GRID：将健康领域网红的言论建模为论证网络以评估可信度

# GRID: Modeling Health Influencer Discourse as Argument Networks for Credibility Assessment
Source: [https://arxiv.org/html/2610.27149](https://arxiv.org/html/2610.27149)
###### Abstract

Health\-related discussions on social media are increasingly shaped by influencers who offer diverse, and often competing perspectives on the same issue\. Yet, such content is typically evaluated in isolation, without considering the broader discourse in which it appears\. We introduceGRID\(GroundedReasoning overInterconnectedDiscourse\), a framework for evaluating influencer content through collective inference over the surrounding discourse\.GRIDrepresents each video as an argumentation graph connecting claims to its central message, ortakeaway, links related claims and takeaways across videos, grounds claims in scientific evidence, and jointly infers credibility over the resulting network\. We evaluateGRIDon an end\-to\-end health\-relatedcontent validationtask using569569influencer videos grounded in expert assessments\. Our results show that accounting for the broader discourse improves influencer content validation\. Finally, we useGRID\-inferred credibility to demonstrate how argumentative strategies and moral framing differ between low\- and high\-credibility content\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2610.27149v1/figure1.png)Figure 1:Treating influencer content in isolation may be insufficient: a claim can be well\-justified within a video yet still be incorrect\. Evaluating it in isolation also ignores the broaderdistributed information environment, where related influencers may provide supporting or competing perspectives\. We contend that incorporating this surrounding discourse enables a more holistic evaluation of the target content\.Social media influencers increasingly shape how people understand and act on important issues such as health, finance, and legal matters[Fabris et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib20);[Hayes and Ben\-Shmuel \(2024\)](https://arxiv.org/html/2610.27149#bib.bib21);[Powell and Pring \(2024\)](https://arxiv.org/html/2610.27149#bib.bib26)\. Their videos often rely on using persuasive narratives[Lee and Theokary \(2021\)](https://arxiv.org/html/2610.27149#bib.bib45);[Georgakopoulou and Garcés\-Conejos Blitvich \(2024\)](https://arxiv.org/html/2610.27149#bib.bib46), combining factual claims with lived experience, recommendations, and carefully framed justifications[Devos et al\. \(2022\)](https://arxiv.org/html/2610.27149#bib.bib28)\. Such framing selectively emphasizes particular aspects of an issue to promote a product, advocate a viewpoint, or reinforce a broader philosophy\. Consequently, information about an issue is often fragmented across multiple actors, each contributing different, and sometimes competing, perspectives[Barzilai\-Nahon \(2008\)](https://arxiv.org/html/2610.27149#bib.bib35), collectively contributing to the formation of adistributed information environment\. For instance, in Figure[1](https://arxiv.org/html/2610.27149#S1.F1), the target influencer video emphasizessunlight’s role in vitamin D productionto justify skipping sunscreen, while others highlight the UV exposure risks or ways to maintain vitamin D without unprotected sun, each contributing a different part of the broader picture\. Viewed in isolation, the target video may appear convincing because it relies on avalidclaim while omitting other information relevant to its recommendation\. Therefore, understanding and evaluating such content requires reasoning not only about how claims support a video’s message, but also how that message relates to the surrounding discourse\.

This view is motivated by the theory ofExplanatory Coherence[Thagard \(1989\)](https://arxiv.org/html/2610.27149#bib.bib1), which holds that hypotheses shouldnotbe evaluated in isolation, but in relation to available evidence and competing explanations\. Influencer discourse exhibits a similar structure: a video’s central message, ortakeaway[Alhindi et al\. \(2018\)](https://arxiv.org/html/2610.27149#bib.bib4);[Sundriyal et al\. \(2023\)](https://arxiv.org/html/2610.27149#bib.bib32);[Nakshatri et al\. \(2026\)](https://arxiv.org/html/2610.27149#bib.bib2), is supported by claims within the video, while claims and takeaways expressed across related videos may reinforce or contradict one another\. Evaluating these signals jointly therefore provides a principled way to assess how well a video’s message is supported within the broader discourse\.

Existing work addresses different aspects of influencer content understanding and evaluation through persuasion modeling[Wang et al\. \(2019\)](https://arxiv.org/html/2610.27149#bib.bib41), argumentation mining[Mochales and Moens \(2011\)](https://arxiv.org/html/2610.27149#bib.bib5);[Lawrence and Reed \(2019\)](https://arxiv.org/html/2610.27149#bib.bib39), structured reasoning[Freedman et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib16);[Zhu et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib37), and claim verification[Dmonte et al\. \(2026\)](https://arxiv.org/html/2610.27149#bib.bib33)\. However, these approaches typically analyze content independently, without accounting for its relationship with the surrounding discourse\. In parallel, network\-based approaches to news veracity detection have demonstrated the value of modeling relationships among users, posts, and claims[Monti et al\. \(2019\)](https://arxiv.org/html/2610.27149#bib.bib19);[Mehta et al\. \(2022\)](https://arxiv.org/html/2610.27149#bib.bib17)\. However, these approaches do not explicitly reason over theargumentativerelationships within influencer discourse—how claims and conclusions expressed across different videos support or contradict one another\.

