Paper deep dive
CogenPVG: Cognitive-Enhanced Reflective Multi-Agent Framework for Persuasive Video Generation
Yuntian Xiao, Shoulong Zhang, Wenfeng Song, Yan Wang, Yi Chen, Shuai Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/23/2026, 5:13:06 AM
Summary
The paper introduces CogenPVG, a novel cognitive-enhanced reflective multi-agent framework designed for Persuasive Video Generation (PVG). It addresses the challenge of creating human-made-like videos with substantial persuasiveness by decoupling the generation process into four sequential stages: argument reasoning, storyboard planning, asset creation, and post-editing. The framework leverages the Elaboration Likelihood Model (ELM) from psychology, utilizing a central route for logical argument reasoning (guided by Critical Thinking theory) and a peripheral route for multimodal asset optimization (guided by heuristics like fluency and affect). Each stage employs a generator-critic reflective mechanism to iteratively refine content, aiming to maximize persuasion performance on general topics.
Entities (14)
Relation Signals (13)
CogenPVG → usestheory → Elaboration Likelihood Model
confidence 98% · CogenPVG... following a reflective refinement scheme grounded in a solid psychological theory of persuasion, the Elaboration Likelihood Model (ELM).
CogenPVG → hasstage → Argument Reasoning
confidence 95% · we decouple the sophisticated generation process into four sequential stages: argument reasoning...
CogenPVG → hasstage → Storyboard Planning
confidence 95% · we decouple the sophisticated generation process into four sequential stages: ... storyboard planning...
CogenPVG → hasstage → Asset Creation
confidence 95% · we decouple the sophisticated generation process into four sequential stages: ... asset creation...
CogenPVG → hasstage → Post-Editing
confidence 95% · we decouple the sophisticated generation process into four sequential stages: ... and post-editing
Argument Reasoning → guidedby → Critical Thinking Theory
confidence 92% · In the argument reasoning stage, we generate highly logical and credible reasoning thoughts under the guidance of critical thinking theory
Argument Reasoning → implements → Central Route
confidence 90% · enabling cognitive enhancement via the central route of the ELM.
Storyboard Planning → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2609.25821v1
- Canonical: https://arxiv.org/abs/2609.25821v1
Trouble viewing inline? Open PDF directly →
Full Text
73,354 characters extracted from source content.
Expand or collapse full text
CogenPVG: Cognitive-Enhanced Reflective Multi-Agent Framework for Persuasive Video Generation Yuntian Xiao 1,2 , Shoulong Zhang 2 , Wenfeng Song 3 , Yan Wang 2 , Yi Chen 4 , Shuai Li 1∗ 1 Beihang University, 2 Zhongguancun Laboratory, 3 Beijing Information Science and Technology University, 4 Beijing Technology and Business University Abstract 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 persua- siveness remains a formidable challenge. In this paper, we pro- pose 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 produc- ers. To ensure high persuasiveness, each stage is equipped with a pair of generator and critic agents, following a reflective re- finement scheme grounded in a solid psychological theory of persuasion, the Elaboration Likelihood Model (ELM). In the argument reasoning stage, we generate highly logical and cred- ible 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 op- timize 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 be- ing 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 videos circulate widely across social networks and media platforms (e.g., TikTok, YouTube, Douyin), wield- ing considerable influence due to their purpose of changing or reinforcing individual attitudes and beliefs (Wittenberg et al. 2021; English, Sweetser, and Ancu 2011; Pressgrove, McKeever, and Collins 2021). As an enriched information medium, persuasive video creation has garnered the inter- est of researchers in commercial advertising (Tucker 2015; Hsieh, Hsieh, and Tang 2012) and holds considerable poten- tial for education (Tuong, Larsen, and Armstrong 2014), news commentary (Nelson, Wood, and Paek 2009), and debunking misinformation (Kim and Chen 2024). However, persuasive video generation (PVG) is a challenging and interdisciplinary task that integrates generative AI technologies with cognitive ∗ Corresponding author. science and psychological theories, yet it remains an under- explored area. Considering its significance for both academic research and daily life, we propose the first practical solution toward AI-empowered persuasive video generation. Prior video generation methods leverage large language models (LLMs) (Achiam et al. 2023; Team et al. 2023; Liu et al. 2024) to build agents for dedicated multi-step planning, followed by generating multimodal assets and assembling into a long-form video (Xiao et al. 2025; Zhuang et al. 2024; Shi et al. 2025; Li et al. 2024a; Xu et al. 2025b; Wu, Zhu, and Shou 2025; Wang et al. 2024), taking advantage of pow- erful generative models for a range of modalities such as image (Zhou et al. 2024; Podell et al. 2023; ByteDance Seed 2025), video clip (Wan et al. 2025; Yang et al. 2024; Seedance et al. 2025) and human speech (Du et al. 2025; Anastassiou et al. 2024). Although existing studies can produce high- quality videos and provide valid multi-agent frameworks, they are mainly confined to narrative-centered video types, such as story visualization (Xu et al. 2025b; Shi et al. 2025; Li et al. 2024a), movies (Wu, Zhu, and Shou 2025), and vlogs (Hou et al. 2025; Zhuang et al. 2024), whose primary concern lies in the consistency of characters, richness of scenes, and vividness of plots, while ignoring their cognitive impact, such as persuasiveness. To tackle the aforementioned challenge, this paper pro- poses a novel cognitive-enhanced reflective multi-agent framework specifically designed for the PVG task, referred to as CogenPVG. We choose to utilize the Elaboration Like- lihood Model (ELM) (Petty and Cacioppo 1986), a well- established persuasion theory in psychology, as guidance in designing the framework. Specifically, the ELM defines two modes of persuasiveness perception: the central route and the peripheral route, corresponding to scenarios that involve deep thinking and superficial perception, respectively. Thus, we derive two principles for enhancing the persuasiveness of generated video: 1) conveying highly credible and logically rigorous information to enhance the central route, and 2) or- chestrating content to utilize heuristics for peripheral route enhancement. As illustrated in Fig. 1, we follow these two principles as high-level guidance to construct our multi-agent framework, which consists of four sequential process stages: argument reasoning, storyboard planning, asset creation, and post-editing. In each stage, we employ a generator-critic re- flective mechanism to optimize the generated content itera- arXiv:2609.25821v1 [cs.M] 22 Sep 2026 Figure 1: Overview of our CogenPVG. We incorporate the well-established dual-route persuading model of ELM for persuasive video generation. ELM Route Psychological Theory Applied Model CogenPVG StageCriteria CentralCritical ThinkingPaul-ElderAR relevance, depth, breath, significance, accuracy, precision, sufficiency, clarity, logic Peripheral Fluency Heuristic Affect Heuristic Multimedia Learning Emotion Contagion SP segmenting, personalization, coherence, affect AC coherence, signaling, voice naturalness, affect PE coherence, spatial contiguity, signaling, affect Table 1: The persuasion theories and applied psychological models used in the four stages of CogenPVG: argument reasoning (AR), storyboard planning (SP), asset creation (AC), and post-editing (PE). tively, ensuring that the outcome of each stage satisfies the persuasion criteria defined in the corresponding psychologi- cal models. The primary contributions of this paper could be summarized as follows: • We introduce a novel cognitive-enhanced reflective multi- agent framework, CogenPVG, for generating videos with high persuasiveness. To the best of our knowledge, Co- genPVG is the first work on the PVG task for the visual persuasion on general topics. • We incorporate the ELM persuasion theory and related applied psychological models throughout our framework by devising a generator-critic reflective mechanism with delicately designed prompts, ensuring the persuasiveness inherited from the persuasion theory. • By conducting extensive experiments and evaluations on metrics that cover persuasiveness comparison, attitude shift, and a series of subjective dimensions, we validate the efficacy of our proposed framework and cognitively enhanced designs. Related Work Persuasiveness Modeling. Modeling the persuasive effects and generating highly persuasive contents are significant re- search topics mainly in the computational linguistic domain. Some works (Zeng et al. 2025; Wang et al. 2019; Yang et al. 2019) model the persuasive strategies from public data and train language models to generate persuasive dialogue. Re- cently, some studies have shifted their focus to visual per- suasion (Liu and Yu 2023; Liu et al. 2019; Ibrahim, Wong, and Shiratuddin 