Persuasiveness Prediction Model

Updated 16 January 2026

Persuasiveness prediction models are quantitative systems that forecast belief changes and behavior adoption using multimodal algorithms.
They integrate psychological theories, annotated datasets, and fusion techniques to capture persuasive effects across text, visual, and conversational domains.
Advanced frameworks employ ensemble methods and transformer architectures to achieve high accuracy and robustness in diverse applications.

A persuasiveness prediction model quantitatively estimates whether a message or communicative artifact will induce belief change, behavioral adoption, or other persuasive effects in its target audience. These systems operationalize “persuasion” in diverse ways, ranging from observable behavioral outcomes (e.g., awarded deltas, donation acts, click-throughs, engagement scores) to probabilistic predictions of attitudinal shift or strategic rhetorical effectiveness. Modern frameworks are multi-modal, theory- and data-driven, and increasingly leverage LLMs, custom feature schemas, and fusion architectures to robustly forecast persuasion success across digital, conversational, and visual domains.

1. Data Sources and Labeling Paradigms

Persuasiveness models rely on annotated datasets where ground-truth labels reflect actual belief change or behavioral compliance. For text-based persuasion, the Change My View (CMV) forum on Reddit serves as a standard: the original poster (OP) awards a “delta” (Δ) to replies that changed their mind, yielding binary labels ( $y\in\{0,1\}$ ) for belief revision (Hoang et al., 27 Nov 2025). Controlled persuasion experiments (Truth Wins) record continuous pre- and post-message belief scores, and human raters supply Likert-scale evaluations of psychological features. Visual persuasion assessments use pairwise A/B datasets: WiserUI-Bench for UI design (ground truth from real A/B engagement tests) (Jeon et al., 8 May 2025), banded image-message judgments with high Fleiss’ $\kappa$ consensus for visual persuasive cues (Park, 21 Nov 2025). Behavioral-outcome centric corpora define persuasiveness by metrics such as percentiles of likes on social media (PersuasionBench) or choice rates in language-based persuasion games (Apel et al., 2020). Personalized models exploit users’ histories—posts, comments, meta-data—and optimize context-aware retrieval and summarization for profile-driven prediction (Park et al., 9 Jan 2026).

2. Feature Engineering and Theoretical Foundations

Feature design in persuasion prediction spans interpretable, theory-grounded psychological dimensions and raw multimodal representations. Core psychological features adapted from persuasion literature include: influence, interest, epistemic emotion (“interesting-if-true”), emotional valence (positive/negative), shareability, truthfulness, and attention-capture (Hoang et al., 27 Nov 2025). Persuasion strategy taxonomies enumerate rhetorical maneuvers such as attack on reputation, distraction, manipulative wording, simplification, justification, and call-to-action (Labruna et al., 15 Jan 2026). Visual persuasiveness models factorize cues into low-level perceptual (colorfulness, brightness), mid-level compositional (saliency entropy, center-bias), and high-level semantic (key object-alignment, human presence) features (Park, 21 Nov 2025). Multimodal fusion systems extract acoustic, facial, and linguistic embeddings with adaptive weighting and alignment loss (Bai et al., 2020).

Personalized systems build dynamic user profiles from context-relevant historical activity, cognitive/emotional traits, and values, using retrieval-augmented LLM summarization (Park et al., 9 Jan 2026). Item Response Theory frameworks introduce latent listener persuadability ( $\alpha_i$ ), argument difficulty ( $\beta_j$ ), and discrimination ( $\phi_j$ ), each parameterized by style (LIWC, affect, polarity, argument cues), content (TF-IDF, ngram), and speaker/recipient attributes; these latent factors enable prediction of individualized response probabilities (Kornilova et al., 2022).

3. Model Architectures and Algorithms

State-of-the-art persuasiveness predictors frequently deploy ensemble classifiers, deep sequence models, and hybrid fusion pipelines. Canonical approaches synthesize:

Random forest classifiers trained on LLM-derived psychological features, with permutation-based variable importance (VI) and majority-vote inference (Hoang et al., 27 Nov 2025).
Multimodal adaptive fusion: Transformer-based per-modality encoders, alignment and heterogeneity fusion modules, and reference-model-soft weighting for acoustic, visual, and language signals (Bai et al., 2020).
Multi-strategy persuasion scoring (MS-PS) leverages zero-shot LLM reasoning over multiple rhetorical strategies, aggregates strategy-specific scores, and feeds them to MLP classifiers for held-out accuracy and regression metrics (Labruna et al., 15 Jan 2026).
Personalized profiling frameworks: Query generator LLMs, retrieval via embedding similarity (e.g., BGE-M3), LoRA-adapted profile summarizers, and classification or regression heads with jointly-trained DPO losses—yielding highly context-sensitive, persuadee-specific predictions (Park et al., 9 Jan 2026).
Linear probe analysis of LLM hidden states in multi-turn conversational persuasion; residual-stream activations at specific layers are fed to lightweight affine classifiers for turn-level and token-level persuasion probability estimation (Jaipersaud et al., 7 Aug 2025).