Building on these insights, we introduceGRID\(GroundedReasoning overInterconnectedDiscourse\), a framework for understanding and evaluating health\-related influencer discourse\. Given atargetvideo, we retrieverelatedinfluencer videos to situate it within the surrounding discourse\. We represent each video as an argumentation graph connecting its claims to itstakeaway, capturing how the influencer justifies the central message rather than treating claims independently\. We then connect related claims and takeaways across videos to model theinterconnected discourse, andgroundits claims in external scientific evidence\. Finally, wereasonjointly over this evidence\-grounded network using probabilistic inference[Beltagy et al\. \(2014\)](https://arxiv.org/html/2610.27149#bib.bib12), allowing evidence and argumentative relations to propagate across videos\. Together, these components move beyond independent validation towardcollective inferenceover the broader information environment\.

For evaluation, we formulate an end\-to\-endcontent validationtask that requires reasoning over influencer content to classify it ascorrect,partially correct, orincorrect\. Our dataset contains569569health\-related influencer videos grounded in expert assessments fromScienceFeedback111[science\.feedback\.org/health\-feedback](https://science.feedback.org/health-feedback/)\.GRIDachieves67\.667\.6Macro\-F1 on this task, outperforming the strongest baseline by\+14\+14F1 points\.

The resulting representation also enables us to study broader patterns in how influencers construct their messages\. We characterize discourse along complementary dimensions: \(a\) argumentative strategies[Walton et al\. \(2008\)](https://arxiv.org/html/2610.27149#bib.bib14)used to support a video’stakeaway; \(b\) the moral values[Haidt and Graham \(2007\)](https://arxiv.org/html/2610.27149#bib.bib15)invoked in its claims\. We use these dimensions to compare communication patterns across differing levels of credibility, illustrating how our framework can support analysis of influencer discourse beyond content validation alone\.

In summary, our contributions are:

- •We proposeGRID, a framework for understanding and evaluating influencer content in the context of its surrounding discourse; jointly modeling within\-video argumentation, cross\-video interactions, and external scientific evidence\.
- •We show that collective inference substantially improves content validation over competing baselines \(by\+14\+14F1\-points\)\.
- •We useGRIDto characterize argumentative and moral framing across credibility levels, finding thatauthority/subversionappears at similar overall rates but in different topics: more often invaccine and maskcontent among low\-credibility videos, and intreatment and dietcontent among high\-credibility videos\.

## 2Related Work

Prior work relevant to understanding and evaluating influencer content has approached the problem from several complementary directions, including persuasive discourse analysis, evidence\-based verification, and structured argumentation\.

![Refer to caption](https://arxiv.org/html/2610.27149v1/work3-framework-output-highlights.png)Figure 2:GRIDcollects related influencer videos, connects their claims and takeaways in an networked argumentation graph, incorporates scientific evidence, and jointly infers per\-video credibility\. Note thatGRIDoperates on influencer video transcripts\.Influencer and persuasive discourse\.Prior work has examined persuasion and rhetorical strategies in online discourse[Wang et al\. \(2019\)](https://arxiv.org/html/2610.27149#bib.bib41);[Lee and Theokary \(2021\)](https://arxiv.org/html/2610.27149#bib.bib45);[Liu and Zheng \(2024\)](https://arxiv.org/html/2610.27149#bib.bib40);[Akhynko et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib42);[Bozdag et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib36), showing how messages are strategically framed to influence audiences\. Related work also highlights the importance of identifying the central assertion expressed in noisy social\-media content before verification[Sundriyal et al\. \(2023\)](https://arxiv.org/html/2610.27149#bib.bib32)\. While these studies characterize how messages are expressed or received, we focus on validating influencer content\.