2015; Aghazadeh and Kovashka 2025; Bai et al. 2023; Liu et al. 2022; Kim et al. 2025). CAP (Aghaz- adeh and Kovashka 2025) evaluates the persuasiveness of generated advertising images. M2P2 (Bai et al. 2023) pre- dicts the persuasive outcome by analyzing the multimodal facial video. PVP (Kim et al. 2025) provides a large-scale image datasets with visual persuasiveness annotations and personal characteristics of the viewers. Despite the advance- ments, the task of directly generate persuasive videos is still underexplored. LLM-based Video Generation. With the fast development of LLM models (Achiam et al. 2023; Team et al. 2023; Liu et al. 2024), recent works start to leverage their ad- vanced semantic reasoning merits for video generation, es- pecially in long narrative video cases (Xiao et al. 2025; Shen and Elhoseiny 2025). StoryGPT-V (Shen and Elho- seiny 2025) harnesses the context memorizing ability to address the ambiguous references, yielding better charac- ter consistency. VideoAuteur (Xiao et al. 2025) proposes a LLM-augmented large-scale narrative video dataset and leverages a multimodal LLM as video director to acquire better visual consistency. Another approach is to leverage the subtask planning and tool calling capabilities in LLM via agentic frameworks (Xu et al. 2025b; Wu, Zhu, and Shou 2025; Shi et al. 2025; Hou et al. 2025; Zhuang et al. 2024; Li et al. 2024a; Wang et al. 2024). M-StoryAgent (Xu et al. 2025b) simulates a discussing process for script plan- ning, followed by a series of tool-calling agents for the as- sembly of multimodal videos. Vlogger (Zhuang et al. 2024) hires a LLM as video director to generate vlog shooting plan and actor settings. Anim-Director (Li et al. 2024a) employs a multimodal LLM to automatically orchestrate the entire animation-making process. While extensive works have been conducted in the LLM-based video generation, a notable re- search gap persists in the PVG task, where the primary chal- lenge is to change the subjective attitudes rather than improve the objective visual qualities. Methodology Preliminary for ELM Theory The Elaboration Likelihood Model (ELM) (Petty and Ca- cioppo 1986) is a dual-process model (Kahneman 2013) of persuasion and attitude change (Petty and Cacioppo 1986, 2012), which suggests information can be processed in two distinct cognitive routes: the central route, which involves high proactivity for careful thought of information credibility and logical soundness, and the peripheral route relies on su- perficial cues with inferior cognitive resource requirement or elevated emotional arousal. The ELM indicates two parallel aspects for modeling persuasiveness, where specific applied psychological models of Critical Thinking theory (Facione 1990) and Heuristic theory (Gilovich, Griffin, and Kahne- man 2002) can be utilized for the central route and peripheral route, respectively. Specifically, we adopt the representative and well- established Paul-Elder (P-E) model (Paul and Elder 2019) as the criteria instance of Critical Thinking in the central route. For the peripheral route, we focus on the fluency heuris- tic (Hertwig et al. 2008; Reber and Schwarz 1999) and af- fect heuristic (Slovic et al. 2007) which are instantiated as concrete creation guidelines based on their corresponding applied psychological models (Mayer 2002; Hatfield, Ca- cioppo, and Rapson 1993). Please Refer to Appendix B for detailed explanations of psychological models and the crite- ria employed by the critic agents. Task Definition and Framework Overview We define the persuasive video generation task as a mapping PV G that projects a topic t and a target stance s to a highly persuasive video V , as follows: PV G : (t,s)→ V.(1) The goal of generated video V is to lead the audience to favor s towards t. To this end, we incorporate the ELM persuasion theory and several practical psychological models as guidance in designing the four stages of our proposed framework, as illustrated in Fig. 2. The four stages include argument reasoning (AR), storyboard planning (SP), asset creation (AC), and post-editing (PE), imitating the workflow of human video creators. In each stage, a pair of generator agent G stage and critic agent C stage execute one or multiple subtasks in an iterative and reflective refinement scheme, guaranteeing the effectiveness of ELM. As shown in Tab. 1, the argument reasoning stage generates content in the central route of persuasion, obeying the criteria of critical thinking, and the other three stages follow the peripheral route by instantiating the fluency and affect heuristics. Due to the page limit, we provide the detailed prompts for the generator and critic for each subtask in each stage in Appendix A. Argument Reasoning Given the input topic t and stance s, the argument reason- ing stage aims to generate a set of high-quality thoughts T as persuasive content. Each thought consists of a view- point, several evidences, and a reasoning process generated by two subtasks: multi-perspective viewpoint generation and evidence-grounded reasoning. T = Stage AR (t,s).(2) Multi-Persepective Viewpoints Generation. To satisfy the breadth criterion of the P-E model, we first prompt G AR to analyze the persuasion goal (t,s) from diverse perspectives, including the benefits of the target stance and the potential drawbacks of the opposite stance. AfterG AR forms the initial viewpoints, C AR reviews and critiques them based on more criteria of the P-E model. Specifically, C AR examines the rel- evance between generated viewpoints and persuasion goal, the depth of viewpoints to ensure a thorough exploration of the core issue, and the significance of viewpoints to cover the most crucial aspects. Evidence-grounded Reasoning. G AR conducts a two-step Chain-of-Thought (CoT) (Wei et al. 2022) to collect eviden- tial facts and generate reasoning based on the previously obtained viewpoints. We equip G AR with a web search tool and execute three queries for each viewpoint. Given the raw message returned from the queries, G AR extracts the critical information such as precise digits and authoritative institu- tions, in accordance with the accuracy and precision criteria of the P-E model. Then, it formulates a logically plausible reasoning based on the evidences to enrich each thought. In C AR , evidences and reasoning are reviewed simultane- ously. C AR evaluates whether the evidences are sufficient to substantiate the corresponding viewpoint, while meeting the sufficiency standard of the P-E model. For the reasoning, C AR inspects the logic and clarity to ensure that the gener- ated thoughtsT have no logical fallacies and are expressed unambiguously. Storyboard Planning Anchored in the argument thoughts T and the persuasion goal (t,s) the storyboard planning generator G SP yields a hierarchical storyboardS in global chapter-level and precise shot-level via CoT prompting. S = Stage SP (t,s,T ).(3) In chapter-level planning, G SP semantically clusters the thoughts T into chapters based on different focused issues, aiming to decrease cognitive load by obeying the segmenting principle in the Multimedia Learning theory. Then, to lever- age the affect heuristic, G SP assigns each chapter a proper Figure 2: The framework of our proposed CogenPVG. The text highlighted in red illustrates the cognitive-guided generated content in the thinking quality, cognitive fluency, and emotional contagion. emotion and articulates a music description for each chap- ter. In shot-level planning, G SP further divides each chapter into multiple shots. Each shot consists of a concrete moti- vation corresponding toT , t, and s, as well as descriptions of the visual scene, camera trajectory, and speech lines. To ameliorate, C SP evaluates all chapters and shots in a sin- gle prompting step, enabling the joint optimization of both levels. C SP first assesses the chapter arrangement and the distinctiveness among different chapters to facilitate com- prehension, as proposed in the segmenting criterion. Then it judges the rationality of the emotional setting and guarantees that the music description can effectively convey the target emotion. For the shot-level optimization, C SP sticks to the personalization and coherence criteria to evaluate whether the speech lines exhibit an approachable language style and contribute to the shot motivation, as well as the emotional grounding of visual scene description. Asset Creation Following the chapter and shot descriptions inS, the agents generate and optimize a range of multimodal video assets by three subtasks: video clip V, human voiceover clip H, and background music clipM generation. V,H,M = Stage AC (S).