Understanding Reliability in LLM-based Human Behavior Simulation¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://yupei-wang.github.io/understand_reliability/
Area: Information Retrieval
Keywords: Large Language Models, Human Behavior Simulation, Reliability Auditing, Hierarchical Generative Process, Population Distribution Alignment
TL;DR¶
ReliMap decomposes LLM-based human behavior simulation into a three-layer generative framework to audit individual and population reliability across model capacity, profile completeness, and population coverage, revealing that individual accuracy gains decouple from macro-distribution alignment and increasing sample coverage reduces variance but never systematic bias.
Background & Motivation¶
Large language models (LLMs) are increasingly leveraged across computational social science, policy research, and market analysis as scalable computational proxies for human survey respondents and behavioral agents. By conditioning generative agents on demographic personas or textual backstories, researchers seek to run rapid, inexpensive synthetic polls and behavioral experiments. However, if synthetic simulations deviate systematically from actual human responses, downstream social science findings and algorithmic recommendations can be critically misled. Existing evaluation studies remain fragmented into isolated camps: one line of research focuses almost exclusively on micro-level per-respondent prediction accuracy (e.g., via persona prompting or alignment fine-tuning), while another line evaluates only macro-level population distribution alignment (e.g., calculating statistical divergence against census surveys). Treating simulation as a monolithic end-to-end black box fails to localize whether failure modes stem from intrinsic model priors, profile conditioning inefficiencies, or aggregation sampling noise.
This fundamental diagnosis gap arises because LLM-based human behavior simulation is treated as a flat, single-step operation when it is inherently a hierarchical generative process. In reality, simulation unfolds across three consecutive tiers: first, pretraining and instruction tuning impart a default, unconditional base behavioral distribution over opinions before any persona is specified; second, conditioning this base prior on individual demographic attributes, psychometrics, and historical behaviors shifts the behavioral trajectory toward a target persona, though attribute interaction and context saturation introduce severe nonlinearities; finally, individual simulated responses are statistically aggregated across a sampled cohort to reconstruct the target population distribution. Because these stages operate sequentially, optimizing any single layer in isolation provides zero mathematical guarantee of global simulation fidelity.
To bridge this critical divide between localized evaluation metrics and systemic simulation behavior, there is an urgent need for a unified diagnostic framework capable of dissecting error propagation across layers and configuration spaces. A principled framework must theoretically bound distributional discrepancy across generative stages while empirically exploring how model scaling, persona informativeness, and respondent sample sizes govern individual and collective reliability. Core idea: Introduce ReliMap, a principled three-layer diagnostic framework that formalizes LLM behavior simulation into base distribution, profile conditioning, and population aggregation, systematically auditing reliability at both individual (R1) and population (R2) levels across model capacity, profile completeness, and population coverage to uncover fundamental decoupling phenomena and variance-bias boundaries.
Method¶
Overall Architecture¶
ReliMap formalizes LLM-based human behavior simulation as a three-layer generative hierarchy and evaluates reliability under a three-dimensional configuration space across two complementary evaluation levels. The generative hierarchy consists of: (1) Layer I: Base Distribution, capturing the model's default, unconditional response tendency \(P_s \triangleq P_{f_\psi}(y \mid s)\) under a specific scenario \(s\) without personal profile information; (2) Layer II: Profile Conditioning, which injects individual profile information \(x\) (e.g., demographic variables, ideological values, historical social media posts) into the context to induce a personalized conditional distribution \(P_{s,x} \triangleq P_{f_\psi}(y \mid s, x)\); and (3) Layer III: Population Aggregation, which aggregates the individualized predictions across a sampled cohort \(X_N = \{x_i\}_{i=1}^N\) to reconstruct the macroscopic population distribution \(P_{s,X_N} \triangleq \frac{1}{N}\sum_{i=1}^N P_{f_\psi}(y \mid s, x_i)\).