Evidence\-grounded and argumentative verification\.Fact\-verification systems commonly evaluate individual claims against retrieved evidence[Alhindi et al\. \(2018\)](https://arxiv.org/html/2610.27149#bib.bib4);[Atanasova et al\. \(2020\)](https://arxiv.org/html/2610.27149#bib.bib43);[Zheng et al\. \(2024\)](https://arxiv.org/html/2610.27149#bib.bib44), including scientific and health evidence[Wadden et al\. \(2020\)](https://arxiv.org/html/2610.27149#bib.bib30);[Kotonya and Toni \(2020\)](https://arxiv.org/html/2610.27149#bib.bib47);[Pradeep et al\. \(2021\)](https://arxiv.org/html/2610.27149#bib.bib29);[Vladika et al\. \(2024\)](https://arxiv.org/html/2610.27149#bib.bib48)\. Recent LLM\-based systems further combine claim decomposition, retrieval, and verification[Dmonte et al\. \(2026\)](https://arxiv.org/html/2610.27149#bib.bib33);[Li et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib8);[Wang et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib31)\. Complementary approaches introduce explicit argumentative structure into verification: ArgLLM[Freedman et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib16)uses argumentation frameworks to support interpretable verification; ArgRAG[Zhu et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib37)organizes retrieved evidence through support and attack relations; MArgE[Ng et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib38)structures evidence from multiple LLMs into argument trees; and TAIGR[Nakshatri et al\. \(2026\)](https://arxiv.org/html/2610.27149#bib.bib2)models influencer content as an argumentation graph grounded in scientific evidence\. These methods, however, primarily reason within individual claims or pieces of content\. We extend structured verification to the surrounding discourse by explicitly connecting claims and takeaways across related videos\.

Reasoning over broader context\.Early work in cognitive science and psychology shows the importance of jointly reasoning over interdependent information for making informed decisions[Thagard \(1989\)](https://arxiv.org/html/2610.27149#bib.bib1);[Thagard and Verbeurgt \(1998\)](https://arxiv.org/html/2610.27149#bib.bib24);[Fraidin \(2004\)](https://arxiv.org/html/2610.27149#bib.bib25)\. Similarly, research in communication emphasizes understanding discourse through competing perspectives[Mercier and Landemore \(2012\)](https://arxiv.org/html/2610.27149#bib.bib27)\. Empirical studies demonstrate the utility of broader context in argumentative discourse[Durmus et al\. \(2019\)](https://arxiv.org/html/2610.27149#bib.bib34);[Zhao et al\. \(2021\)](https://arxiv.org/html/2610.27149#bib.bib22)and misinformation detection, particularly through social interactions and propagation networks[Monti et al\. \(2019\)](https://arxiv.org/html/2610.27149#bib.bib19);[Han et al\. \(2020\)](https://arxiv.org/html/2610.27149#bib.bib18);[Mehta et al\. \(2022\)](https://arxiv.org/html/2610.27149#bib.bib17)\. Our work instead models theargumentative relationshipsamong claims and takeaways across influencer videos, grounding them in scientific evidence and jointly inferring credibility over the discourse network\.

## 3GRID

Motivated by the theory of explanatory coherence, which holds that hypotheses should be assessed in relation to competing explanations and available evidence[Thagard \(1989\)](https://arxiv.org/html/2610.27149#bib.bib1),GRIDvalidates a target video in the context of related influencer content and scientific evidence\. Figure[2](https://arxiv.org/html/2610.27149#S2.F2)summarizes its four stages: \(a\) retrieve related videos, \(b\) construct a networked argumentation graph, \(c\) ground its claims in scientific evidence, and \(d\) jointly infer video credibility through constrained optimization\.

As input,GRIDoperates on video transcriptions generated using Whisper[Radford et al\. \(2023\)](https://arxiv.org/html/2610.27149#bib.bib50)\. The following subsections describe each stage, with implementation details in Appendix[C\.1](https://arxiv.org/html/2610.27149#A3.SS1)\.

### 3\.1Influencer Content Collection

Letxix\_\{i\}denote the content of a target video by influencerii\. To evaluatexix\_\{i\}in context[Thagard \(1989\)](https://arxiv.org/html/2610.27149#bib.bib1), we obtain related influencer videos that may offer supporting claims or alternative explanations\. Following[Alhindi et al\. \(2018\)](https://arxiv.org/html/2610.27149#bib.bib4);[Nakshatri et al\. \(2026\)](https://arxiv.org/html/2610.27149#bib.bib2), we guide this search by the video’stakeaway,txit\_\{x\_\{i\}\}: the core recommendation or message that the audience is meant to retain and act upon\. Focusing on thetakeawaylets us find videos about the main message itself, even when they express it through different claims\. We use an LLM to identifytxit\_\{x\_\{i\}\}and two associated keywords fromxix\_\{i\}\. Using the TikTok Research API, we perform a keyword\-based search and filter the results for relevance to obtain the set of related videos𝒴=\{ai,bj,ck,…\}\\mathcal\{Y\}=\\\{a\_\{i\},b\_\{j\},c\_\{k\},\\ldots\\\}\. In the example in Figure[2](https://arxiv.org/html/2610.27149#S2.F2), the target takeaway is todrink apple cider vinegar \(ACV\) daily to lose weight\. The retrieved videos may offer contrasting perspectives, such as the claims thatresearch has not proved that ACV causes meaningful weight lossor thatdrinking ACV causes nausea\.