(4) Video Clip Generation. Instead of generating video clips directly via text-to-video models, we use images as an in- termediate modality to fully leverage image-editing capabil- ities. Firstly, we prompt G AC to call the text-to-image tool with an optimized emotion-centered prompt and generate several first-frame image candidates. G AC then selects the best one with the fewest artifacts. The generated image is reviewed by C AC based on the coherence and signaling cri- teria in the Multimedia Learning theory, which suggests the image to contain only necessary elements relevant to the shot motivation and emphasize the core visual element that pow- erfully conveys the motivation. It also provides feedback on the emotional properties of the image based on the affect criteria, as outlined in the Emotional Contagion theory. No- tably, the reflective iteration in first-frame image generation is achieved by switching to the image-editing tool, which directly optimizes the visual contents in image scope rather than prompt scope. Then, G AC creates video clips using the image-to-video tool. Voiceover and Music Clip Generation. We directly utilize the speech lines inS as the voiceover scripts, while prompting G AC with the emotion to infer emotion-centered instructions regarding the speaking tone. Following the same process of rewriting the description to an emotion-centered prompt and generating assets by calling generation tool, G AC creates the chapter-level background music clips. As the primary objective of auditory assets is to amplify the emotional at- mosphere, we adopt the affect criteria inC AC for the subtasks of voiceover and music clip generation to enhance emotional resonance. Figure 3: Distribution of human-based persuasiveness ratings in ablation study. Post-Editing After acquiring the multimodal assets, the post-editing stage aims to enhance the persuasiveness of the final video through two subtasks: onscreen text R and highlighted caption C generation, and then assembles all the video components into a complete multimodal persuasive video V . R,C = Stage PE (V,H).(5) Onscreen Text Generation. According to the signaling cri- terion, we prompt G PE with the first frame image and shot motivation to synthesize an onscreen text with attributes of font color, background color, and pixel coordinates. Based on the coherence and spatial contiguity criteria, C PE pro- vides recommendations to convey the core motivation within the text content concisely and to position it adjacent to the relevant visual element. Regarding the signaling and affect criteria, color settings are adjusted to enhance text readabil- ity and to harmonize the emotional tone with the chapter’s overall sentiment. Highlighted Caption Generation. To explicitly reduce the cognitive load, G PE segments the speech lines into captions where key elements, such as precise statistical numbers and emotive words, are highlighted in different colors and fonts. C PE critiques the captionsC according to the criteria of co- herence and signaling to optimize the selection of the most concise and significant words. After all the video componentsV,H,M,R,C are gener- ated and optimized, we combine shot-level components, and apply a smooth fading transition between chapters for the video and audio channel, to form the final persuasive video V . V = Assemble(V,H,M,R,C).(6) Experiment Data Preparation. Since there is no conventional benchmark for the PVG task, we construct the input topic and stance set from the Personalized Visual Persuasion (PVP) dataset (Kim et al. 2025). After removing two potentially sensitive themes (defense and veteran affairs), we retain 18 themes and select two messages per theme to form our topic set. For each topic, we then assign two stances (i.e., supporting and opposing), resulting in 72 persuasion goals in total. Given that persuasive generation task is inherently more subjective than aesthetic generation task, we construct an evaluation set several times larger than those used in other story visualization works, where about 10 samples are typically considered sufficient to draw solid conclusions (Shi et al. 2025; Zhuang et al. 2024; Wu, Zhu, and Shou 2025; Wang et al. 2025). Implementation Details. We utilize GPT-5.4 for the genera- tor agents and GPT-4o for the critic agents, which are widely regarded as top-tier large multimodal models (LMMs). Given that GPT-4o API lacks native audio processing ability, we ad- ditionally employ Qwen3-Omni (Xu et al. 2025a) to analyze voiceover and music clips. For the equipped tools, we em- ploy Tavily Search (AI n.d.) for evidence searching in G AR , Seedream-4.5 (ByteDance Seed 2025) for image generation and editing, Seedance-1.5-pro (Seedance et al. 2025), Seed- tts-2.0 (Anastassiou et al. 2024), and Suno-4.5 for the gener- ation of video clips, human voiceover and background music clips in G AC respectively. All tools are wrapped as MCP ser- vices and proactively invoked by the agents. Please refer to Appendix A for our detailed implementations, including the system prompts and subtask executions. Baselines. Because CogenPVG is the first framework for persuasive video generation on general topics, it is hard to find existing methods strictly aligned with our task setting for a fully fair comparison. Therefore, we select three of the most relevant agentic long-form video synthesis meth- ods and make minor adaptations for comparative discussion. Specifically, we use 1) M-StoryAgent (Xu et al. 2025b) and 2) Anim-Director (Li et al. 2024a) as representative narrative-centered video generation methods for compari- son, whose input synopses are generated by prompting GPT- 5.4(OpenAI 2026) with our persuasion goals. We further include 3) VideoGen (VideoGen 2026), a prevailing com- mercial platform that synthesizes videos by retrieving con- textually relevant clips from a vast stock media library, as a representative industrial baseline. To better evaluate our method, we propose two degraded versions of our method as additional comparison baselines, where 4) DirectAgent removes all cognitive-enhanced designs and retains only the generators in the last three processing stages, serving as a minimal video synthesis framework, and 5) AutoPVG con- solidates the four stages into a single agent that acomplishes the PVG task by autonomously planning and tool execution. Metrics. Following (Nozue et al. 2025; Qiu et al. 2025; Jin et al. 2024), we evaluate our method and the selected baselines via Win-Tie-Loss persuasiveness comparison and attitude shift scores. Both categories of metrics synergize human-based user studies and LMM-based automated as- sessments. Human evaluation serves as the gold standard for real-world persuasive efficacy, whereas LMM is explorato- rily employed as a potential scalable and objective estimation that mitigates the influence of subjective individual biases. Attitude is quantified using a 5-point Likert scale, ranging from 1 (strongly disagree) to 5 (strongly agree). Average At- titude Shift (AAS) denotes the average increase in human participants’ attitudes. AS i measures the increase of atti- tude value from initial score i. Since human participants’ pre-existing attitudes are non-manipulable, we utilize LMM with predefined initial states for AS i to estimate the per- suasive efficacy with different initial attitudes. To alleviate self-preference bias (Wataoka, Takahashi, and Ri 2024), we employ Gemini-2.5-pro as the automatic evaluator, a model distinct from the LMMs applied in our framework. It is in- Method Persuasiveness (Human) Persuasiveness (LMM) Attitude Shift Win (↑) Tie Loss (↓) Win (↑) Tie Loss (↓) AAS (↑) AS 1 (↑) AS 2 (↑) AS 3 (↑) AS 4 (↑) AS 5 (↑) M-StoryAgent 0.636 0.143 0.221 0.653 0.042 0.305 0.243 2.648 1.944 1.225 0.676 -0.254 Anim-Director 0.788 0.120 0.092 0.901 0.056 0.043 0.127 2.606 1.873 1.099 0.732 -0.254 VideoGen0.688 0.118 0.194 0.915 0.042 0.043 0.257 2.451 1.802 1.014 0.690 -0.155 DirectPVG0.541 0.264 0.195 0.662 0.000 0.338 0.347 3.197 2.229 1.443 1.000 0.000 AutoPVG0.493 0.285 0.222 0.606 0.014 0.380 0.444 3.127 2.142 1.451 0.930 -0.042 CogenPVG------0.604 3.268 2.296 1.592 1.000 0.000 Table 2: Quantitive results of all methods. Our CogenPVG consistently prevails over baseline methods, with a win rate notably higher than the loss rate across every comparison setting and the largest positive attitude shift toward the persuasion goal. Method PersuArgCredAffCompNatMem MHMHMHMHMHMHMH w/o both4.667 3.569 4.639 3.327 3.028 2.957 4.361 3.020 4.986 3.783 4.944 3.652 4.028 2.980 w/o central4.750 3.417 4.618 3.245 2.882 3.060 4.500 3.264 4.944 3.740 4.861 3.560 4.056 3.038 w/o peripheral 4.676 3.167 4.694 3.241 2.917 3.135 4.444 2.981 4.917 3.596 4.861 3.481 4.028 2.852 Ours4.801 3.764 4.722 3.542 3.083 3.278 4.382 3.556 4.971 4.069 4.882 3.944 4.000 3.431 Table 3: Quantitative results of the ablation study, containing both LMM-based (M) and human-based (H) rating results. structed to act as a lay audience without psychology expertise and provide first-person reasoning for the assessment results. To further analyze how different cognitive-enhancement routes affect persuasive outcomes, we additionally adopt a set of subjective rating metrics in the ablation study, using five-point Likert scale to quantify multidimensional influenc- ing factors. Similar to related cognitive research (Gilovich, Griffin, and Kahneman 2002; Petty and Cacioppo 2012), we include overall persuasiveness (Persu), argument quality (Arg), credibility (Cred), affective potency (Aff), compre- hension fluency (Comp), perception naturalness (Nat), and memorability (Mem). Please refer to the Appendix C for their definitions and evaluation prompts. Baseline Comparison. For pairwise persuasion comparison, we combine videos generated by five baseline methods with those generated by CogenPVG