PaperlayerView exact revision on arXiv

Source-linked AI summary

CogenPVG: Cognitive-Enhanced Reflective Multi-Agent Framework for Persuasive Video Generation

Yuntian Xiao, Shoulong Zhang, Wenfeng Song, Yan Wang, Yi Chen, Shuai Li

arXiv:2609.25821v1cs.MMcs.AI

TL;DR

Persuasive video generation remains underexplored and challenging, despite the advances in multimodal generation. CogenPVG uses a four-stage, ELM-guided reflective multi-agent framework; the authors report stronger persuasion performance than degraded baselines. Its evaluation by language-model-mediated evaluators warrants further granular investigation.

  • Problem

    Persuasive video generation remains an underexplored, challenging task that combines generative AI with cognitive and psychological theories.

    Source: Introduction · p. 1

  • Method

    CogenPVG uses four sequential stages with ELM-guided generator–critic refinement, applying central-route reasoning and peripheral-route cues.

    Source: Introduction · pp. 1–2

  • Results

    CogenPVG outperforms two degraded baselines and achieves the best persuasiveness performance, with stronger emotion leading, higher argument density, and more effective information delivery.

    Source: Experiment · p. 7

  • Takeaways & Limitations

    Experiments support the effectiveness of combining ELM’s central and peripheral routes for producing highly persuasive videos.

    Source: Conclusion and Future work · p. 8

    The role of language-model-mediated evaluators in assessing video persuasion warrants granular investigation.

    Source: Experiment · p. 8

Abstract

from arXiv · show

Persuasive video generation (PVG) is a valuable yet under-explored research topic. Despite the significant advances in multimodal content generation, AI-empowered automated creation of human-made-like videos with substantial persuasiveness remains a formidable challenge. In this paper, we propose CogenPVG, a novel Cognitive-Enhanced reflective multi-agent framework tailored for Persuasive Video Generation task. Given the topic and stance from the user, we decouple the sophisticated generation process into four sequential stages: argument reasoning, storyboard planning, asset creation, and post-editing, imitating the workflow of human video producers. To ensure high persuasiveness, each stage is equipped with a pair of generator and critic agents, following a reflective refinement scheme grounded in a solid psychological theory of persuasion, the Elaboration Likelihood Model (ELM). In the argument reasoning stage, we generate highly logical and credible reasoning thoughts under the guidance of critical thinking theory, enabling cognitive enhancement via the central route of the ELM. For the other three stages, we generate and optimize multimodal assets, assembling them into a persuasive video guided by theories of heuristics, as the peripheral route of the ELM. To the best of our knowledge, CogenPVG is the first work focused on general persuasive topics, without being confined to commercial purposes. Extensive experiments and comprehensive analysis demonstrate that our framework achieves the best persuasion performance, thereby proving the effectiveness of our proposed multi-agent framework for the PVG task.

Introduction

Persuasive video generation remains underexplored despite its influence and potential across domains, while existing video-generation work largely prioritizes narrative qualities over cognitive impact. CogenPVG applies the Elaboration Likelihood Model (ELM) through four stages and reflective generator–critic refinement.

Introduction

  • Persuasive video generation remains underexplored despite persuasive videos’ influence on attitudes and beliefs and potential in education, news commentary, and misinformation debunking.The paper describes PVG as an interdisciplinary challenge involving generative AI and cognitive science.Source: Introduction · p. 1
  • Existing multi-agent video-generation methods mainly target narrative forms and prioritize character consistency, scene richness, and vivid plots over cognitive impact such as persuasiveness.Source: Introduction · p. 1
  • CogenPVG uses ELM-guided generator–critic refinement across argument reasoning, storyboard planning, asset creation, and post-editing.The framework targets general-topic persuasive video generation, not only commercial purposes.Source: Introduction · pp. 1–2, Table 1 · p. 2, Figure 1 · p. 2
  • The framework follows ELM’s central route for credible, logical information and peripheral route for heuristic content orchestration.The stages iteratively refine generated content against persuasion criteria.Source: Introduction · pp. 1–2

Related Work

Prior persuasive-generation research has focused largely on language and visual persuasion, while direct persuasive video generation remains underexplored. LLM-based video agents largely target narrative-video qualities rather than changing subjective attitudes.

Persuasiveness Modeling

  • Persuasive-generation work models strategies for persuasive dialogue, while visual-persuasion studies assess advertising images, multimodal facial video, and annotated image datasets.Source: Related Work · p. 2
  • Direct generation of persuasive videos remains underexplored despite recent work on visual persuasion and persuasive-outcome prediction.Source: Related Work · p. 2

LLM-based Video Generation

  • LLM-based video-generation agents address narrative tasks such as character consistency, visual consistency, and multimodal video assembly.Source: Related Work · pp. 2–3
  • The stated PVG gap is changing subjective attitudes rather than improving objective visual qualities.Source: Related Work · p. 3

Methodology

CogenPVG maps a topic and stance to a persuasive video through four reflective stages, guided by the Elaboration Likelihood Model (ELM) and related psychological criteria. Argument reasoning follows the central route; storyboard planning, asset creation, and post-editing follow peripheral-route heuristics.