### 3\.2Modeling Networked Discourse

Letℰ=\{xi\}∪𝒴\\mathcal\{E\}=\\\{x\_\{i\}\\\}\\cup\\mathcal\{Y\}denote the ecosystem with the target video and its related videos\. We therefore modelℰ\\mathcal\{E\}in two steps\. Following prior work in claim\-premise modeling and argumentation mining[Mochales and Moens \(2011\)](https://arxiv.org/html/2610.27149#bib.bib5);[Sviridova et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib6);[Freedman et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib16), we first map each video to an argumentation graph that represents the claims used to support its takeaway \(see Figure[2](https://arxiv.org/html/2610.27149#S2.F2)–b\.1\)\. We then connect related claims across these individual graphs to capture how videos support or challenge one another\. The resulting networked argumentation graph \(Figure[2](https://arxiv.org/html/2610.27149#S2.F2)–b\.2\) represents both reasoning within videos and relations across videos, allowing us to evaluatexix\_\{i\}in the broader discourse context\.

#### 3\.2\.1Individual Argumentation Graphs

For each videoz∈ℰz\\in\\mathcal\{E\}with takeawaytzt\_\{z\}, we use an LLM to extract claims𝒞z=\{c1,…,cm\}\\mathcal\{C\}\_\{z\}=\\\{c\_\{1\},\\ldots,c\_\{m\}\\\}from its transcript and caption\. For each claim, we also identify itsspecificity\(high or low\) andsource invoked\(institutional/scientific evidence, experiential evidence, social testimony, or unsupported assertion\)\.Specificitycaptures whether a claim makes concrete, bounded assertions, whilesourcecaptures the evidentiary basis invoked\. These attributes characterize how a claim is presented, and provide priors for trust inference \(Section[3\.2\.4](https://arxiv.org/html/2610.27149#S3.SS2.SSS4)\)\.

To capture how these claims jointly justify the video’stakeaway, we organize them into a directed argumentation graphGz=\(Vz,Az\)G\_\{z\}=\(V\_\{z\},A\_\{z\}\), whereVz=\{tz\}∪𝒞zV\_\{z\}=\\\{t\_\{z\}\\\}\\cup\\mathcal\{C\}\_\{z\}\. Edges represent either direct support from a claim to the takeaway,\(c,tz\)\(c,t\_\{z\}\), or support between claims,\(c,q\)\(c,q\)\. Modeling both allows us to capture reasoning that reaches the takeaway directly or through intermediate claims\.

Graph Construction:For each claimcc, we first consider a candidate edge\(c,tz\)\(c,t\_\{z\}\)\. An LLM classifies the relation as \{direct support,weak support,none\}\. Using token log\-probabilities[Kadavath et al\. \(2022\)](https://arxiv.org/html/2610.27149#bib.bib7), we score the edge asw⁡\(c,tz\)=P⁡\(direct∣c,tz\)\+0\.5​P​\(weak∣c,tz\)w\(c,t\_\{z\}\)=P\(\\mathrm\{direct\}\\mid c,t\_\{z\}\)\+0\.5P\(\\mathrm\{weak\}\\mid c,t\_\{z\}\)and retain it when the score exceeds a threshold\. We then expand the graph recursively\. Each claim connected to the takeaway becomes a new target node, and the remaining unattached claims are evaluated as potential supporters of that target using the same procedure\. Expansion continues until no connected claim remains to be explored\.

#### 3\.2\.2Networked Argumentation Discourse

Individual graphs capture how each influencer justifies their takeaway, but remain disconnected from arguments expressed in other videos\. To represent this broader discourse, we connect them into a networked argumentation graphGℰ=\(Vℰ,Aℰ\)G\_\{\\mathcal\{E\}\}=\(V\_\{\\mathcal\{E\}\},A\_\{\\mathcal\{E\}\}\), whereVℰ=⋃z∈ℰVzV\_\{\\mathcal\{E\}\}=\\bigcup\_\{z\\in\\mathcal\{E\}\}V\_\{z\}andAℰA\_\{\\mathcal\{E\}\}contains the within\-video edges⋃z∈ℰAz\\bigcup\_\{z\\in\\mathcal\{E\}\}A\_\{z\}plus cross\-video edges encoding equivalence, support, or contradiction\.

We identify cross\-video relations at two levels\. First, we compare takeaways across all videos inℰ\\mathcal\{E\}, capturing relations between their central messages even when they rely on different supporting claims\. Second, we connect claims across individual graphs\. To avoid exhaustively comparing all cross\-video claim pairs, we embed claims withall\-MiniLM\-L6\-v2[Reimers and Gurevych \(2019\)](https://arxiv.org/html/2610.27149#bib.bib49)and cluster semantically related claims using BERTopic[Grootendorst \(2022\)](https://arxiv.org/html/2610.27149#bib.bib3)\. For each claim, we consider the top\-2020most similar claims from other videos from the same cluster\.

