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

A four-stage, dual-agent pipeline grounded in the Elaboration Likelihood Model for general-purpose persuasive video synthesis

Yuntian Xiao; Shoulong Zhang; Wenfeng Song; Yan Wang; Yi Chen; Shuai Li· 2026· DOI 10.48550/arXiv.2609.25821

The core problem

Persuasive video generation (PVG) is presented as a valuable yet under-explored research topic. Although multimodal content generation has advanced significantly, the authors argue that AI-empowered automated creation of human-made-like videos with substantial persuasiveness remains a formidable challenge. The paper introduces **CogenPVG**, a Cognitive-Enhanced reflective multi-agent framework tailored for PVG. Given a user-supplied topic and stance, the framework decouples the sophisticated generation process into four sequential stages that imitate the workflow of human video producers: (1) argument reasoning, (2) storyboard planning, (3) asset creation, and (4) post-editing. To ensure high persuasiveness, each stage is equipped with a pair of **generator and critic agents** following a reflective refinement scheme grounded in a psychological theory of persuasion, the **Elaboration Likelihood Model (ELM)**. The authors position CogenPVG as the first work focused on general persuasive topics, without being confined to commercial purposes. The claimed contribution is a framework that achieves the best persuasion performance in extensive experiments and comprehensive analysis, there

Innovation

The abstract reports that extensive experiments and comprehensive analysis demonstrate that the framework achieves the **best persuasion performance**. The authors state that this proves the effectiveness of the proposed multi-agent framework for the PVG task. The evaluation is described as covering general persuasive topics rather than being confined to commercial purposes. No specific quantitative metrics, benchmark names, dataset sizes, or numerical scores are provided in the supplied source text. The reported result is therefore a comparative claim of best persuasion performance supported by extensive experiments and comprehensive analysis, without disclosed numeric values in the available material.

Persuasive video generation (PVG) is presented as a valuable yet under-explored research topic. Although multimodal content generation has advanced significantly, the authors argue that AI-empowered automated creation of human-made-like videos with substantial persuasiveness remains a formidable challenge. The paper introduces **CogenPVG**, a Cognitive-Enhanced reflective multi-agent framework tailored for PVG. Given a user-supplied topic and stance, the framework decouples the sophisticated generation process into four sequential stages that imitate the workflow of human video producers: (1) argument reasoning, (2) storyboard planning, (3) asset creation, and (4) post-editing. To ensure high persuasiveness, each stage is equipped with a pair of **generator and critic agents** following a reflective refinement scheme grounded in a psychological theory of persuasion, the **Elaboration Likelihood Model (ELM)**. The authors position CogenPVG as the first work focused on general persuasive topics, without being confined to commercial purposes. The claimed contribution is a framework that achieves the best persuasion performance in extensive experiments and comprehensive analysis, thereby proving the effectiveness of the proposed multi-agent framework for the PVG task.

The methodology is organized as a staged multi-agent pipeline with reflective refinement. The input is a topic and stance from the user. The output is a persuasive video. The four stages are argument reasoning, storyboard planning, asset creation, and post-editing. Each stage contains a generator agent and a critic agent. The critic provides feedback that the generator uses to refine its output, forming a reflective loop.

Why it matters

The central analytical claim is that decoupling PVG into four sequential stages, each with a generator-critic pair, improves persuasiveness by mirroring the workflow of human video producers. The use of the Elaboration Likelihood Model provides a dual-route account: the central route is operationalized through argument reasoning guided by critical thinking theory, while the peripheral route is operationalized through heuristic-guided multimodal asset generation and optimization in storyboard planning, asset creation, and post-editing. This separation is the framework's main theoretical contribution, linking cognitive enhancement to logical credibility and heuristic cues to multimodal assembly. The authors emphasize that CogenPVG is the first work focused on general persuasive topics, without being confined to commercial purposes. This positions the work as broadening PVG beyond advertising-style applications. The reflective refinement scheme is presented as the mechanism that ensures high persuasiveness at every stage. The paper's stated conclusion is that the framework achieves the best persuasion performance, proving the effectiveness of the proposed multi-agent framework for the PVG task. Limitations, failure cases, and detailed ablation results are not described in the supplied source text. The taxonomy candidates provided (Architecture, Cybersecurity, Network, Cryptography) do not directly correspond to the paper's stated topic of persuasive video generation; the work is best categorized under multimodal generation and multi-agent systems.

Who should read this

CS practitioners and researchers

Opening member content…