Preliminary for ELM Theory

  • ELM distinguishes careful central-route processing from peripheral cues, and the framework applies critical-thinking criteria centrally and fluency and affect heuristics in the other stages.The method adopts the Paul-Elder model for critical thinking, plus fluency and affect heuristics for peripheral persuasion.Source: Preliminary for ELM Theory · p. 3, Task Definition and Framework Overview · p. 3

Task Definition and Framework Overview

  • CogenPVG maps topic t and stance s to a persuasive video through argument reasoning, storyboard planning, asset creation, and post-editing, each using generator–critic reflective refinement.The stages imitate human video creators; argument reasoning follows the central route, while the other three follow peripheral-route heuristics.Source: Task Definition and Framework Overview · p. 3, Figure 2 · p. 4

Argument Reasoning

  • Argument reasoning builds persuasive thoughts from multi-perspective viewpoints and evidence-grounded reasoning, with critics reviewing relevance, depth, significance, evidential support, logic, and clarity.The generator uses web search, running three queries per viewpoint to gather evidence.Source: Argument Reasoning · p. 3

Storyboard Planning

  • Storyboard planning organizes argument thoughts into chapters and shots, then critics jointly assess chapter structure, emotional settings, speech, and visual-scene coherence.Chapters group focused issues and receive emotions and music descriptions; shots specify motivation, scene, camera trajectory, and speech.Source: Storyboard Planning · pp. 3–4

Asset Creation

  • Asset creation generates video, voiceover, and music, using reviewed first-frame images as an intermediate step before image-to-video generation.Critics assess visual coherence, signaling, and affect; voiceover and music instructions are emotion-centered.Source: Asset Creation · p. 4

Post-Editing

  • Post-editing creates onscreen text and highlighted captions, then combines the video components and applies smooth fades between chapters across video and audio.Critics refine text placement and readability, and select concise, significant caption highlights.Source: Post-Editing · p. 5

Experiment

The experiments compare CogenPVG with adapted video-generation baselines and evaluate its persuasion outcomes through human and LMM assessments. Both evaluator types favor CogenPVG, while route ablations reveal trade-offs and the authors note cost, computational-efficiency, and evaluator-investigation limits.

Experimental Setup

  • The evaluation covers 72 persuasion goals drawn from 18 themes and compares CogenPVG with adapted narrative and industrial baselines plus two degraded variants.The study includes 360 video pairs, each assessed twice by 36 human participants; 432 videos were used for attitude-shift evaluation.Source: Experiment · pp. 5–6

Baseline Comparison

  • Human and LMM evaluations favor CogenPVG over baselines in pairwise persuasiveness, with win rates exceeding loss rates across every comparison.The authors also report the largest positive attitude shift toward the persuasion goal.Source: Table 2 · p. 6, Experiment · p. 6

Ablation Study

  • Route ablations were non-cumulative: removing central critics boosted emotional expressiveness but hurt credibility, while removing peripheral critics improved credibility but weakened fluency and affective potency.Removing both critic agents produced natural, fluent content but weaker emotional arousal and evidential credibility.Source: Experiment · pp. 6–7

Qualitative Analysis

  • Qualitative examples indicate that story-centered methods can dilute viewpoint delivery, while DirectPVG adheres more closely to the goal but has shallow, weakly grounded arguments.VideoGen relies mainly on cost-saving arguments; AutoPVG improves cognitive-load control and attention guidance through visual cues, but lacks critic-agent gains. Figure 4 marks goal-relevant speech and visuals.Source: Experiment · p. 7, Figure 4 · p. 7

Discussion and Limitations

  • The authors identify cost and computational efficacy as limitations of using pretrained generative models, and say LMM evaluation of video persuasion warrants granular investigation.The framework uses off-the-shelf models for each modality.Source: Experiment · pp. 7–8

Conclusion and Future work

CogenPVG applies a reflective multi-agent framework to persuasive video generation, guided by the Elaboration Likelihood Model’s dual routes. The authors identify extending the framework with retrieval-augmented generation and developing other persuasiveness factors as future work.

Conclusion and Future work

  • CogenPVG combines central-route rational reasoning and peripheral-route multimodal cues with generator-critic reflection to produce persuasive videos.The conclusion describes this as expanding persuasive generation from linguistic content to multimodal video.Source: Conclusion and Future work · p. 8
  • Future work includes adding retrieval-augmented generation and developing personal traits and persuasion strategies as persuasiveness factors.The authors say retrieval is intended to further enhance fluidity and credibility.Source: Conclusion and Future work · p. 8

A Implementation Details

The implementation uses dedicated prompts and iterative subtasks, while its critic criteria draw on critical-thinking and peripheral-route heuristics. The system prompts and multi-perspective viewpoint prompt are illustrated in Figures 5 and 6.