For each candidate pair\(u,v\)\(u,v\), an LLM classifies the directed relation fromuutovvasequivalent,supports,contradicts, ornone\. Similar to individual graph construction, we derive class probabilities from the model’s token log\-probabilities and assign the edge score:w⁡\(u,v\)=P⁡\(supports∣u,v\)\+P⁡\(equivalent∣u,v\)−P⁡\(contradicts∣u,v\)\.w\(u,v\)=P\(\\mathrm\{supports\}\\mid u,v\)\+P\(\\mathrm\{equivalent\}\\mid u,v\)\-P\(\\mathrm\{contradicts\}\\mid u,v\)\.

For every relation classified as other thannone, we add a cross\-video edge with its relation type and score\. Positive scores encode support or equivalence, whereas negative scores encode contradiction\. The resulting graph combines the reasoning expressed within individual videos with relations across their takeaways and claims, providing the connected discourse structure used for subsequent evidence augmentation and joint trust inference\.

#### 3\.2\.3Factual Grounding

The networked graphGℰG\_\{\\mathcal\{E\}\}\(Section[3\.2\.2](https://arxiv.org/html/2610.27149#S3.SS2.SSS2)\) represents influencer discourse but does not establish factual validity\. We therefore augment it with scientific evidence from PubMed, following the evidence\-retrieval pipeline in[Li et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib8)\.

Checkworthy nodes\.Since not every claim is verifiable against biomedical literature, we use an LLM to identifyPubMed\-checkworthyclaims\. Only these undergo evidence retrieval; all other claims remain inGℰG\_\{\\mathcal\{E\}\}\.

Query expansion and retrieval\.For each checkworthy claim, an LLM generates up to55supporting and55opposing queries to retrieve evidence bothforandagainstit\. UsingColBERTv2[Santhanam et al\. \(2022\)](https://arxiv.org/html/2610.27149#bib.bib9), we retrieve the top\-2020PubMed documents per query, aggregate them across queries, and retain the top\-2020documents in each supporting and opposing pool\. An LLM then filters these documents for relevance\.

Evidence classification and graph augmentation\.An LLM classifies each candidate as \{strongly supporting\(SS\),weakly supporting\(WS\),weakly opposing\(WO\),strongly opposing\(SO\),neutral, orirrelevant\}\. We retain up to three articles per pool, prioritizing stronger evidence \(SS or SO\) and excluding neutral or irrelevant articles\. Each retained articleeeis added as a node linked to claimccby an evidence edge\(e,c\)\(e,c\)with scorew⁡\(e,c\)=P⁡\(SS∣e,c\)\+0\.5​P​\(WS∣e,c\)−0\.5​P​\(WO∣e,c\)−P⁡\(SO∣e,c\)\.w\(e,c\)=P\(\\mathrm\{SS\}\\mid e,c\)\+0\.5P\(\\mathrm\{WS\}\\mid e,c\)\-0\.5P\(\\mathrm\{WO\}\\mid e,c\)\-P\(\\mathrm\{SO\}\\mid e,c\)\.The score lies in\[−1,1\]\[\-1,1\], with positive values indicating support and negative indicating opposition\. The evidence\-augmented graph is then used to jointly infer trust and estimate target video’s credibility\.

#### 3\.2\.4Trust Inference

Given the evidence\-augmented networked argumentation graph, we jointly infer trust over its claims and takeaways using probabilistic soft logic \(PSL\)[Bach et al\. \(2013\)](https://arxiv.org/html/2610.27149#bib.bib11);[Bach et al\. \(2017\)](https://arxiv.org/html/2610.27149#bib.bib10)\. PSL assigns soft truth values in\[0,1\]\[0,1\]and captures relational dependencies through first\-order logical rules[Beltagy et al\. \(2014\)](https://arxiv.org/html/2610.27149#bib.bib12)\. LetT⁡\(v\)∈\[0,1\]T\(v\)\\in\[0,1\]denote the trust of nodevv\. We treat retrieved PubMed evidence as trusted, fixingT⁡\(e\)=1T\(e\)=1, while jointly inferring trust for claims and takeaways\.

