Skip to content

PRISM: A Personality-Driven Multi-Agent Framework for Social Media Simulation

🕒 Published (v1): 2025-12-22 23:31 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

PRISM couples stochastic differential equations (SDE) for continuous emotional evolution with a Personality-Conditional Partially Observable Markov Decision Process (PC-POMDP) for discrete interaction logic, assigning MBTI-typed cognitive policies to MLLM agents initialized from 47,207 real social media topics. It outperforms homogeneous and Big Five baselines, achieving a 66.7% reduction in sentiment polarity error (MAE 0.14 vs. 0.42) and Spearman ρ = 0.782 against human trajectories.

Problem

Traditional agent-based models of opinion dynamics assume homogeneous agents with uniform interaction rules, failing to capture psychological heterogeneity—particularly the interplay between individual cognitive biases (rational suppression, affective resonance) and information propagation that drives polarization on social media. Existing MLLM-driven simulations lack data-driven personality initialization and suffer from "state amnesia" (no long-term behavioral coherence across turns).

Method

PRISM assigns each of 16 MBTI personality types a distinct agent, fine-tuned from Qwen3-8B-VL via Role-Conditioned Instruction Tuning (RoleCIT) for 5 epochs. Emotional state is modeled as a multivariate Ornstein-Uhlenbeck process (continuous SDE layer); discrete interaction decisions (reply vs. silence) are governed by a PC-POMDP whose activation threshold is personality-conditioned. Empirical emotion prior distributions P(e_k | T_i) are estimated from the dataset using Dirichlet-smoothed frequency counts and coarse-grained to ε = 0.05 resolution. When an agent receives a message, a jump term ΨTi · v(ω) (susceptibility matrix × valence-arousal vector) perturbs the continuous emotional state; agents exceeding their threshold Γ_Ti generate MLLM responses conditioned on dialogue history, emotional state projection φ(e), and MBTI linguistic style ρ(T_i). GPT-4o annotated the corpus with MBTI and sentiment labels to build the priors.

Key Contributions

  • First multi-agent simulation grounded in real-world multimodal social media data (47,207 topics) with MBTI profiling validated at scale.
  • Hybrid SDE + PC-POMDP formalism that resolves state amnesia in ABMs by maintaining continuous emotional state across discrete interaction events.
  • Threshold-based activation model replacing round-robin turn-taking, where emotional agitation (||e - μ_T||²) directly raises interaction probability.
  • Empirical Bayesian parameterization: Dirichlet-smoothed affective priors per MBTI type with Shannon entropy analysis distinguishing T-types (low entropy, dominant Neutral) from F-types (high entropy, diffuse affect).

Results

  • PRISM MAE = 0.14 ± 0.03 vs. BFPM 0.21 ± 0.05, HRA 0.42 ± 0.08, RBM 0.55 ± 0.11 (lower is better).
  • Consistency ρ = 0.782 (p < 0.01) vs. BFPM 0.613 (p = 0.05), HRA 0.154 (p = 0.12).
  • 66.7% reduction in distributional divergence vs. homogeneous baseline.
  • EQ-Bench: ENFJ agent scores 81.48 (+11.36 over Qwen3-8B-VL base 70.12); ISTP scores 61.97 (−8.15), confirming functional divergence between F- and T-type agents.
  • Ablation: removing T/F dimension degrades MAE by +128.6% and ρ by −42.1%; removing E/I degrades ρ by −24.8% (p < 0.01).

Limitations

  • MBTI's scientific validity is contested in clinical psychology; the paper does not address whether MBTI categories are the most principled choice over alternatives.
  • Personality labels are inferred by GPT-4o from text via distant supervision—no ground-truth MBTI annotations exist for the dataset.
  • Evaluation is on one large dataset with topical controversy; generalizability to other platforms, languages, or cultural contexts is untested.
  • The 16 agent profiles require separate fine-tuning runs, making the framework computationally expensive to extend or retrain.
  • Social graph structure is treated as fixed input; dynamic network evolution (follow/unfollow, algorithmic feed) is not modeled.

Relevance to Harnesses / Meta-Harnesses

PRISM is an instantiation of a multi-agent simulation harness where agent behavior is governed by typed cognitive policies rather than a single monolithic prompt—a pattern directly analogous to role-specialized sub-agent pools in meta-harness architectures. The PC-POMDP activation mechanism (threshold-gated agent invocation based on internal state + social context) mirrors the routing and dispatch logic in harnesses that selectively invoke agents based on task state. The empirical prior initialization pipeline—GPT-4o annotation → Dirichlet estimation → agent fine-tuning → simulation loop—is itself a multi-stage harness for constructing heterogeneous agent populations. For researchers building meta-harnesses, PRISM demonstrates how typed agent specialization with continuous state tracking can reduce behavioral drift in long-horizon multi-agent interactions.