to construct 360 video pairs. Each video pair is evaluated twice in random order by 36 human participants, yielding a total of 720 evaluation sam- ples. The Cohen’s κ between the two evaluation rounds is 0.325, which is typical for subjective tasks (Qiu et al. 2025). For LMM-based automatic evaluation, we swap the inter- nal order across two evaluation rounds to mitigate position bias (Li et al. 2024b).A similar protocol is adopted for atti- tude shift evaluation on 432 generated videos, yielding 864 attitude shift samples in the human-based evaluation and 864 samples per initial LMM attitude score. Tab. 2 presents the quantitative results, where win-tie-loss ratios are computed by comparing our CogenPVG against baseline methods. Results demonstrate that our method con- sistently achieves a higher win rate than loss rate, underscor- ing the superior persuasiveness of CogenPVG. Although Di- rectPVG and AutoPVG are degraded variants, they still pre- serve a design tailored exclusively to the nuances of persua- sive video generation, therefore show comparatively lower win rate and higher tie rate than other baselines. We observe that LMM shows a pronounced preference for our method over Anim-Director, indicating that LMM tends to disfavor animated visual styles for persuasion. A similar LMM pref- erence occurs in the comparison against VideoGen, which can be attributed to the inferior alignment between its visual synthesis and vocal commentary. While the magnitude of preference differs, human and LMM-based evaluations con- vergently indicate that our method outperforms the baselines in the pairwise persuasiveness comparison. Our analysis of attitude shift in LMM reveals a ceiling ef- fect: as initial attitude scores ascend, the marginal efficacy of persuasive in further augmenting endorsement diminishes. Notably, baseline methods often struggle to consolidate ex- isting support and may even trigger adverse outcomes. Never- theless, our method along with its degraded variants remains highly effective in eliciting attitude shifts towards the ceiling. The highest average attitude shift in human-based evaluation further demonstrates the effectiveness. Ablation Study. To better understand our generator-critic re- flective mechanism, we conduct three ablation experiments on 36 persuasion goals through removing the critic agent in the central route (i.e., C AR ), the peripheral route (i.e., C SP ,C AC ,C PE ), and both routes, forming w/o central, w/o peripheral and w/o both respectively. The ablated results are shown in Tab. 3. 1) w/o both: The framework tends to generate natural and fluent content when no critic agent is involved, but remains less compelling in emotional arousal and evidential credibility, yielding a suboptimal persuasive effect. 2) w/o central: The generated videos exhibit enhanced emotional expressiveness, albeit at the cost of credibility. 3) w/o peripheral: The results show a noticeable increase in the content credibility, but the poor fluency and affective potency performance indicates that the generated videos are difficult to follow and memorize with light cognitive resources, re- sulting in lower persuasiveness. Interestingly, we observe Figure 4: Qualitative results. We present the generated video snapshots and the corresponding speech lines on the topic Buy a used car with stance supporting as the persuasion goal. The speech contents that contribute to the persuasion goal are rendered in bold, and the ELM-related cognitive-enhanced contents are highlighted in red. The visual centers that closely serve the persuasion goal are marked with red boxes. that the two ELM routes do not contribute to persuasion in a cumulative manner. Instead of incremental gains, the inde- pendent enhancement of single route appears to undermine overall efficacy. This implies that persuasion depends on a synergistic balance, where over-emphasizing one route may trigger cognitive interference in some subjective dimensions, thus lead to diminished persuasiveness. Through grounding the persuasion in more precise and credible information collected from factual evidence and evoking higher affective potency together, CogenPVG achieves superior persuasive efficacy, as validated by both human-based evaluations and LMM-based estimations. Ad- ditionally, we present the distribution of human ratings in Fig. 3, identifying that our method yields persuasiveness scores that are more concentrated in the high-score range. Qualitative Analysis. We present the generation results in Fig. 4. M-StoryAgent and Anim-Director converts the per- suasion goal into story-based character settings and com- pelling plots. These methods prioritize story progression over viewpoint delivery, thereby diluting persuasive impact. In addition, their lack of persuasion sensitivity can lead to the omission of goal-critical content: Anim-Director overlooks the explicitly stated new-versus-used car comparison in the synopsis and concludes merely by recommending the pur- chase of a favored car. We attribute this phenomenon to the intrinsic difference between narrative stories and persuasive scripts, which reveals the notable gap between existing video synthesis tasks and PVG task, validating the necessity of our work. VideoGen adopts a relatively direct presentation strat- egy with limited viewpoint coverage, relying mainly on cost- saving arguments in an explanatory persuasion style, which weakens its overall persuasive efficacy. Benefiting from its structured script orchestration process, DirectPVG demon- strates stronger adherence to the persuasion goal. However, without the dual-route enhancements of ELM, its argumen- tation remains relatively shallow and weakly grounded in factual evidence. Furthermore, it fails to establish informa- tion saliency. By incorporating ELM principles, AutoPVG achieves better cognitive load control and attention guid- ance through visual cues, thereby yielding stronger persua- siveness. Nevertheless, in the absence of critic agents, its cognitive-enhancement gains remain limited. By decoupling the workflow into four stage-specialized agents and applying ELM-guided generator–critic reflective refinement at each stage, our CogenPVG achieves stronger emotion leading, higher argument density, and more effective information delivery. The cognitive enhancements enable our framework to outperform the two degraded baselines and achieve the best persuasiveness performance. Discussion and Limitations. Our framework leverages pre- trained, off-the-shelf generative models for each modality, enabling convenient integration of latest models (Appendix D.1). However, it introduces limitations in terms of cost and computational efficacy (Append D.2). We have also probed the viability of LMMs as neutral evaluators, free from the subjective idiosyncrasies of individual human raters. Nev- ertheless, their role in assessing video persuasion warrants granular investigation. Conclusion and Future work This paper presents CogenPVG, a cognitive-enhanced reflec- tive multi-agent framework specifically tailored for the per- suasive video generation task, expanding persuasive genera- tion research from linguistic domain into multimodal videos. Inspired by the well-established ELM persuasion theory, we govern our framework with two primary principles: the cen- tral route for high-quality rational reasoning, and the periph- eral route for multimodal cues that facilitate mental shortcuts. By incorporating a generator-critic reflective mechanism, CogenPVG effectively amplifies the persuasiveness rooted in solid persuasion psychological models. Extensive experi- ments demonstrate that our framework effectively synergizes the dual routes, resulting in highly persuasive video outputs. Our near-term research goal is to extend our framework with RAG capability, further enhancing the fluidity and cred- ibility by retrieving real-world content. We regard the devel- opment for other persuasiveness factors, such as personal traits and persuasion strategies, as a future long-term work. References Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. Aghazadeh, A.; and Kovashka, A. 2025. Cap: Evaluation of persuasive and creative image generation. In Proceedings of ICCV, 16970–16980. AI, T. n.d. Tavily Search. https://w.tavily.com/. Anastassiou, P.; Chen, J.; Chen, J.; Chen, Y.; Chen, Z.; Chen, Z.; Cong, J.; Deng, L.; Ding, C.; Gao, L.; et al. 2024. Seed-tts: A family of high-quality versatile speech generation models. arXiv preprint arXiv:2406.02430. Bai, C.; Chen, H.; Kumar, S.; Leskovec, J.; and Subrahma- nian, V. 2023. M2P2: Multimodal Persuasion Prediction Using Adaptive Fusion. IEEE Transactions of Multimedia, 25: 942–952. ByteDance Seed. 2025.Seedream 4.5. https://seed.bytedance.com/en/seedream4_5. Du, Z.; Gao, C.; Wang, Y.; Yu, F.; Zhao, T.; Wang, H.; Lv, X.; Wang, H.; Ni, C.; Shi, X.; et al. 2025. Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training. arXiv preprint arXiv:2505.17589. English, K.; Sweetser, K. D.; and Ancu, M. 2011. YouTube- ification of political talk: An examination of persuasion ap- peals in viral video. American Behavioral Scientist, 55(6): 733–748. et al, Y. 2023. Prompt-based Monte-Carlo tree search for goal-oriented dialogue policy planning. In Proceedings of EMNLP. Facione, P. 1990. Critical thinking: A statement of expert consensus for purposes of educational assessment and in- struction (The Delphi Report). Gilovich, T.; Griffin, D.; and Kahneman, D. 2002. Heuristics and biases: The psychology of intuitive judgment. Cambridge university press. Hatfield, E.; Cacioppo, J. T.; and Rapson, R. L. 1993. Emo- tional contagion. Current directions in psychological sci- ence, 2(3): 96–100. Hertwig, R.; Herzog, S. M.; Schooler, L. J.; and Reimer, T. 2008. Fluency heuristic: A model of how the mind exploits a by-product of information retrieval. Journal of Experimental Psychology: Learning, memory, and cognition, 34(5): 1191. Hou, X.; Ma, B.; Cheng, J.; Ren, X.; Yu, K.; Li, W.; Zheng, T.; and Lu, Q. 2025. PersonaVlog: Personalized Multimodal Vlog Generation with Multi-Agent Collaboration and Itera- tive Self-Correction. arXiv preprint arXiv:2508.13602. Hsieh, J.-K.; Hsieh, Y.-C.; and Tang, Y.-C. 2012. Exploring the disseminating behaviors of eWOM marketing: persuasion in online video. Electronic Commerce Research, 12(2): 201– 224. Ibrahim, N.; Wong, K. W.; and Shiratuddin, M. F. 2015. Per- suasive impact of online media: investigating the influence of visual persuasion. In Asia Pacific Conference on Multimedia and Broadcasting, 1–7. Jin, C.; Ren, K.; Kong, L.; Wang, X.; Song, R.; and Chen, H. 2024. Persuading across diverse domains: a dataset and persuasion large language model. In Proceedings of ACL (Volume 1: Long Papers), 1678–1706. Kahneman, D. 2011. Thinking, fast and slow. Farrar, Straus and Giroux. Kahneman, D. 2013. A perspective on judgment and choice: Mapping bounded rationality. Psychology Press. Kim, J.; Han, J.; Choi, D.; Yoon, J.; Lee, E.-J.; and Jo, Y. 2025. PVP: An Image Dataset for Personalized Vi- sual Persuasion with Persuasion Strategies, Viewer Char- acteristics, and Persuasiveness Ratings. arXiv preprint arXiv:2506.00481. Kim, S. J.; and Chen, K. 2024. The use of emotions in con- spiracy and debunking videos to engage publics on YouTube. New Media & Society, 26(7): 3854–3875. Li, Y.; Shi, H.; Hu, B.; Wang, L.; Zhu, J.; Xu, J.; Zhao, Z.; and Zhang, M. 2024a. Anim-director: A large multimodal model powered agent for controllable animation video generation. In SIGGRAPH Asia Conference Papers, 1–11. Li, Z.; Wang, C.; Ma, P.; Wu, D.; Wang, S.; Gao, C.; and Liu, Y. 2024b. Split and merge: Aligning position biases in LLM-based evaluators. In Proceedings of EMNLP, 11084– 11108. Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; et al. 2024. Deepseek-v3 technical report. arXiv:2412.19437. Liu, C.; Dong, Y.; Yu, H.; Shen, Z.; Gao, Z.; Wang, P.; Zhang, C.; Ren, P.; Xie, X.; Cui, L.; et al. 2019. Generat- ing persuasive visual storylines for promotional videos. In ACM International Conference on Information and Knowl- edge Management, 901–910. Liu, C.; and Yu, H. 2023. Ai-empowered persuasive video generation: A survey. ACM Computing Surveys, 55(13s): 1–31. Liu, Z.; Guo, M.; Dai, Y.; and Litman, D. 2022. ImageArg: A Multi-modal Tweet Dataset for Image Persuasiveness Min- ing. In Workshop on Argument Mining, 1–18. Mayer, R. E. 2002. Multimedia learning. In Psychology of learning and motivation, volume 41, 85–139. Elsevier. Nelson, M. R.; Wood, M. L.; and Paek, H.-J. 2009. Increased persuasion knowledge of video news releases: Audience be- liefs about news and support for source disclosure. Journal of Mass Media Ethics, 24(4): 220–237. Nozue, S.; Nakano, Y.; Watanabe, Y.; Takasaki, M.; Moriya, S.; Akama, R.; and Suzuki, J. 2025. Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Com- munication Strategies. In Proceedings of EMNLP: Industry Track, 2287–2312. OpenAI. 2026. GPT-5.4: Advancing Multimodal Reasoning and Video-Audio Understanding. https://openai.com/. Paul, R.; and Elder, L. 2019. The miniature guide to critical thinking concepts and tools. Rowman & Littlefield. Petty, R. E.; and Cacioppo, J. T. 1986. The elaboration likelihood model of persuasion. In Advances in experimental social psychology, volume 19, 123–205. Elsevier. Petty, R. E.; and Cacioppo, J. T. 2012. Communication and persuasion: Central and peripheral routes to attitude change. Springer Science & Business Media. Podell, D.; English, Z.; Lacey, K.; Blattmann, A.; Dockhorn, T.; Müller, J.; Penna, J.; and Rombach, R. 2023. Sdxl: Im- proving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952. Pressgrove, G.; McKeever, R.; and Collins, E. 2021. Effec- tiveness of persuasive frames in advocacy videos. Public Relations Review, 47(4): 102060. Qiu, H.; Zhou, Y.; Venkit, P. N.; Huang, K.-H.; Zhang, J.; Peng, N.; and Wu, C.-S. 2025. MMPersuade: A Dataset and Evaluation Framework for Multimodal Persuasion. arXiv preprint arXiv:2510.22768. Reber, R.; and Schwarz, N. 1999. Effects of perceptual flu- ency on judgments of truth. Consciousness and cognition, 8(3): 338–342. Seedance, T.; Chen, H.; Chen, S.; Chen, X.; Chen, Y.; Chen, Y.; Chen, Z.; Cheng, F.; Cheng, T.; Cheng, X.; et al. 2025. Seedance 1.5 pro: A Native Audio-Visual Joint Generation Foundation Model. arXiv preprint arXiv:2512.13507. Shen, X.; and Elhoseiny, M. 2025. Storygpt-v: Large lan- guage models as consistent story visualizers. In CVPR, 13273–13283. Shi, H.; Li, Y.; Chen, X.; Wang, L.; Hu, B.; and Zhang, M. 2025. AniMaker: Multi-Agent Animated Storytelling with MCTS-Driven Clip Generation. In Proceedings of the SIGGRAPH Asia 2025 Conference Papers, 1–11. Slovic, P.; Finucane, M. L.; Peters, E.; and MacGregor, D. G. 2007. The affect heuristic. European journal of operational research, 177(3): 1333–1352. Team, G.; Anil, R.; Borgeaud, S.; Alayrac, J.-B.; Yu, J.; Sori- cut, R.; Schalkwyk, J.; Dai, A. M.; Hauth, A.; Millican, K.; et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805. Tucker, C. E. 2015. The reach and persuasiveness of viral video ads. Marketing Science, 34(2): 281–296. Tuong, W.; Larsen, E. R.; and Armstrong, A. W. 2014. Videos to influence: a systematic review of effectiveness of video- based education in modifying health behaviors. Journal of behavioral medicine, 37(2): 218–233. VideoGen. 2026. VideoGen: An Automated Video Orches- tration and Content Creation Platform. https://videogen.io/. Wan, T.; Wang, A.; Ai, B.; Wen, B.; Mao, C.; Xie, C.-W.; Chen, D.; Yu, F.; Zhao, H.; Yang, J.; et al. 2025. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314. Wang, J.; Du, Z.; Zhao, Y.; Yuan, B.; Wang, K.; Liang, J.; Zhao, Y.; Lu, Y.; Li, G.; Gao, J.; et al. 2024. Aesopagent: Agent-driven evolutionary system on story-to-video produc- tion. arXiv preprint arXiv:2403.07952. Wang, W.; Zhao, C.; Chen, H.; Chen, Z.; Zheng, K.; and Shen, C. 2025. Autostory: Generating diverse storytelling images with minimal human efforts. IJCV, 133(6): 3083– 3104. Wang, X.; Shi, W.; Kim, R.; Oh, Y.; Yang, S.; Zhang, J.; and Yu, Z. 2019. Persuasion for good: Towards a personalized persuasive dialogue system for social good. arXiv preprint arXiv:1906.06725. Wataoka, K.; Takahashi, T.; and Ri, R. 2024. Self-Preference Bias in LLM-as-a-Judge. In Neurips Safe Generative AI Workshop 2024. Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompt- ing elicits reasoning in large language models. NeurIPS, 35: 24824–24837. Whittlesea, B. W. 1993. Illusions of familiarity. Journal of Experimental Psychology: Learning, Memory, and Cogni- tion, 19(6): 1235. Wittenberg, C.; Tappin, B. M.; Berinsky, A. J.; and Rand, D. G. 2021. The (minimal) persuasive advantage of political video over text. Proceedings of the National Academy of Sciences, 118(47): e2114388118. Wu, W.; Zhu, Z.; and Shou, M. Z. 2025. Automated movie generation via multi-agent cot planning. arXiv preprint arXiv:2503.07314. Xiao, J.; Cheng, F.; Qi, L.; Gui, L.; Zhao, Y.; Lin, S.; Cen, J.; Ma, Z.; Yuille, A.; and Jiang, L. 2025. Videoauteur: Towards long narrative video generation. In ICCV, 19163–19173. Xu, J.; Guo, Z.; Hu, H.; Chu, Y.; Wang, X.; He, J.; Wang, Y.; Shi, X.; He, T.; Zhu, X.; et al. 2025a. Qwen3-omni technical report. arXiv preprint arXiv:2509.17765. Xu, X.; Mei, J.; Li, C.; Wu, Y.; Yan, M.; Lai, S.; Zhang, J.; and Wu, M. 2025b. Mm-storyagent: Immersive narrated sto- rybook video generation with a multi-agent paradigm across text, image and audio. arXiv preprint arXiv:2503.05242. Yang, D.; Chen, J.; Yang, Z.; Jurafsky, D.; and Hovy, E. 2019. Let’s make your request more persuasive: Model- ing persuasive strategies via semi-supervised neural nets on crowdfunding platforms. In Proceedings of NAACL, 3620– 3630. Yang, Z.; Teng, J.; Zheng, W.; Ding, M.; Huang, S.; Xu, J.; Yang, Y.; Hong, W.; Zhang, X.; Feng, G.; et al. 2024. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv:2408.06072. Zeng, D.; Legaspi, R.; Sun, Y.; Dong, X.; Ikeda, K.; Spirtes, P.; and Zhang, K. 2025. Generative Framework for Per- sonalized Persuasion: Inferring Causal, Counterfactual, and Latent Knowledge. In ACM Conference on User Modeling, Adaptation and Personalization, 83–93. Zhou, Y.; Zhou, D.; Cheng, M.