Each claimccis also assigned a priorπ⁡\(c\)=s⁡\(c\)​b​\(c\)\\pi\(c\)=s\(c\)b\(c\)based on its extractedspecificityandsourceattributes\.222We uses⁡\(c\)∈\{1,0\.5\}s\(c\)\\in\\\{1,0\.5\\\}for high/low specificity andb⁡\(c\)∈\{1,0\.5,0\.25,0\.1\}b\(c\)\\in\\\{1,0\.5,0\.25,0\.1\\\}for institutional/scientific, experiential, social testimony, and unsupported sources, respectively\.These priors capture how a claim is presented rather than its factual validity; takeaways receive no metadata\-based prior\. LetS⁡\(u,v\)S\(u,v\),A⁡\(u,v\)A\(u,v\), andE⁡\(u,v\)E\(u,v\)denote observed support, attack, and equivalence relations, with values given by their edge score magnitudes\. We define:

R​1:T⁡\(u\)∧S⁡\(u,v\)→T⁡\(v\),R​2:T⁡\(u\)∧A⁡\(u,v\)→¬T⁡\(v\),R​3:E⁡\(u,v\)∧T⁡\(u\)→T⁡\(v\),R​4:E⁡\(u,v\)∧T⁡\(v\)→T⁡\(u\),R5:Π\(c\)→T\(c\),R6:¬Π\(c\)→¬T\(c\),\\begin\{array\}\[\]\{rcl\}R1:T\(u\)\\land S\(u,v\)&\\rightarrow&T\(v\),\\\\ R2:T\(u\)\\land A\(u,v\)&\\rightarrow&\\neg T\(v\),\\\\ R3:E\(u,v\)\\land T\(u\)&\\rightarrow&T\(v\),\\\\ R4:E\(u,v\)\\land T\(v\)&\\rightarrow&T\(u\),\\\\ \\lx@intercol\\hfil R5:\\Pi\(c\)\\rightarrow T\(c\),\\quad R6:\\neg\\Pi\(c\)\\rightarrow\\neg T\(c\),\\hfil\\lx@intercol\\end\{array\}whereΠ⁡\(c\)\\Pi\(c\)is observed with valueπ⁡\(c\)\\pi\(c\)\.R​1R1–R​2R2propagate support and attack from trusted nodes,R​3R3–R​4R4encourage equivalent nodes to receive consistent trust, andR​5R5–R​6R6softly anchor claims to their metadata priors\.

Under PSL’s soft\-logic relaxation, each grounded rulerrinduces a hinge\-loss potentialϕr​\(𝐓\)=max⁡\{0,ℓr​\(𝐓\)\}pr\\phi\_\{r\}\(\\mathbf\{T\}\)=\\max\\\{0,\\ell\_\{r\}\(\\mathbf\{T\}\)\\\}^\{p\_\{r\}\}, whereℓr​\(𝐓\)\\ell\_\{r\}\(\\mathbf\{T\}\)is its linear rule violation andpr∈\{1,2\}p\_\{r\}\\in\\\{1,2\\\}\. MAP inference then finds the joint trust minimizing total weighted rule violation[Bach et al\. \(2017\)](https://arxiv.org/html/2610.27149#bib.bib10):

𝐓∗=arg​min𝐓∈\[0,1\]n⁡∑r∈ℛλr​ϕr​\(𝐓\)\.\\mathbf\{T\}^\{\*\}=\\mathop\{\\mathrm\{arg\\,min\}\}\_\{\\mathbf\{T\}\\in\[0,1\]^\{n\}\}\\sum\_\{r\\in\\mathcal\{R\}\}\\lambda\_\{r\}\\phi\_\{r\}\(\\mathbf\{T\}\)\.We use unit rule weights and linear hinge loss \(λr\\lambda\_\{r\}=11,prp\_\{r\}=11\)\. The resulting𝐓∗\\mathbf\{T\}^\{\*\}jointly resolves scientific evidence, discourse relations, and claim priors over the network\. We useT∗​\(tz\)T^\{\*\}\(t\_\{z\}\)as the credibility score of videozzand map it tocorrect/partially correct/incorrectwith fixed thresholds\.

## 4Evaluation Task

We evaluateGRIDthrough acontent validationtask: given an influencer content, classify it as \{correct,partially correct,incorrect\}\. We choose this task because successful prediction requires both understanding how the video’s claims support its message and evaluating those claims against evidence, providing an end\-to\-end evaluation\.

### 4\.1Annotated Dataset

Content validationtask requires gold\-standard judgments of factual validity\. We therefore construct an evaluation set from TikTok videos whose central message \(takeaway\) aligns with expert\-annotated claims fromScienceFeedback\. Medical experts onScienceFeedbackassign health\-related claims one of1818fine\-grained veracity labels \(Appendix[C\.2](https://arxiv.org/html/2610.27149#A3.SS2)\)\. As these are sparsely represented, we map them to our three task labels: \{correct,partially correct,incorrect\}\.

For each expert\-annotated claim, we extract up to two representative keywords and retrieve candidate videos through the TikTok Research API\. Because topical similarity alone does not guarantee that the expert judgment applies to a video’stakeaway, we first filter candidates for relevance and then use an LLM to verify alignment between the video’s takeaway and the correspondingScienceFeedbackclaim\. Aligned videos inherit the expert veracity label\. We validate this procedure throughhuman evaluationon9090randomly sampled videos: extracted takeaways are judged correct for91%91\\%of videos, with90%90\\%annotator agreement, while84%84\\%align with the correspondingScienceFeedbackclaim, with80%80\\%agreement \(Appendix[B](https://arxiv.org/html/2610.27149#A2)\)\. We further restrict retrieval to creators with at least1,0001\{,\}000followers to focus on influencers with an established audience\.