Across these generative layers, ReliMap investigates three key operational configuration dimensions: Model Capacity (\(M\)), Profile Completeness (\(c\), parameterized by the fraction of provided attributes relative to all available features), and Population Coverage (\(N\), the simulated cohort size). Reliability is audited at two distinct levels: individual-level reliability (R1), quantified via discrete prediction accuracy (ACC), and population-level reliability (R2), quantified via total variation distance (TVD) alongside KL divergence, Jensen-Shannon divergence (JSD), and Earth Mover's Distance (EMD) against ground-truth human populations.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input: Scenario Context s & Target Population Panel"] --> B["Three-Layer Generative Decomposition: Base, Profile, Aggregation"]
B --> C["Attribute-Specific Predictive Gain (ASPG) & Information Modeling"]
C --> D["Two-Level Reliability Decoupling Audit (R1 vs R2)"]
D --> E["Population Coverage Scaling & Variance-Bias Decomposition"]
E --> F["Output: ReliMap Diagnostic Audit & Simulation Strategy"]
Key Designs¶
1. Three-Layer Generative Decomposition: Base, Profile, Aggregation
To resolve the inability of end-to-end evaluation scores to localize error sources within multi-stage simulation pipelines, this design establishes a formal mathematical error-propagation bound. Let the ground-truth human population distribution under scenario \(s\) be decomposed as \(p^*(y \mid s) = \sum_u \sum_x p^*(u) p^*(x \mid u) p^*(y \mid s, x, u)\), and let the simulator induce an approximate distribution \(p(y \mid s) = \sum_u \sum_x p(u) p(x \mid u) p(y \mid s, x, u)\). By applying the triangle inequality to the global \(\ell_1\) population discrepancy \(D(s) \triangleq \|p^*(\cdot \mid s) - p(\cdot \mid s)\|_1\), the overall error is strictly upper-bounded by three orthogonal components:
where \(A^* = p^*(y \mid s, x, u)\) represents scenario-conditioned behavior generation, \(B^* = p^*(x \mid u)\) denotes persona profile extraction, and \(C^* = p^*(u)\) represents individual population sampling weights. This theoretical derivation formally proves that simulation convergence to ideal fidelity requires the coordinated convergence of individual sampling (\(p(u) \to p^*(u)\)), profile representation (\(p(x \mid u) \to p^*(x \mid u)\)), and conditional generative reasoning (\(p(y \mid s, x, u) \to p^*(y \mid s, x, u)\)); isolating or optimizing any single component alone cannot guarantee systemic simulation reliability.
2. Attribute-Specific Predictive Gain (ASPG) & Information Modeling
To address the practical reality that persona attributes are not uniformly useful and simply packing prompts with exhaustive features leads to diminishing or negative returns, this design introduces Attribute-Specific Predictive Gain (ASPG) to quantify the information-theoretic utility of each profile attribute \(a\) with respect to target behavior \(Y\):
where \(I(a; Y)\) denotes mutual information and \(H(Y)\) denotes label Shannon entropy. Calculating the Gini coefficient over ASPG values across demographic datasets reveals heavy right-skewed concentration: the top 20% most informative attributes capture over 50% of total predictive information, whereas the bottom 50% account for less than 22%. By comparing quantity-based completeness correlation (\(r_Q\)) with information-weighted completeness correlation (\(r_I\)), the analysis reveals that \(r_I\) exhibits substantially stronger correlation with individual accuracy (\(\Delta = r_I - r_Q\) reaching up to +0.1449 on Party and +0.1752 on Immigration). This proves that attribute informativeness and semantic quality govern simulation performance far more than raw attribute count, while multi-attribute prompt interference explains performance dips under dense profiles.
3. Two-Level Reliability Decoupling Audit (R1 vs R2)
To test the long-standing assumption in synthetic polling that higher individual accuracy directly translates to superior population-level fidelity, this design tracks R1 (ACC) and R2 (TVD) concurrently across all models, datasets, and completeness increments. Across all profile completeness steps, 26% of incremental steps exhibit outright decoupling, where individual accuracy improves while population TVD degrades (or vice versa). By conducting option-level error tracking, the mechanism behind this decoupling is exposed: providing additional persona attributes frequently causes LLMs to become overconfident and collapse their predictions into dominant majority classes (e.g., escalating the over-prediction of "Quite good" from +4.00% to +13.06% while severely under-predicting intermediate neutral stances). Consequently, while individual discrete accuracy ticks upward on certain subgroups, the aggregated macro-distribution suffers severe polarization and structural distortion.
4. Population Coverage Scaling & Variance-Bias Decomposition
To determine whether scaling virtual cohort size \(N\) can continuously eliminate population-level simulation error, this design benchmarks simulated population curves against an empirical ground-truth human sampling baseline (aggregating actual human responses subsampled at size \(N\)). By systematically expanding simulated cohort sizes from \(N=10\) up to \(N=460+\), the framework isolates Monte Carlo sampling variance from systematic behavioral bias. While the human ground-truth baseline decreases monotonically toward zero as \(N\) expands, LLM simulation TVD curves rapidly level off into a non-zero error plateau between \(N \approx 50\) and \(100\). This empirical plateau proves that increasing simulated sample size only eliminates aggregation variance; the persistent residual gap is dominated by systematic cognitive and alignment biases inherent to the base model and prompting conditioning, which cannot be eradicated by scaling virtual respondents alone.
