Skip to content

Can Persona-Conditioned Language Agents Simulate Counterfactual Online Discourse Dynamics?

Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: Social Simulation, State-Space Modeling, Affective Dynamics, Counterfactual Intervention, Online Discourse Forecasting

TL;DR

Addressing the failure of prompt-driven language agents to capture longitudinal discourse evolution, this work presents PRISM, a state-space framework decoupling latent affective SDE evolution, POMDP participation control, and controlled text generation to deliver high-fidelity discourse forecasting and counterfactual interventions on 47k social media threads.

Background & Motivation

Large language models are increasingly adopted as social simulators, synthetic annotators, and proxy human subjects in computational social science. However, current evaluations of language agents predominantly focus on isolated utterance plausibility and local fluency rather than their ability to reproduce longitudinal, temporal discourse dynamics. A truly functional online discourse simulator must accomplish far more than producing coherent individual turns: it must accurately forecast who decides to speak, who stays silent, how emotional affect diffuses across social network topologies, when disagreements escalate into conflict, and how conversational cadence shapes downstream polarization over extended horizons.

Existing simulation paradigms reveal severe conceptual shortcomings when faced with this objective. Traditional agent-based models (ABMs) impose homogeneous or low-dimensional rule sets that fail to capture the subtle, heterogeneous cognitive and emotional reactions shaping natural online discussions. Conversely, contemporary LLM-powered multi-agent architectures achieve surface-level eloquence, but unconstrained prompt-only persona conditioning suffers from chronic persona drift over long conversational horizons, falls into stereotypical caricatures, and critically lacks calibrated mathematical mechanisms for affective memory decay, participation timing control, and rigorous intervention testing. Furthermore, while continuous trait frameworks like the Big Five provide scalar conditioning, they cannot directly govern discrete communicative decisions such as whether to reply, withdraw into silence, provide emotional comfort, challenge, or amplify a specific message.

To bridge this fundamental divide, this work reformulates online discourse simulation as a calibrated state-space language-agent problem: diverse behavioral profiles systematically filter identical social inputs through distinct latent dynamics, prompting individualized internal state transitions and discrete action choices before rendering responses. Core idea: decouple online discourse simulation into a three-layer state-space architecture comprising profile-conditioned SDE affective evolution, partially observed participation activation, and controlled language generation, grounding heterogeneous responses in empirical behavioral priors to achieve robust longitudinal trajectory forecasting and counterfactual discourse intervention.

Method

Overall Architecture

PRISM formulates online discourse simulation as a conditional forecasting and generation problem over complex interaction topologies. Given an observed conversational prefix history \(H_{1:k}\), an exposure graph \(G=(V, E)\), and profile anchors \(z_{1:n}\) for \(n\) participating agents, the simulator sequentially forecasts future discrete actions \(a_{k+1:T}\), textual utterances \(x_{k+1:T}\), and affective observations \(y_{k+1:T}\).

The framework operates via three seamlessly coordinated functional tiers: First, empirical affective prior centroids and posterior distributions are constructed from large-scale historical forum data via Dirichlet smoothing. Second, during each simulation interval, an agent's continuous latent affect evolves under a multivariate Ornstein–Uhlenbeck (OU) stochastic differential equation driven by mean-reverting drift, Brownian diffusion, and instantaneous message impulses from active conversational neighbors. Third, a partially observed participation module aggregates local exposure, affective displacement, and direct user mentions to determine whether an agent replies or remains silent. Finally, the backbone language model renders the selected state-action configuration into natural language, feeding newly generated affective impulses back into the social graph.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Observed Thread Prefix H1:k & Interaction Graph G"] --> B["Data-Driven Profile Priors<br/>Dirichlet smoothed posterior & continuous affective centroids"]
    B --> C["Continuous Affective Evolution SDE<br/>OU mean reversion + diffusion volatility + message impulses"]
    C --> D["Partially Observed Participation Activation<br/>Exposure + affective displacement + mention gating"]
    D -->|Activation Triggered| E["Controlled State-Action Generation<br/>Context ⊕ Affect ⊕ Profile ⊕ Action prompt conditioning"]
    D -->|Withhold Response| F["Maintain Silence & Natural Affective Decay"]
    E --> G["Feedback Impulses & Network Affective Propagation"]
    G --> C