The aligned videos serve as labeled target instances, while unlabeled related videos are collected using the procedure in §[3\.1](https://arxiv.org/html/2610.27149#S3.SS1)to provide surrounding discourse\. The final dataset contains1245videos from1050unique influencers:569labeled target videos and676unlabeled related videos\.

## 5Experiments

We evaluateGRIDon thecontent validationtask against several baselines and conduct ablations to validate the contribution of its components\. Unless otherwise stated, all methods usegpt\-5\.4\-mini[OpenAI \(2026\)](https://arxiv.org/html/2610.27149#bib.bib51)for takeaway and claim extraction, andQwen3\.5\-9B[Yang et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib13)for all remaining LLM\-based components\. We report results from a single model run\.

### 5\.1Baselines

We compareGRIDagainst four baselines:

\(a\)Direct Prompting:We promptQwen3\.5\-9Bwith the target transcript to directly predict \{correct,partially correct,incorrect\}, relying only on the model’s parametric knowledge\.

\(b\)LOKI:We adapt LOKI[Li et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib8)to our task\. LOKI extracts a flat set of checkworthy claims, retrieves evidence, verifies each claim, and aggregates claim\-level scores into a video\-level prediction\.

\(c\)Per\-video Pragmatic Inference:Following[Nakshatri et al\. \(2026\)](https://arxiv.org/html/2610.27149#bib.bib2), we apply our structured reasoning pipeline to the target video independently: construct its argumentation graph, ground claims with PubMed evidence, and infer takeaway trust\. This captures within\-video argumentative structure but excludes cross\-video relations\.

\(d\)Per\-video Score Aggregation:We independently apply the same per\-video inference to the target and related videos, then average their takeaway trust scores; serving as a continuous version ofmajority voting\. This tests whether aggregating predictions across related videos is sufficient without explicitly modeling cross\-video interactions\.

For the latter two baselines, trust scores are mapped to classes using tuned thresholds\. Implementation details are provided in Appendix[C\.3](https://arxiv.org/html/2610.27149#A3.SS3)\.

Table 1:Main results for content validation:GRIDoutperforms all baselines in Macro\-F1, showing the benefit of the context of broader influencer discourse\. Inc\./Part\./Corr\.:incorrect/partially correct/correct\.Figure 3:Larger networks protectGRIDfrom very poor predictions\.In networks of up to2020videos, prediction quality spans nearly the full range from00to100%100\\%; in larger networks, every prediction stays above40%40\\%\. Points show individual target videos, plotted by number of videos in their network\.
### 5\.2Results

Extended results are reported in Appendix[A\.1](https://arxiv.org/html/2610.27149#A1.SS1)\.

Main Results\.Table[1](https://arxiv.org/html/2610.27149#S5.T1)presents results on thecontent validationtask\.GRIDoutperforms the strongest baseline by1414F1\-macro points\. The claim\-centric LOKI baseline performs substantially worse, partly due to limited coverage[Nakshatri et al\. \(2026\)](https://arxiv.org/html/2610.27149#bib.bib2): it identifies checkworthy claims for only≈75%\\approx 75\\%of target videos, restricting the evidence available for validation\. Modeling the target video as an argumentation graph withper\-video pragmatic inferenceimproves F1\-macro, showing the benefit of reasoning over its internal argumentative structure\. Incorporating related videos through score aggregation yields a further77\-point gain, indicating that the surrounding discourse provides useful context\. Our full method performs best by explicitly modeling interactions across videos, supporting our premise that influencer content should be evaluated jointly within the broader discourse rather than in isolation\.

Figure 4:DoesGRIDconform to the majority view when flooded with information?While majority voting’s prediction degrades with information flooding,GRIDmaintains and even improves in content validation\.Impact of Network Size\.We examine how the number of related videos affects prediction quality\. Figure[3](https://arxiv.org/html/2610.27149#S5.F3)plots prediction quality, defined as1−E1\-E\(EEis the absolute distance between the predicted trust score and the numeric gold score\), againstnetwork size, measured as the number of videos in the network including the target video and its related videos\. For smaller network size, prediction quality varies widely, spanning nearly the full range from00to11\. As network size increases,GRIDis protected from bad predictions, suggesting that related videos provide useful context for inference\. We further test this by randomly removing25%25\\%,50%50\\%, and75%75\\%of related videos; performance consistently decreases as more videos are removed \(Table[4](https://arxiv.org/html/2610.27149#A1.T4)\)\.