System Prompts

  • Dedicated system prompts define each agent’s cognitive-enhanced role, tasks, execution instructions, and constraints or critiquing principles.Figure 5 shows system prompts for the argument-reasoning stage.Source: System Prompts · p. 10, Figure 5 · p. 11

Subtask Executions

  • Subtasks exchange input-bearing user prompts and parse JSON responses, while retaining conversational context through reflective iterations.Figure 6 presents the user-prompt template for multi-perspective viewpoint generation.Source: Subtask Executions · p. 10, Figure 6 · p. 12

B Cognitive-Enhanced Criteria

  • The criteria section connects ELM’s central and peripheral routes to the psychological models used to define critic criteria.The central route concerns rational deliberation; the peripheral route uses heuristics.Source: B Cognitive-Enhanced Criteria · p. 10

The Central Route: Critical Thinking

  • The central route uses the Paul-Elder model to evaluate thought elements, including purpose, information, and viewpoint, against critical-thinking criteria.The criteria include relevance, depth, breadth, significance, accuracy, precision, sufficiency, clarity, and logic.Source: The Central Route: Critical Thinking · p. 10

The Peripheral Route: Heuristics

  • Peripheral-route guidance uses fluency and affect heuristics, translated into criteria for segmentation, conversational language, coherence, signaling, natural voice, spatial contiguity, and affect.The fluency guidance draws on multimedia learning; the affect guidance uses emotion in speech, visuals, and audio.Source: The Peripheral Route: Heuristics · p. 11

C Metrics and Likert Scale

The evaluation uses seven subjective dimensions on a five-point Likert scale, with metrics linked to central- and peripheral-route influences. The scale and generated videos were in Chinese because participants were native Chinese speakers.

C Metrics and Likert Scale

  • Seven subjective dimensions are rated on a five-point Likert scale, with overall persuasiveness designed to directly indicate video persuasiveness.The dimensions are overall persuasiveness, argument quality, credibility, affective potency, comprehension fluency, perception naturalness, and memorability.Source: C Metrics and Likert Scale · p. 12
  • Argument quality and credibility assess central-route thinking, while affective potency, comprehension fluency, and naturalness reflect peripheral-route factors.Memorability measures persistence of persuasion.Source: C Metrics and Likert Scale · p. 12
  • The Likert scale and generated persuasive videos were in Chinese because all participants were native Chinese speakers.Source: C Metrics and Likert Scale · p. 12

D Experimental Results

The authors report that CogenPVG scales with the capabilities of its underlying LMM and generalizes across model families. Cross-model evaluation compared it with DirectPVG and with its GPT-5.4-powered version on 16 sampled persuasion goals.

LLM Sensitivity

  • The cross-model results indicate that CogenPVG scales with underlying LMM capabilities and generalizes across different model families.Source: LLM Sensitivity · p. 12
  • The evaluation compared CogenPVG with DirectPVG using the same LMM and with GPT-5.4-powered CogenPVG, on 16 randomly sampled persuasion goals.The authors cite cost and time overhead as the reason for this sample size.Source: LLM Sensitivity · p. 12

Cost Analysis

Reflective refinement increases API cost and time overhead, primarily because of asset generation. The authors accept this trade-off for an ELM-integrated, training-free approach and identify a lightweight version as a short-term focus.

Cost Analysis

  • Reflective refinement incurs higher API cost and additional time overhead, primarily due to asset generation.Measured time depends on network conditions and provider throughput and may exceed ideal-deployment observations.Source: Cost Analysis · p. 12
  • The authors accept the computational overhead as a trade-off for exploring an ELM-integrated, training-free method to enhance persuasiveness.They identify improving a lightweight version as their primary short-term focus.Source: Cost Analysis · p. 12
  • The cost-analysis table lists input tokens, output tokens, cost, and time as its measures.The supplied passage gives the column headings but no values.Source: Cost Analysis · p. 16

Win-Tie-Loss on Reflection Ablations

The full method achieves the best persuasive effect among the reflection ablations, with pairwise comparisons providing evidence of improvement despite small absolute differences. Visual examples show generated videos for contrasting stances on the same topic.

Win-Tie-Loss on Reflection Ablations

  • The full method achieves the best persuasive effect among reflection ablations, with pairwise comparisons showing improved persuasiveness despite relatively small absolute differences.The authors characterize small absolute gains as expected and acceptable given the difficulty of inducing cognitive impact in persuasion tasks.Source: Win-Tie-Loss on Reflection Ablations · p. 16

Visualization on Contrasting Stances

E Ethical Risk

The authors aim to streamline intelligent video creation, but the method imposes no additional restrictions on user input. They warn this may enable malicious users to synthesize content with high cognitive influence.

E Ethical Risk

  • The method adds no restrictions to user input, which may enable malicious users to synthesize content with high cognitive influence.Source: E Ethical Risk · p. 16
  • The stated objective is to streamline intelligent video creation and explore AI’s role in information production.Source: E Ethical Risk · p. 16
  • Figure 11 presents CogenPVG videos for two contrasting stances, illustrating chapter focus and shot motivation.Source: Figure 11 · p. 17
Loading 2609.25821v1…