Key Designs

1. Data-Driven Profile Priors: Calibrated Anchors Bridging Weak Labels and Continuous Dynamics

Addressing the issue where unstructured persona prompts lead to behavioral drift and lack empirical grounding, this design establishes empirical prior distributions over behavioral response styles. The framework utilizes 16 discrete MBTI-like types as operational behavioral anchors (treated strictly as response priors rather than psychological diagnostic truth). Over an emotion inventory \(E=\{e_1, \dots, e_K\}\), Dirichlet smoothing with parameter \(\alpha=1.0\) estimates empirical posterior distributions from large-scale weak annotations: \(\hat{\theta}_{z,k} = \frac{N(e_k, z) + \alpha}{\sum_j N(e_j, z) + K\alpha}\), refined via a coarse-graining rounding step (\(\epsilon=0.05\)) to prevent overfitting to label noise, yielding \(\hat{\theta}^{\text{final}}_{z,k}\). These discrete priors are projected into continuous space via fixed emotion embeddings \(q(e_k)\) to establish baseline affective centroids: $\(\mu_z = \sum_{k=1}^K \hat{\theta}^{\text{final}}_{z,k} q(e_k)\)$ The distribution sharpness is quantified via Shannon entropy \(H(z) = -\sum_k \hat{\theta}^{\text{final}}_{z,k} \log_2 \hat{\theta}^{\text{final}}_{z,k}\), providing a stable, mathematically grounded basin of attraction for each agent profile.

2. Continuous Affective Evolution SDE: Mean Reversion and Impulse-Driven Affective Trajectories

Addressing the limitation of static sentiment tags that ignore temporal emotional decay, affective contagion, and sudden conversational shocks, this design models an agent's continuous latent affect \(e_i(t) \in \mathbb{R}^d\) via a multivariate Ornstein–Uhlenbeck stochastic differential equation: $\(de_i(t) = \Theta_{z_i} (\mu_{z_i} - e_i(t)) dt + \Sigma_{z_i} dW_t + J_i(t)\)$ Here, \(\Theta_{z_i}\) is a profile-specific stiffness matrix controlling the rate of mean reversion toward centroid \(\mu_{z_i}\), \(\Sigma_{z_i}\) dictates diffusion volatility against Wiener increment \(dW_t\), and \(J_i(t) = \sum_{\omega \in M_i} \delta(t - t_\omega) \Psi_{z_i} v(\omega)\) models incoming impulses from received messages \(M_i\). The term \(v(\omega)\) encodes the valence-arousal profile of message \(\omega\), while \(\Psi_{z_i}\) denotes the profile's affective susceptibility matrix. Between message events, the state evolves under an analytical closed-form trajectory, providing an exact and physically interpretable description of how individuals react, flare up, and gradually calm down over conversation turns.

3. Partially Observed Participation Activation: Adaptive Silence and Gated Response Thresholds

Addressing the standard unrealistic assumption in multi-agent systems where every agent responds in lockstep round-robin schedules, this design formulates participation control as a partially observed Markov decision process (POMDP). Agents observe only their local immediate neighborhood rather than the global platform state, computing an activation utility \(U_i(t)\): $\(U_i(t) = w_g \sum_{j \in \mathcal{N}_i} G_{ij} \mathbb{I}[j \in \mathcal{A}_{t-1}] + w_e \|e_i(t) - \mu_{z_i}\|_2 + w_m \mathbb{I}_{\text{mention}}(i, t)\)$ This formulation linearly synthesizes structural exposure from recently active neighbors, internal affective displacement from the equilibrium baseline, and direct conversational mentions. Discrete action selection is governed by a profile-specific activation threshold \(\Gamma_{z_i}\): the agent executes \(\text{REPLY}\) if \(U_i(t) > \Gamma_{z_i}\), and remains in \(\text{SILENCE}\) otherwise. This allows extroverted and emotionally volatile agents to dominate fast-paced conversational cascades while reserved agents remain passive until provoked.

4. Controlled State-Action Generation: Decoupled Realization of Internal Dynamics into Natural Text

Addressing the pitfall of end-to-end LLMs conflating state tracking with linguistic generation, this component restricts the LLM to a controlled surface text renderer. When an agent is triggered to speak, the observed dialogue history \(H_t\), projected affective state \(\phi(e_i(t))\), profile behavioral constraints \(\rho(z_i)\), and selected action intent \(\eta(a_i^t)\) are serialized into an instruction-tuned prompt: $\(h_t = \mathcal{E}(H_t) \oplus \phi(e_i(t)) \oplus \rho(z_i) \oplus \eta(a_i^t), \quad P(x_t \mid H_t, e_i(t), z_i, a_i^t) = \text{LM}_\theta(x_t \mid h_t)\)$ The system prompt strictly forbids explicit verbalization of internal variables (such as MBTI codes or numeric state coordinates), instructing the model to reflect latent emotion and persona purely through lexical choices, tone, and rhetorical style. This clean separation prevents semantic leakage while guaranteeing that generated dialogue faithfully reflects underlying mathematical states.