-M.; Feng, J.; and Hou, Q. 2024. Storydiffusion: Consistent self-attention for long- range image and video generation. NeurIPS, 37: 110315– 110340. Zhuang, S.; Li, K.; Chen, X.; Wang, Y.; Liu, Z.; Qiao, Y.; and Wang, Y. 2024. Vlogger: Make your dream a vlog. In IEEE/CVF CVPR, 8806–8817. A Implementation Details System Prompts We implement the generator and critic agents in our frame- work by contextual prompting where the cognitive-enhanced functional specialization of agents is defined by their respec- tive dedicated system prompts. We display the main content of these system prompts in Fig. 5, 7, 8 and 9. In the beginning of each system prompt, the main role of the agent is defined by one-sentence description, followed by a series of tagged regions, where <Tasks> encapsulates the primary objective, <Instructions> lists the execution chain and describes the basic requirements for each subtask, and <Important Guide- lines> specifies the detailed constraints for generator agents and related critiquing principles for critic agents. Subtask Executions In each processing stage, we execute the subtasks by sending a user prompt that contains the necessary input arguments to the agents, and receiving the json-structured response that can be parsed to extract the output results of the subtask. The dynamic conversational context of each subtask is main- tained and utilized in the reflective iteration to avoid inef- fective update. After the reflection converges, next subtask is executed samely. We display the user prompt template for multi-perspective viewpoints generation subtask in Fig. 6. For more details on the user prompts designed to trigger subtasks, please refer to our code files in the supplement. B Cognitive-Enhanced Criteria In this section, we introduce the psychological roles of dual routes in ELM theory and demonstrate the specific definition of criteria used in critic agents. The Central Route: Critical Thinking Critical Thinking (Facione 1990) is a fundamental cognitive theory that examines rational deliberation and logical infer- ence, which are inherently compatible with the central route of the ELM. Among its various applied models, we select the most representative and well-established one, the Paul- Elder (P-E) model (Paul and Elder 2019). In the central route of ELM, the P-E model offers a comprehensive set of crite- ria for evaluating the quality of critical thinking, providing a foundational logical framework for assessing the quality of information and its logical plausibility. Specifically, the P-E model analyzes the thinking process by extracting the elements of thought such as purpose, information, and point of view, then evaluates the quality of critical thinking man- ifested in these elements based on a set of criteria, e.g., relevance, significance, precision. In the central route of the ELM theory, we utilize the P-E Model to enhance the critical thinking quality of our persua- sive contents. The carefully crafted criteria are as below: • Relevance: The content of thoughts should be directly related to the core issue, excluding the details that are irrelevant to the problem. • Depth: The content of thoughts should address the root causes of the problem, avoiding the superficial analysis. • Breadth: The content of thoughts should consider from multiple perspectives, avoiding the biases from each sin- gle perspective. • Significance: The content of thoughts should contain the most impactful elements that directly influence the va- lidity and practical outcome, avoid focusing on trivial details. • Accuracy: The content of thoughts should be true, correct with factual evidences, avoiding misleading or unproven assumptions. • Precision: The content of thoughts should provide exact numbers or context with specificity, avoid any kind of vagueness. • Sufficiency: The content of thoughts should consist of evidences, reasoning to adequately justify the intended argument, avoiding any gaps in proof. • Clarity: The content of thoughts should be easy to un- derstand for viewers to grasp the meaning without misin- terpretation. • Logic: The content of thoughts should be consistent and coherent, with no contradictions or gaps in the reasoning process. The Peripheral Route: Heuristics In the peripheral route of the ELM, the heuristics play a piv- otal role in enhancing the persuasiveness. In this work, we particularly emphasize two kinds of heuristics: the fluency System Prompt for 푮 퐀퐑 You are an argument thinker and a supporting material collector. <Tasks> Your goal is to analyze the given topic and stance, and construct several argumentative viewpoints along with supporting evidence materials. </Tasks> <Instructions> •When you are asked to give viewpoints of a given topic and stance, you should: •Analyze the given topic and stance. •Construct arguments with about 3 viewpoints that are persuasive to the audience, affecting them to agree with the stance. •When you are asked to give reasoning process, you should: •Construct and execute web search queries to collect necessary facts and data supporting the viewpoint reasoning. •Extract important information from each search query's result as evidence. •Generate the logical reasoning process which utilizes all the collected evidences to argue for the main topic and stance. •When you are given some feedback suggestions, you should consider seriously reflect on the given feedbacks and complete the task again. </Instructions> System Prompt for 푪 퐀퐑 You are a topic commentator with profound critical thinking skills. <Tasks> Your job is to evaluate the quality of given thoughts strictly according to the critical thinking theory, and give some feedbacks to guide the improvement. </Tasks> <Important Guidelines> •For viewpoints, follow the critical thinking criteria below: •[Relevance] How do the viewpoints relate to the topic and stance? Does it help to convince the audience? •[Depth] Do the viewpoints cover the deeper implications and meanings of the topic, rather than staying at superficial phenomena? •[Breadth] Can the viewpoints cover a broad range of aspects, including both the supporter's perspective and the opponent's perspective? •[Significance] Does the viewpoints cover the most important arguments for the persuasion? Are they incisive enough to hit the nail on the head? •For evidences, follow the critical thinking criteria below: •[Accuracy] Is the evidence credible and reliable, endorsed by sufficiently authoritative institutions or experts? •[Precision] Is the evidence specific and exact, avoiding any kind of vagueness or ambiguity? •[Sufficiency] Is the evidence sufficient to support the viewpoint? •For reasoning process, follow the critical thinking criteria below: •[Clarity] Has the reasoning well illustrated the viewpoint? Use examples if necessary. •[Logic] Is the reasoning process logical and coherent, avoiding any kind of logical fallacies? •After you evaluate for each input content, rate them from 1 to 10 and give some feedbacks to guide the improvement. </Important Guidelines> Figure 5: System prompts in argument reasoning stage. heuristic (Hertwig et al. 2008; Reber and Schwarz 1999) and the affect heuristic (Slovic et al. 2007), which have been proven to have a pronounced impact on imperceptibly al- tering an individual’s attitude (Kahneman 2011; Whittlesea 1993). The fluency heuristic suggests that people tend to be more susceptible to persuasive messages that involve flu- ent cognitive processing experiences (Reber and Schwarz 1999; Whittlesea 1993). To trigger the fluency heuristic, the Multimedia Learning theory (Mayer 2002) offers a practical approach by managing the viewer’s cognitive load through carefully designed perceptual elements, such as message seg- mentation, visual emphasis, and language style. On the other hand, the affect heuristic provides another perspective, where affective feelings strongly influence attitudes and personal judgments (Slovic et al. 2007). To leverage the affect heuris- tic, we introduce the Emotional Contagion theory (Hatfield, Cacioppo, and Rapson 1993), which provides theoretical guidance on eliciting a certain affective feeling by integrating the intended emotion deeply into implicit perception cues, such as speech lines, visual scenes, and auditory sensations. In total, we form several criteria in the peripheral route as follows: • Segmenting: The video content should be well separated into semantic consistent user-paced chunks, reducing the cognitive load for comprehension. • Personalization: The words should be in conversational style rather than formal style, increasing the engagement during cognitive processing. • Coherence: The extraneous words and visual objects that have no contribution to the persuasive goal should be excluded, avoiding redundant cognitive load. • Signaling: The essential contents of the video should be emphasized with highlighting cues to guide the attention of viewers. • Natural Voice: The voiceover should be spoken in a nat- ural human voice, making the persuasive content easier to process. • Spatial Contiguity: The onscreen text should be placed near the related visual elements, making it more efficient to integrate the verbal and visual contents. • Affect: The video content should evoke the emotional res- onance that is relevant to the persuasive goal, enhancing the viewer’s empathy. User Prompt for Multi-Perspective Viewpoints Generation Analyze the input topic and stance displayed below, and give me some related viewpoints. <Input> Topic: topic Stance: stance </Input> % if feedback is defined