Specialized paradigms include Drift Diffusion Models (DDMs) for agent-level behavioral modeling with stochastic evidence accumulation, mapping drift rate $\nu$ to persuasiveness, boundary separation $a$ to cautiousness, and starting bias $z$ to prior belief (Alvarez-Zuzek et al., 2024). Decision calibration frameworks enforce no-regret best-responses in population-level persuasion games via Lagrangian-constrained optimization over randomized predictors (Tang et al., 22 May 2025). Targeted Persuasion Score (TPS) uses Wasserstein optimal transport between pre- and post-context answer distributions to quantify the contextual influence on LLM behavior (Nguyen et al., 22 Sep 2025).

4. Training, Evaluation Metrics, and Empirical Results

Training pipelines integrate both transparent and black-box architectures, supervised with direct outcome metrics or proxy labels. Common protocols include cross-entropy loss (classification), mean squared error (continuous outcome), and DPO for preference ranking in user profiling (Park et al., 9 Jan 2026). Cross-modal systems utilize alignment loss and reference-model-weighted fusion (Bai et al., 2020). Item Response Theory models leverage hierarchical Bayesian fitting and cold-start linear parameterization (Kornilova et al., 2022).

Evaluation employs macro-F1, AUC, RMSE, accuracy, and Jensen-Shannon distance (strategy distribution alignment). For CMV classification, hybrid RF+LLM models reach up to 82.3% accuracy—approaching the upper bound for human annotation—while theory-only or logistic baselines lag at ~56–57% (Hoang et al., 27 Nov 2025). PersuasionBench models achieve 61–81% accuracy in tweet engagement bin prediction (Singh et al., 2024). Multimodal IPP models deliver reduced MSE (0.011 vs 0.015) in predicting vote swings (Bai et al., 2020). Strategy-aware LLM scoring outperforms monolithic and direct-comparison baselines, yielding higher interpretability and task-wise robustness (Labruna et al., 15 Jan 2026). Personalized profiling confers up to +13.77% F1 gain over static or demographic-only baselines in persuasion prediction (Park et al., 9 Jan 2026).

UI-persuasiveness frameworks (G-FOCUS) substantially increase consistent accuracy (CA) by >10pp over debate-style and zero-shot VLM inferencing (Jeon et al., 8 May 2025). Wasserstein-based TPS surface nuanced sub-threshold contextual persuasion beyond greedy answer flips (Nguyen et al., 22 Sep 2025).

5. Feature Attribution, Interpretability, and Causal Analysis

Feature-level interpretation is paramount for actionable persuasion modeling. Permutation-based variable importance, SHAP analysis, and strategy-wise score attribution are standard (Hoang et al., 27 Nov 2025 Chen et al., 2021). AutoPersuade introduces SUN supervised semi-nonnegative matrix factorization for topic-driven latent decomposition; average marginal component effects estimate per-topic causal impact on out-of-sample persuasiveness (Saenger et al., 2024).

Strategy-wise persuasion scores facilitate direct inspection of rhetorical levers, with topic-wise domain analysis revealing the context-dependent strength of different tactics (e.g. attack/simplification in cultural debates; justification/distraction in religion/ethics) (Labruna et al., 15 Jan 2026). Sequential modeling of interaction prefixes in decision games yields dynamic prediction and highlights causal language patterns (e.g. strong positive words, negation, and topic mentions) (Apel et al., 2020). Personalized profiles demonstrate that flexibility—emphasizing cognitive or emotional traits as the task demands—outperforms static attributes for adaptation (Park et al., 9 Jan 2026).

6. Applications, Limitations, and Prospects

Persuasiveness prediction models underpin applications in online influence detection, misinformation mitigation, UI optimization, recommender systems, and operational assessment of LLM persuasiveness for safety and controllability (Hoang et al., 27 Nov 2025 Jeon et al., 8 May 2025 Singh et al., 2024).

Limitations persist around domain specificity (e.g. CMV forum may not generalize), LLM rating biases (inconsistency and anchoring), and incomplete feature sets (omitted arousal, discourse structure) (Hoang et al., 27 Nov 2025). Agent-based models currently assume population homogeneity and undirected networks, suggesting future expansion to individual-level and community-structured interaction (Alvarez-Zuzek et al., 2024). Computational overhead (e.g., optimal transport in TPS, multi-call inference in G-FOCUS) and finite candidate answer spaces pose scalability constraints (Nguyen et al., 22 Sep 2025).

Best practices include modular, interpretable architectures, targeted data augmentation, LoRA-based fine-tuning, and staged human validation. Forthcoming work will address multi-turn persuasion, integrate richer behavioral and cognitive profiles, extend to cross-lingual scenarios, and refine causal feature discovery and intervention design.

Principal References:

(Hoang et al., 27 Nov 2025, Park et al., 9 Jan 2026, Park, 21 Nov 2025, Kornilova et al., 2022, Bai et al., 2020, Labruna et al., 15 Jan 2026, Alvarez-Zuzek et al., 2024, Jaipersaud et al., 7 Aug 2025, Nguyen et al., 22 Sep 2025, Tang et al., 22 May 2025, Singh et al., 2024, Saenger et al., 2024, Chen et al., 2021, Jeon et al., 8 May 2025, Apel et al., 2020).