Loss & Training

System parameter estimation and training adhere strictly to modularity and data isolation: 1. SDE Parameter Estimation: On the training split, discrete transitions \(A_z = \exp(-\Theta_z \Delta t)\) are estimated via ridge-regularized regression over consecutive affective states, recovering stiffness \(\Theta_z = -\frac{1}{\Delta t} \log(A_z)\). Volatility \(\Sigma_z\) is obtained from empirical residual covariance, while susceptibility \(\Psi_z\) is learned by regressing post-message residual shifts against incoming valence-arousal embeddings. 2. Activation Policy Optimization: The participation module is trained as a profile-conditioned logistic model with globally shared weights \(w_g, w_e, w_m\) and profile-specific thresholds \(\Gamma_z\) and slopes \(\beta_z\). Parameters are optimized via \(\ell_2\)-regularized negative log-likelihood: $\(\mathcal{L}_{\text{act}} = -\sum_{i,t} \left[ r_i(t) \log p_i(t) + (1 - r_i(t)) \log(1 - p_i(t)) \right] + \lambda \|w\|_2^2\)$ 3. Language Generator Tuning: Built upon a Qwen3-8B-VL backbone, role-conditioned instruction tuning is applied across 8 NVIDIA A100 (80GB) GPUs. Inference employs temperature 0.7, Top-p 0.9, and a strict 96-token maximum budget to mirror concise social-media comment structures.

Key Experimental Results

Main Results

Evaluation is performed across 47,207 social media discussion threads spanning 159 trending topics and 356,593 hierarchical comments. Models are evaluated in held-out prefix forecasting (observing prefix \(H_{1:k}\) and predicting remaining sentiment trajectories, reply graphs, and timestamps).

Method Sent. MAE ↓ DTW Distance ↓ Consistency (Spearman \(\rho\)) ↑ Reply F1 ↑ Tree Edit Dist. (TED) ↓ Time MAE ↓
Homogeneous Rational (HRA) \(0.42 \pm 0.03\) \(0.64 \pm 0.04\) \(0.154\) \(0.16 \pm 0.02\) \(0.83 \pm 0.04\) \(0.54 \pm 0.04\)
Random Behavior (RBM) \(0.55 \pm 0.04\) \(0.82 \pm 0.05\) \(0.021\) \(0.11 \pm 0.01\) \(0.92 \pm 0.05\) \(0.68 \pm 0.05\)
Big Five Personality (BFPM) \(0.21 \pm 0.02\) \(0.44 \pm 0.03\) \(0.613\) \(0.26 \pm 0.02\) \(0.67 \pm 0.04\) \(0.46 \pm 0.03\)
Prompt-only MBTI Agents \(0.27 \pm 0.02\) \(0.41 \pm 0.03\) \(0.66 \pm 0.03\) \(0.31 \pm 0.02\) \(0.59 \pm 0.03\) \(0.42 \pm 0.03\)
Big Five Prompted Agents \(0.28 \pm 0.02\) \(0.43 \pm 0.03\) \(0.63 \pm 0.04\) \(0.29 \pm 0.02\) \(0.61 \pm 0.03\) \(0.44 \pm 0.03\)
Memory-only LLM Agents \(0.30 \pm 0.03\) \(0.47 \pm 0.04\) \(0.58 \pm 0.04\) \(0.34 \pm 0.02\) \(0.57 \pm 0.03\) \(0.39 \pm 0.03\)
No-SDE Agents \(0.20 \pm 0.02\) \(0.35 \pm 0.02\) \(0.71 \pm 0.02\) \(0.39 \pm 0.02\) \(0.51 \pm 0.03\) \(0.37 \pm 0.02\)
No-Activation Agents \(0.21 \pm 0.02\) \(0.37 \pm 0.02\) \(0.70 \pm 0.03\) \(0.32 \pm 0.02\) \(0.63 \pm 0.03\) \(0.49 \pm 0.03\)
PRISM (Ours) \(\mathbf{0.14 \pm 0.01}\) \(\mathbf{0.27 \pm 0.02}\) \(\mathbf{0.79 \pm 0.02}\) \(\mathbf{0.48 \pm 0.02}\) \(\mathbf{0.41 \pm 0.02}\) \(\mathbf{0.28 \pm 0.02}\)

In blind human evaluations (200 thread prefixes, 1,000 pairwise comparisons, Krippendorff's \(\alpha \in [0.57, 0.61]\)), PRISM substantially outscores Prompt-only MBTI across realism (4.18 vs. 3.71), profile consistency (4.27 vs. 3.89), and emotional coherence (4.21 vs. 3.55), winning a 63.5% human preference win rate.