and feedback % <Feedback> feedback </Feedback> % endif % <Output Format> When you have decided the final answer, the output should ONLY contain a JSON string with schema: “thoughts": ["id": STRING <unique id for each thought>, "viewpoint": STRING <description of the viewpoint>, ...] </Output Format> Figure 6: User prompt for triggering viewpoint generation subtask. C Metrics and Likert Scale Persuasiveness Metrics. The persuasiveness of the gener- ated videos is evaluated via a 5-point Likert scale across 7 subjective dimensions that can effectively measure a wide range of cognitive-level influences of the persuasive videos. We specify the definitions of the related scale items as below: • Overall Persuasiveness (Persu): The video is strongly persuasive on the whole, and can effectively change my attitude. • Argument Quality (Arg): The reasoning and argumen- tation in the video are clearly expressed and logically strong. • Credibility (Cred): The information presented in the video is verifiable and grounded in credible facts. • Affective Potency (Aff): The ambiance crafted in the video is highly conductive to evoke a sense of connection and empathy. • Comprehension Fluency (Comp): It is effortless for the viewers to understand the content and perspectives ex- pressed in the video. • Perception Naturalness (Nat): The audiovisual presen- tation of the multimodal video contents makes a pleasant, fluid and effortless processing experience. • Memorability (Mem): The video leaves a lasting impres- sion and enables the viewers to recall a majority of the presented information. Among these metrics, Persu is designed to indicate the level of persuasiveness directly. Arg and Cred are used to evaluate the quality of critical thinking, referring to the central route of ELM. Aff, Comp, and Nat can reflect the emotional reso- nance and cognitive load, linking to the peripheral route of ELM. An additional metric, Mem, is used to measure the per- sistence of persuasion. These items constitute a Likert scale, where participants are asked to rate each item from 1 to 5 based on their extent of agreement. To better simulate human cognitive processes when using the LMM as a persuasive ef- ficacy estimator, we instruct it to first provide a rationale for each dimension and then output the corresponding score, as demonstrated in Fig. 10. It is worth noting that since the par- ticipants are all native Chinese speakers, we use Chinese as the language for the Likert scale and the generated persuasive videos. D Experimental Results LLM Sensitivity We analyze the sensitivity of the proposed framework to the underlying LLM by evaluating two open-source and one pro- prietary LMMs under our evaluation protocol. We compare CogenPVG against DirectPVG using the same LMM (C vs. D), and compare CogenPVG powered by the given LMM against CogenPVG powered by GPT-5.4 (C vs. C ∗ ). Due to the cost and time overhead, this cross-model evaluation was conducted on 16 randomly sampled persuasion goals. These results demonstrate that our framework scales with the ca- pabilities of the underlying LMM and generalizes effectively across different model families. LMM C vs. DC vs. C ∗ Win Tie Loss Win Tie Loss qwen3-vl-32b 0.563 0.125 0.312 0.250 0.063 0.687 gemma-4-31b 0.500 0.188 0.312 0.188 0.250 0.562 kimi-k2.60.563 0.062 0.375 0.250 0.312 0.438 Cost Analysis We present a cost analysis in the table below. The measured time overhead is affected by network conditions and provider throughput, and may therefore be higher than that observed in an ideal deployment environment. As an inference-time scal- ing approach, reflective refinement inevitably incurs higher api cost and additional time overhead, primarily due to the as- set generation stage. To explore an ELM-integrated, training- free method for enhancing persuasiveness, we accepted a trade-off in computational overhead. The improvement to- wards a lightweight version will be our primary short-term focus. System Prompt for 푮 SP You are an experienced short-form video director who knows the best practices to make a persuasive short-form video. <Tasks> Your job is to direct a short-form video which is targeted for social media platforms, aiming at persuading people to believe the given topic and stance. </Tasks> <Instructions> •The generated storyboard should contain several chapters, each chapter should have a primary focus and plays an important role in the video. •Each chapter shares a common emotional atmosphere (such as calm, sad, happy, depressed, comfort, tender, etc.) and background music description, etc. •Each chapter should contain one or several shots, each shot consists of title, motivation, camera motion, duration, scene description and speech lines, etc. •You should ONLY use the given viewpoints, evidences and reasoning as inspiration for the storyboard design. •You should ONLY use the user input information as reference material and source of factual knowledge for the storyboard design. •When you receive feedback suggestions, you need to follow the suggestion to redo the task and generate better results. </Instructions> <Important Guidelines> •The motivation should clearly indicate the purpose of setting this shot, and point out the ideal effect of this shot. The motivation should be treated as the core guidance for asset creation. •The designed visual scene should be highly consistent with the motivation of the shot, visually compelling and rich in detail. •The speech lines should be concise and consistent with both the scene description and the target stance. •The background music description should contain detailed information about the instrument usage, music style and emotional atmosphere. •The chapters should be properly arranged to form natural connections and coherent transitions. •Each chapter should combine as more shots as possible, only separate the shots where the emotional contrast is very distinct. •The emotional atmosphere of each chapter should be highly aligned with the focus and contents of the chapter's shots. •The overall video design should be attractive and engaging. </Important Guidelines> System Prompt for 푪 SP You are a talented video director who knows how to apply the cognitive bias of Fluency Heuristic and Affect Heuristic to make a video more persuasive. <Tasks> Your job is to assess the quality of the given storyboard strictly according to the cognitive bias of Fluency Heuristic (Multimedia Learning Principles) and Affect Heuristic (Emotional Contagion Theory), and give some feedbacks to guide the improvement. </Tasks> <Important Guidelines> •For the design of shots, follow the Mayer's Multimedia Learning principles and Emotional Contagion Theory below to evaluate the cognitive load: •[Segmenting Principle] Is all the content appropriately segmented into digestible units to fit the viewer's understanding pace? •[Personalization Principle] Is the speech lines' vocabulary and language style conversational and approachable? Avoid technical jargon or complex expressions. •[Coherence Principle] Are the storyboard elements (e.g. shot motivation, speech lines) helpful for the purpose of persuasion, with irrelevant contents excluded? •[Affective Scene] Does the scene description contain visual elements that are consistent with the emotion of the shot? •[Affective Language] Does the vocabulary and language style of speech lines trigger the target emotional feeling of the viewer? •For the design of chapters, follow the Affect Heuristic criteria below to evaluate the emotional engagement: •[Segmenting Principle] Is the chapter division clear and suitable? •[Affective Emotion] Is the emotion type consistent with the implication of chapter focus? •[Affective Music] Can the music design effectively complement the emotional atmosphere of the chapter? •After you evaluate for each shot and chapter, rate them from 1 to 10 and give some feedbacks to guide the improvement. </Important Guidelines> Figure 7: System prompts in storyboard planning stage. System Prompt for 푮 퐀퐂 You are an experienced AIGC model user and musician who knows the best practices to generate artistic assets by calling multimedia AIGC models. <Tasks> Your job is to call related generation tools to produce multimedia assets from input descriptions. </Tasks> <Instructions> •You should consider what makes a text description suitable for conditioning the corresponding generative model. •Then you should rewrite the input description to increase the generation quality for each asset. •For image, video, music generation, use the specially rewritten text prompts to call related generation tools to produce the asset. •You should call the generation tools for asset generation. •When you receive feedback suggestions, you need to follow the suggestion to generate better results. </Instructions> <Important Guidelines> •You should keep the core information and content of the input descriptions NOT changed while rewriting. •For image generation: •You should rewrite the scene description to a structured text prompt for calling image generation tools. •The text prompt is expected to contain but not limited to the appearance, color, artistic style. •The words of rewritten text prompt should be simple and plain. Remove the argumentation and keep only