Benchmark & Evaluation Protocol¶
The empirical evaluation spans four diverse benchmarks across structured surveys and dynamic social events: 1. Party: European Social Survey (ESS 11), predicting German voter alignment across 9 political parties (100 demographic/attitudinal attributes, 1,000 validated respondents); 2. Immigration: World Values Survey (WVS Wave 7), evaluating public perception of immigrants on country development across a 5-point Likert scale (200 attributes, 9,994 validated respondents); 3. Religion: SocioBench, assessing whether churches and religious organizations hold excessive institutional power across a 5-point scale (80 attributes, 500 validated respondents); 4. Media: A dynamic Chinese social media corpus capturing netizen attitudes toward the "Luckin Coffee stock collapse" across 5 categories (Loyal Advocate, Opportunist, Moral Outrage, Rational Evaluator, Bystander Entertainer; profiles constructed from 30 historical Weibo posts, 150 validated users).
The model testbed encompasses 11 representative LLMs spanning diverse architectures: proprietary frontier models (GPT-4.1, Gemini-2.5-Pro), open-source general and reasoning models (DeepSeek-V3.2-Exp, DeepSeek-R1), and the Qwen3 family across dense and MoE configurations (Qwen3-8B, 14B, 32B, 30B-A3B-Instruct/Thinking, 235B-A22B-Instruct/Thinking).
Key Experimental Results¶
Main Results: Base Distribution Bias and Profile-Conditioned Alignment¶
Without persona profile conditioning (Layer I, Base), all models exhibit severe population-level distributional bias. Adding complete persona conditioning (Layer II/III, c=100%) drastically improves individual accuracy and reduces population TVD across tasks.
| Model | Party Base TVD↓ | Party c=100% (ACC%↑ / TVD↓) | Immig. Base TVD↓ | Immig. c=100% (ACC%↑ / TVD↓) | Relig. Base TVD↓ | Relig. c=100% (ACC%↑ / TVD↓) | Media Base TVD↓ | Media c=100% (ACC%↑ / TVD↓) |
|---|---|---|---|---|---|---|---|---|
| GPT-4.1 | 0.670 | 41.6% / 0.155 | 0.368 | 38.6% / 0.160 | 0.542 | 50.8% / 0.223 | 0.747 | 56.7% / 0.147 |
| Gemini-2.5-Pro | 0.589 | 39.8% / 0.180 | 0.297 | 37.8% / 0.171 | 0.549 | 49.2% / 0.228 | 0.527 | 50.7% / 0.173 |
| DeepSeek-V3.2-Exp | 0.483 | 42.1% / 0.162 | 0.128 | 36.4% / 0.155 | 0.342 | 51.6% / 0.245 | 0.613 | 55.3% / 0.150 |
| DeepSeek-R1 | 0.664 | 38.4% / 0.221 | 0.579 | 35.1% / 0.224 | 0.345 | 48.0% / 0.260 | 0.480 | 50.0% / 0.210 |
| Qwen3-235B-A22B-Inst | 0.271 | 41.3% / 0.154 | 0.482 | 37.1% / 0.165 | 0.263 | 52.4% / 0.181 | 0.684 | 56.0% / 0.180 |
| Qwen3-32B | 0.664 | 32.9% / 0.315 | 0.408 | 33.2% / 0.211 | 0.474 | 45.6% / 0.252 | 0.653 | 46.7% / 0.220 |
| Qwen3-14B | 0.722 | 31.4% / 0.334 | 0.326 | 32.8% / 0.218 | 0.500 | 44.2% / 0.265 | 0.578 | 44.0% / 0.233 |
| Qwen3-8B | 0.689 | 28.8% / 0.315 | 0.304 | 32.0% / 0.225 | 0.540 | 44.8% / 0.270 | 0.620 | 43.3% / 0.247 |
Ablation Study & Dynamic Decoupling: R1 vs R2 Movement under Profile Scaling¶
Tracking relative percentage changes in accuracy (\(\Delta\text{ACC}\)) and population divergence (\(\Delta\text{TVD}\)) relative to preceding completeness levels highlights persistent decoupling across models and datasets.