Ablation Study

Controlled ablations isolating the architectural layers and persona trait dimensions clarify their specific structural roles:

Variant / Configuration Polarity Error ↓ Relative Change \(\Delta\) Consistency (Spearman \(\rho\)) ↑ Relative Change \(\Delta\) Primary Impact & Role
PRISM Full Model \(\mathbf{0.140}\) — \(\mathbf{0.782}\) — Full state-space SDE + POMDP + 4D persona
w/o Sensing/Intuition \(0.160\) \(+14.3\%\) \(0.741\) \(-5.2\%\) Minor impact on granular information processing
w/o Judging/Perceiving \(0.190\) \(+35.7\%\) \(0.692\) \(-11.5\%\) Regulates decision timing and trajectory stability
w/o Extraversion/Introversion \(0.220\) \(+57.1\%^*\) \(0.588\) \(-24.8\%^*\) Dictates participation intensity and cadence
w/o Thinking/Feeling \(0.320\) \(+128.6\%^{**}\) \(0.453\) \(-42.1\%^{**}\) Primary driver of emotional polarity & affect direction
w/o SDE Dynamics (No-SDE) \(0.180\) \(+28.6\%\) \(0.710\) \(-9.2\%\) Degrades continuous affective alignment & DTW
w/o Activation (No-Activation) \(0.190\) \(+35.7\%\) \(0.700\) \(-10.5\%\) Severe drop in Reply F1 (down to 0.32) and tempo

Counterfactual platform interventions simulated by PRISM show that combining Balanced Exposure, Affect Damping, and Reply Friction lowers community polarization from \(0.62\) to \(0.38\), cuts cascade depth from \(6.8\) to \(3.9\), and reduces toxicity from \(0.11\) to \(0.05\).

Key Findings

  • Thinking/Feeling is the primary driver of affective dynamics: Ablating the T/F dimension induces a massive 128.6% surge in polarity error and a 42.1% plunge in behavioral consistency, proving that the tension between analytical detachment and emotional resonance is foundational to online polarization.
  • Clear modular division of labor: The SDE layer primarily refines continuous trajectory matching (reducing DTW from 0.35 to 0.27), while the POMDP activation policy controls network structural emergence (boosting Reply F1 from 0.32 to 0.48 and cutting Time MAE from 0.49 to 0.28).
  • Invariance to semantic label naming: Replacing MBTI codes with neutral IDs (PROFILE-A~P, Table 13) leaves Reply F1 unchanged (0.47) and shifts Polarity MAE only marginally (\(0.14 \to 0.15\)), confirming that gains stem from mathematical state-space dynamics rather than superficial LLM semantic associations with personality names.

Highlights & Insights

  • Decoupled cognitive state-space paradigm: Replaces monolithic black-box prompting by treating the LLM strictly as a controlled natural language renderer, while SDEs and POMDPs handle internal emotional calculus and discrete behavioral activation.
  • Rigorous bridging of discrete text and continuous calculus: Seamlessly interfaces discrete natural language forum posts with continuous Ornstein–Uhlenbeck equations through Dirichlet-smoothed empirical affective centroids.
  • In-silico platform intervention sandbox: Enables counterfactual policy exploration—such as damping high-arousal content susceptibility or injecting response friction—to observe systemic polarization shifts without invasive real-world platform experimentation.

Limitations & Future Work

  • Validity bounds of persona anchors: MBTI archetypes are employed purely as discrete computational response priors and must not be misinterpreted as validated psychological ground truth.
  • Residual weak-supervision circularity: Despite extensive human audits and independent-model validation, initial priors rely on LLM annotations, leaving potential room for subtle machine-annotation biases.
  • Simulation scope constraints: Focus is currently confined to emotional valence, interaction trees, and response intervals; future expansions should encompass factual misinformation propagation, ideological stance, multimodal memes, and cross-cultural network dynamics.
  • vs. Classical Agent-Based Models (ABMs, e.g., Mueller-Frank, 2013): Classical ABMs rely on rigid algebraic heuristics and cannot parse or generate nuanced natural language; PRISM preserves continuous dynamical rigor while infusing rich contextual text generation via LLMs.
  • vs. Prompt-Only Generative Agents (e.g., Park et al., 2023): Prompt-only agents lack formal state-space tracking, experiencing severe persona drift and uncalibrated participation rates over multi-turn horizons; PRISM introduces explicit SDE and POMDP layers to maintain bounded, realistic long-term behavioral trajectories.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Pioneering state-space formulation combining stochastic differential equations, POMDP activation gating, and controlled language generation.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Exceptionally rigorous validation across 47k threads, 5 baselines, ablation studies, blind human evaluations, and semantic-label controls.
  • Writing Quality: ⭐⭐⭐⭐⭐ Flawless mathematical formulation, clear narrative architecture, and disciplined treatment of behavioral priors.
  • Value: ⭐⭐⭐⭐⭐ An invaluable methodological framework for computational social science, online content moderation, and algorithmic governance research.