the descriptions related to the image content. •You should analyze the scene description's static initial setting, and only based on the initial setting of the scene, rewrite the prompt for image generation task. •If you receive feedback in image generation, you should translate the feedback into clear image editing instructions as prompt, avoid repeating the prompts used for initial image generation. •For video generation: •You should rewrite the scene description and static image description to a structured text prompt for calling video generation tools. •The text prompt is expected to contain but not limited to the dynamic motion, appearance change, color reference, artistic style. •The words of rewritten text prompt should be simple and plain. Remove the argumentation and keep only the descriptions related to the video content. •You should analyze the scene description's dynamic behavior process, and based on the dynamic behavior process, rewrite the prompt for video generation task. •For speech generation: •You should analyze the text content and input emotion to generate proper instruction for calling the speech generation tool. •For music generation: •You should use your knowledge of music theory to create descriptions of specific instrumental techniques and musical structures that can effectively and significantly express the input emotion. •You should explicitly point out the target emotion in the rewritten prompt. •The music duration should be around 30 seconds, the musical expression should be completed within this short duration. •Make the music generation prompt as short and concise as possible, only contains the description related to the music content. </Important Guidelines> System Prompt for 푪 퐀퐂 You are a persuasive multimedia content critic who knows how to apply the cognitive bias of Fluency Heuristic and Affect Heuristic to make multimedia content more persuasive. <Tasks> Your job is to assess the quality of the given visual/audio content strictly according to the Mayer's Multimedia Learning Principles and Emotional Contagion Theory, and give some feedbacks to guide the improvement. </Tasks> <Important Guidelines> •For the generated image and video, follow the Mayer's Multimedia Learning Principles and the Emotional Contagion Theory to evaluate the cognitive load and emotional engagement: •[Coherence Principle] Are all the visual elements relevant to the persuasive motivation? Should remove redundant elements that are weakly linked to the persuasive motivation. •[Signaling Principle] Has the core visual element that best expresses the persuasive motivation been emphasized? Should edit the image or video to make the core element more prominent. •[Affective Scene] Does the visual contents effectively and significantly express the input emotion, so as to enhance the persuasiveness of the content? •For the generated music and speech, follow the Emotional Contagion Theory to evaluate the emotional engagement: •[Voice Principle] Is the human voice of generated speech natural and personable, making the audience feel like they are listening to a real person with empathy? •[Affective Voice] Does the generated speech voice show the target emotion, can it trigger the same emotion of the viewers? •[Affective Music] Do the music melody and rhythm effectively and significantly express the input emotion? •After you evaluate for each input content, rate them from 1 to 10 and give some feedbacks to guide the improvement. </Important Guidelines> Figure 8: System prompts in asset creation stage. System Prompt for 푮 퐏퐄 You are a video editor who knows how to compose multimedia assets into a coherent video with lower cognitive load and higher cognitive engagement. <Tasks> Your job is to post-process the input raw multimedia assets for the purpose of video editing, and edit them into a short-form video ready for media distribution. </Tasks> <Instructions> •You should segment the speech lines into chunks to be displayed as captions according to the semantic structure of each line. •You should design an on-screen text clip for each shot to highlight the core information. </Instructions> <Important Guidelines> •For caption segmentation: •Each segment should contain no more than 8 words. •The speech line segmentation should avoid splitting a phrase into multiple captions. •As the captions displayed on the screen do not contain pause-indicating punctuation, you should only use spaces to indicate pauses instead of these marks. •For each caption chunk, you need to recognize and highlight the important words or phrases which express the core information of the caption and should be emphasized. •You can not throw words away during caption segmentation, combining all caption chunks together should reform the input speech line. •For on-screen text design: •You should analyze the visual scene, and recognize one key object which is most impressive and best conveys the motivation of the shot. •The text content should highlight the core message conveyed by the object with around 4 words, no more than 6 words. •The text content should be natural and in conversational style. •The text should be placed near the object and avoid occluding the object or overlaying off screen. Assuming that the image size is 1000*1000, generate the coordinate of the left-top corner of the text. •The RGBA color (0-255 for each channel) of the text and its background should be harmonious and consistent with the overall color style of the image, and should make the text easy to read. </Important Guidelines> System Prompt for 푪 퐏퐄 You are a video editor who knows how to design and arrange the visual contents to form a harmonious video with lower cognitive load and higher engagement. <Tasks> Your job is to assess the quality of the on-screen text in the given image and highlighted captions strictly according to the cognitive bias of Fluency Heuristic and Affect Heuristic, and give some feedbacks to guide the improvement. </Tasks> <Important Guidelines> •For on-screen text, follow the Mayer's Multimedia Learning Principles to decrease the cognitive load and the Affect Heuristic to increase the emotional engagement: •[Coherence Principle] Is the text content concise and clear to convey the key message of the object and the shot motivation? •[Spatial Contiguity Principle] Is the position of the text appropriate, neither obstructing important objects nor too far from the relevant object? •[Signaling Principle] Is the color of the text and background of the on-screen clip aesthetic and fitting, and making the text content easy to distinguish as well? •[Affect Heuristic] Is the color theme of the text and background suitable for the emotional atmosphere of the shot? •For captions, follow the criteria below: •[Coherence Principle] Is the highlighted keywords concise and consists of the shortest core information? •[Signaling Principle] Is the selected keywords highly relevant to the persuasive content and need to be highlighted? •After you evaluate for each input content, rate them from 1 to 10 and give some feedbacks to guide the improvement. </Important Guidelines> Figure 9: System prompts in post-editing stage. Figure 10: Prompt for LMM-based evaluation on subjective dimensions. Method Input Tok Output Tok Cost ($) Time (min) DirectPVG 6.34k2.44k0.8918.2 CogenPVG 60.21k 6.52k2.2133.7 Win-Tie-Loss on Reflection Ablations In Table 2 of the main paper, we present the pointwise rating values from the ablation study on critic-agent reflection. We found that although our method achieves the best persuasive effect, its absolute differences from the other variants may appear relatively small. Given the difficulty of inducing cog- nitive impact in persuasion tasks, a relatively small absolute gain is expected and acceptable(Jin et al. 2024; et al 2023). We also converted human scores into pairwise comparisons between the full method and each ablation, providing clear evidence of improved persuasiveness for our method. Ours v.s.Win TieLoss w/o both0.458 0.292 0.250 w/o central0.472 0.361 0.167 w/o peripheral 0.570 0.236 0.194 Visualization on Contrasting Stances Additionally, we provide the visual results generated by our CogenPVG framework for two contrasting stances on the same topic Buy a used car with the main focus of each chapter and the motivation of each shot in Fig. 11. As observed, CogenPVG can search for precise evidences and construct reasonable chapter organizations for either side of the two stances and produce compelling persuasive videos. This demonstrates the flexibility of proposed framework re- gardless of the input stance. E Ethical Risk Our objective is to streamline the workflow for intelligent video creation and explore the potential of AI in the field of information production. However, as purely academic re- search, our method does not provide additional restrictions to user input, which may enable malicious users to synthesize content with high cognitive influence. (a) Topic: Buy a used car, Stance: Supporting (b) Topic: Buy a used car, Stance: Opposing Figure 11: The generated persuasive videos from our CogenPVG framework with two contrast stances. To better illustrate the persuasive effect, we display the focus of each chapter and the motivation of each shot.