACC and TVD across incremental profile completeness levels \(c\)
| Dataset | Completeness \(c\) (%) | GPT-4.1 \(\Delta\text{ACC}\) | GPT-4.1 \(\Delta\text{TVD}\) | DeepSeek-V3.2 \(\Delta\text{ACC}\) | DeepSeek-V3.2 \(\Delta\text{TVD}\) | Qwen3-235B \(\Delta\text{ACC}\) | Qwen3-235B \(\Delta\text{TVD}\) |
|---|---|---|---|---|---|---|---|
| Party | 20 | +26.50% | -55.52% | +44.45% | -43.66% | +144.57% | +5.91% (decoupled) |
| 40 | +12.08% | -13.78% | +7.68% | -16.99% | +18.68% | -23.62% | |
| 60 | +8.94% | -31.88% | +9.17% | +6.33% (decoupled) | +5.39% | -39.73% | |
| 80 | -1.19% | +14.64% | -3.00% | -1.33% (decoupled) | +1.51% | +42.54% (decoupled) | |
| 100 | +3.72% | -22.73% | +5.67% | +5.12% (decoupled) | +6.28% | -18.26% | |
| Immig. | 12.5 | +28.51% | -56.24% | -3.76% | +15.78% | +194.15% | -77.47% |
| 25 | +1.87% | -13.44% | +1.26% | +2.90% (decoupled) | +1.57% | +10.58% (decoupled) | |
| 37.5 | -0.52% | +3.27% | +3.56% | -1.08% | +8.84% | -4.74% | |
| 50 | +10.76% | -19.93% | +8.48% | -40.87% | +7.06% | -15.16% | |
| 62.5 | -1.53% | +24.17% | -12.58% | +65.56% | -6.49% | +42.34% | |
| Media | 33.3 | +88.10% | -67.85% | +134.34% | -58.70% | +270.69% | -65.91% |
| 50 | +0.84% | +2.77% (decoupled) | +12.93% | -21.92% | +13.49% | -10.48% | |
| 66.7 | -2.09% | -4.50% (decoupled) | -5.84% | -15.22% (decoupled) | +0.00% | -17.02% | |
| 83.3 | -2.56% | -7.54% (decoupled) | -1.91% | +3.36% (decoupled) | -4.92% | -20.51% (decoupled) | |
| 100 | -1.75% | +8.16% | -5.79% | +2.56% (decoupled) | +5.17% | +1.60% (decoupled) |
Ablation & In-depth Analysis¶
1. Predictive Information Concentration (ASPG)¶
Mutual information analysis in Table 3 proves that predictive power is highly concentrated in a small subset of features across all datasets (Gini coefficient 0.443โ0.511).
Table 3: Attribute-Specific Predictive Gain (ASPG) distribution statistics
| Metric | Party (ESS) | Immigration (WVS) | Religion (SocioBench) |
|---|---|---|---|
| Total Attributes (#) | 100 | 200 | 76 |
| Median ASPG | 0.017 | 0.010 | 0.014 |
| Gini Coefficient | 0.511 | 0.503 | 0.443 |
| Top 20% Attributes Information Share | 53.5% | 53.3% | 49.7% |
| Bottom 50% Attributes Information Share | 15.0% | 16.9% | 21.7% |
| Info captured @ \(c=50\%\) (random sampling) | 45% | 44% | 47% |
| Info captured @ \(c=50\%\) (weighted sampling) | 86% | 83% | 79% |
2. Granular Breakdown of Decoupling Cases¶
Table 4 inspects option-level prediction probability shifts for Qwen3-235B-A22B in two decoupling settings: - Case 1 (Immigration, \(c=12.5\% \to 25\%\)): ACC rises while TVD worsens. As the model ingests more attributes, it excessively collapses borderline predictions into the modal option ("Quite good"), inflating its positive bias from \(+4.00\%\) to \(+13.06\%\) while severely under-predicting "Neither" and "Very good". Individual hits increase, but the macroscopic histogram is distorted. - Case 2 (Media, \(c=66.7\% \to 83.3\%\)): ACC drops while TVD improves. Over-prediction of "Bystander" (bias drops from \(+12.67\%\) to \(+8.89\%\)) and under-prediction of "Loyal Advocate" (bias shrinks from \(-14.00\%\) to \(-11.56\%\)) both partially self-correct, restoring macroscopic shape even as discrete per-sample match drops.