| Dataset | Step \(c\) (%) | 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 \(\Delta\text{TVD}\) | Diagnostic Decoupling Observation |
|---|---|---|---|---|---|---|---|---|
| Party | \(0 \to 20\) | +26.50% | -55.52% | +44.45% | -43.66% | +144.57% | +5.91% | Qwen3 individual ACC explodes while TVD worsens |
| Party | \(60 \to 80\) | -1.19% | +14.64% | -3.00% | -1.33% | +1.51% | +42.54% | Qwen3 ACC improves +1.51% while TVD surges +42.54% |
| Immigration | \(12.5 \to 25\) | +1.87% | -13.44% | +1.26% | +2.90% | +1.57% | +10.58% | Both open models show ACC gain alongside worse TVD |
| Immigration | \(50 \to 62.5\) | -1.53% | +24.17% | -12.58% | +65.56% | -6.49% | +42.34% | Feature clutter triggers universal collapse across models |
| Religion | \(37.5 \to 62.5\) | +7.19% | -3.18% | +7.29% | +1.96% | +2.56% | +19.91% | DeepSeek & Qwen exhibit decoupling (ACC↑, TVD↑) |
| Media | \(66.7 \to 83.3\) | -2.56% | -7.54% | -1.91% | +3.36% | -4.92% | -20.51% | GPT & Qwen ACC drops but population TVD improves |
Key Findings¶
- Finding 1: Diminishing Returns in Profile Conditioning. Across all tasks, individual accuracy (ACC) scales positively with attribute coverage \(c\), but marginal gains taper off rapidly due to uninformative features and prompt-level cognitive load.
- Finding 2: Capacity Bottlenecks in Profile Scaling. Frontier scale models (GPT-4.1, DeepSeek-V3.2-Exp, Qwen3-235B) continuously exploit dense profiles, whereas smaller models (Qwen3-8B, 14B) plateau early. Crucially, long-chain reasoning models (DeepSeek-R1, Qwen3-Thinking) fail to systematically outperform standard instruction models on behavioral simulation, demonstrating that sociological perspective-taking demands distinct representations from formal mathematical reasoning.
- Finding 3 & 4: Attribute Informativeness Dominates Raw Quantity. ASPG exhibits extreme Gini inequality (0.44 to 0.51). Information-weighted completeness correlations (\(r_I\)) surpass raw count correlations (\(r_Q\)) by up to +0.2312, proving that selecting high-leverage traits matters far more than expanding attribute lists.
- Finding 5: Decoupling of Individual Accuracy and Population Alignment. In 26% of incremental steps, R1 and R2 diverge in opposite directions. Optimizing per-respondent accuracy alone cannot reliably improve population-level simulation fidelity.
- Finding 6: Population Coverage Reduces Variance but Not Systematic Bias. Scaling simulated population size \(N\) levels off at \(N \approx 50 \sim 100\). While sampling variance contracts, a substantial non-zero plateau of systematic distributional bias persists regardless of sample scale.
Highlights & Insights¶
- Formal Generative Decomposition of Social Simulation: Establishes the first rigorous multi-layer error-propagation inequality bounding macro-discrepancy by individual sampling, profile construction, and scenario generation errors.
- Empirical Proof of R1/R2 Decoupling: Disproves the intuitive belief that micro-level accuracy gains automatically repair macro-distribution fidelity by presenting concrete option-shift mechanisms behind distributional skew.
- Pragmatic Sample Scaling Bound (\(N \in [50, 100]\)): Identifies the precise coverage plateau where Monte Carlo variance vanishes, preventing wasteful over-allocation of computational budgets to sample sizes rather than prompt-level debiasing.
Limitations & Future Work¶
- Single-Turn Discrete Survey Boundary: Experiments focus on discrete, single-turn multiple-choice surveys and categorical stance detection, leaving multi-turn social interaction, group dynamics, and agentic negotiations unaddressed.
- Winner-Takes-All Accuracy Metric: Evaluating R1 strictly via discrete exact match fails to account for inherent subjectivity and probabilistic ambivalence in human cognition.
- Observational Evaluation Scope: While the framework maps out how configuration levers shape reliability, internal mechanistic explanations (e.g., representation attention routing across persona attributes) remain open for future interpretability studies.
Related Work & Insights¶
- vs SimBench / SocioBench: Prior benchmarks focus on dataset expansion and static per-respondent accuracy scoring without modeling layer-wise error propagation or population-level distribution convergence; ReliMap unifies three layers, three configuration axes, and two reliability levels.
- vs Pluralistic / Distributional Alignment (MaxMin-RLHF, DPO-Group): Alignment literature treats group preference matching as a training objective; ReliMap operates as a diagnostic audit suite to uncover failure loci, decoupling trends, and bias plateaus in existing pre-trained systems.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ [Pioneering three-layer decomposition and formal error-propagation inequality for LLM behavior simulation]
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ [11 diverse LLMs, 4 benchmarks, full sweeps over profile completeness, sample sizes, and alternative divergence metrics]
- Writing Quality: ⭐⭐⭐⭐⭐ [Rigorous mathematical foundation, clear narrative progression, and insightful micro-to-macro case analyses]
- Value: ⭐⭐⭐⭐⭐ [Provides essential diagnostic principles and practical scaling guidelines for computational social science and synthetic polling]