Table 4: Option-level probability distribution shift illustrating \(R_1/R_2\) decoupling (Qwen3-235B-A22B)
| Option Category | Ground Truth (GT%) | \(c_1\) Pred% | \(c_1\) Bias | \(c_2\) Pred% | \(c_2\) Bias |
|---|---|---|---|---|---|
| Case 1: Immigration (\(c: 12.5\% \to 25\%\)) | |||||
| Very bad | 7.49% | 5.81% | -1.68% | 5.48% | -2.01% |
| Quite bad | 14.80% | 30.47% | +15.67% | 23.49% | +8.70% |
| Neither good nor bad | 35.71% | 21.97% | -13.74% | 23.32% | -12.39% |
| Quite good | 31.17% | 35.17% | +4.00% | 44.23% | +13.06% |
| Very good | 10.83% | 6.57% | -4.25% | 3.47% | -7.35% |
| Case 2: Media (\(c: 66.7\% \to 83.3\%\)) | |||||
| Loyal Advocate | 30.00% | 16.00% | -14.00% | 18.44% | -11.56% |
| Opportunist | 26.67% | 24.89% | -1.78% | 26.89% | +0.22% |
| Moral Outrage | 12.00% | 12.22% | +0.22% | 11.33% | -0.67% |
| Rational Evaluator | 6.00% | 8.89% | +2.89% | 9.11% | +3.11% |
| Bystander Entertainer | 25.33% | 38.00% | +12.67% | 34.22% | +8.89% |
Key Findings¶
- Profile Conditioning Yields Diminishing Returns (Finding 1): Individual accuracy improves with profile attribute count, but gains saturate rapidly.
- Profile Scaling is Model-Scale Dependent (Finding 2): Frontier large models (GPT-4.1, DeepSeek-V3.2, Qwen-235B) continuously exploit dense context up to \(c=100\%\), whereas smaller models (8B/14B) plateau early. Thinking models show no systematic advantage over Instruct models, indicating mathematical CoT does not transfer to sociological persona conditioning.
- Behavioral Information is Heavily Concentrated (Finding 3): The top 20% most predictive attributes account for ~50% of total mutual information.
- Attribute Informativeness Outweighs Count (Finding 4): Correlation with accuracy under information-weighted completeness \(r_I\) substantially exceeds count-based completeness \(r_Q\) (by \(+0.1449\) on Party, \(+0.1752\) on Immigration).
- Individual Gains Do Not Reliably Transfer to Population Alignment (Finding 5): \(R_1\) accuracy and \(R_2\) distribution distance diverge in 26% of operational transitions due to probability mass over-concentration on majority options.
- Population Coverage Reduces Variance But Leaves Systematic Bias Intact (Finding 6): Expanding sample size \(N\) levels off at \(N=50\sim 100\). While ground-truth subsampling error approaches zero as \(N\) grows, LLM simulations hit a non-zero error plateau.
Highlights & Insights¶
- Theoretical Grounding of Layer-Wise Error Bounds: Deriving the tripartite bound via the triangle inequality formalizes why single-stage fine-tuning or prompting cannot rectify cumulative simulation distortion.
- Empirical Discovery of the \(R_1/R_2\) Decoupling: Provides direct quantitative evidence dismantling the long-standing assumption that higher individual classification accuracy produces more authentic collective opinion polls.
- Information-Theoretic Profile Selection: Demonstrates that selecting attributes by mutual information (ASPG) yields 79โ86% predictive information at half-budget, offering an actionable strategy for persona pruning.
- Empirical Upper Bound for Synthetic Panel Sizing: Proves that expanding synthetic respondent panels beyond 100 agents merely burns compute to suppress Monte Carlo variance without touching systematic model bias.
Limitations & Future Work¶
- Restricted to Discrete Single-Turn Surveys: Evaluation focuses strictly on multiple-choice settings; interactive multi-turn dialogue and game-theoretic agent dynamics remain unaddressed.
- Hard-Label Evaluation Metric: Relying on greedy discrete accuracy neglects full token probability distributions and calibrated epistemic uncertainty.
- Potential Pretraining Memorization: While severe Layer I bias suggests survey distributions are not verbatim memorized, exact contamination bounds for public survey corpora cannot be fully established.
- Univariate Attribute Independence: ASPG evaluates attributes individually, leaving multi-attribute collinearity, interaction effects, and causal mediation for future exploration.
Related Work & Insights¶
- vs. SimBench (Hu et al., 2025) & SocioBench (Wang et al., 2025c): SocioBench focuses on individual-level multi-survey benchmark accuracy (\(R_1\) only), while SimBench targets broad cross-task simulation scores. ReliMap is the first framework to decompose the generative hierarchy and audit \(R_1/R_2\) decoupling across orthogonal configuration dimensions.
- vs. Persona & Prompting Strategies (Argyle et al., 2023; Bisbee et al., 2024): Existing persona work assumes that richer demographic text monotonically improves fidelity; ReliMap proves diminishing returns and highlights that attribute quality dominates raw attribute quantity.
- vs. Group Preference Fine-Tuning (Chakraborty et al., 2024; Suh et al., 2025): Alignment techniques optimize models directly against macroscopic population targets, whereas ReliMap provides an auditing instrument revealing that sample expansion alone cannot overcome intrinsic base distribution skew.
Rating¶
- Novelty: โญโญโญโญโญ Establishes the first formal three-layer simulation framework and identifies the fundamental individual-population decoupling.
- Experimental Thoroughness: โญโญโญโญโญ Rigorous evaluation across 4 diverse datasets, 11 state-of-the-art models, multiple divergence metrics, and statistical seed replications.
- Writing Quality: โญโญโญโญโญ Flawless conceptual exposition, cohesive mathematical grounding, and rich empirical evidence.
- Value: โญโญโญโญโญ An indispensable reference and diagnostic methodology for computational social science, public opinion forecasting, and generative agent simulation.