Varying Network Composition\.We vary the composition of the related\-video network by controlling the proportion of videos that oppose the target, which we refer to asinformation flooding\. At0%0\\%flooding, all related videos express a perspective consistent with the target; as flooding increases, opposing perspectives increasingly dominate the network\. This allows us to test whether our networked inference merely tracks the dominant signal in the neighborhood, as simple score aggregation over neighbors would, or instead reasons over the structure of cross\-video interactions\.

Figure[4](https://arxiv.org/html/2610.27149#S5.F4)shows the median prediction quality, with the2525–75%75\\%IQR band, across flooding levels\. We observe that our method andPer\-video Score Aggregationincreasingly diverge once flooding exceeds50%50\\%, when opposing perspectives become the majority\. While score aggregation degrades under this shift, our method remains comparatively robust, indicating that its predictions are not determined by simply aggregating per\-video scores\.

Impact of different interactions\.To evaluate the contribution of different edge types, we remove \(a\) takeaway–takeaway, \(b\) cross\-video claim–claim, and \(c\) PubMed–claim connections, while retaining all other edges\. As shown in Table[2](https://arxiv.org/html/2610.27149#S5.T2), removing cross\-video claim–claim connections reduces F1\-macro by≈7\\approx 7points, while removing takeaway–takeaway connections causes a larger1616\-point drop\. These results show that explicit cross\-video interactions, particularly between takeaways, are critical:GRID’s gains go beyond simply conforming to the majority\.

Table 2:Every connection type contributes, and scientific evidence contributes most\.F1 \(%\) after removing each connection type; fullGRIDreaches67\.667\.6Macro F1 \(Table[1](https://arxiv.org/html/2610.27149#S5.T1)\)\. Without PubMed–claim connections, Macro F1 drops by29\.829\.8points andpartially correctF1 nearly vanishes\. Inc\./Part\./Corr\.:incorrect/partially correct/correct\.These results show that: \(i\) related videos provide useful context for evaluating the target, and \(ii\) explicitly modeling cross\-video interactions is more effective than simple score aggregation, even when opposing perspectives dominate the network\.

## 6Analysis of Argumentative & Moral Framing by Credibility

Figure 5:Which framings are most credible?Claims citing empirical data or causal mechanisms, and claims framed around care/harm, receive above\-average trust, while claims framed around fairness, loyalty, or liberty receive the least\. Points show mean claim trust relative to the mean over all claims with95%95\\%bootstrap intervals\.The trust scores inferred byGRIDalso enable analysis of how influencers argue and frame their claims, including unlabeled videos without expert judgments\. We use an LLM to label each of the13,71713\{,\}717extracted claims with at most one argumentative strategy and any number of moral foundations \(definitions in Table[3](https://arxiv.org/html/2610.27149#S6.T3); human validation in Appendix[B](https://arxiv.org/html/2610.27149#A2)\), and relate these labels to claim trustT∗​\(c\)T^\{\*\}\(c\)and the credibilityT∗​\(tz\)T^\{\*\}\(t\_\{z\}\)of the corresponding video\.

Table 3:Argumentative\-strategy and moral\-foundation labels used in the discourse analysis\.Moral foundations separate claim trust more than argumentative strategies\.Figure[5](https://arxiv.org/html/2610.27149#S6.F5)reports the mean trust of claims in each category relative to the mean over all claims\. Among argumentative strategies, associational/empirical and causal/mechanistic claims receive the highest trust \(\+12\.7\+12\.7and\+9\.0\+9\.0points\), and source credibility claims the lowest \(−9\.3\-9\.3\)\.Differences across moral foundations are larger\.Care/harm is the only foundation above the mean \(\+5\.4\+5\.4\), while claims invoking fairness/cheating, loyalty/betrayal, or liberty/oppression fall1919–2323points below it\. These claims typically concern rights, mandates, or unfair treatment, which the biomedical literature rarely addresses, so they are seldom linked to supporting evidence\. When the two dimensions are combined,the moral foundation matters more than the strategy: causal/mechanistic claims are11\.511\.5points above the mean when they invoke care/harm, but18\.218\.2points below it when they invoke fairness/cheating \(Appendix Figure[8](https://arxiv.org/html/2610.27149#A1.F8)\)\. Health claims that invoke fairness, loyalty, or liberty therefore tend to lack evidential support regardless of how they are argued\.

Figure 6:Do low\-credibility videos invoke authority more often?On average, no \(all topics≈0\\approx 0\)\. But inCOVID\-19 vaccineandmasktopics they invoke authority more, whereas intreatmentanddiet, high\-credibility videos do\.Authority is invoked by both low\- and high\-credibility videos\.Theauthority/subversionfoundation includes both deferring to and challenging institutions\. To examine how its use relates to credibility, we manually group the145145expert\-reviewed claims in our dataset into health topics and, within each topic, compare the share of claims invoking authority in low\-credibility \(T∗​\(tz\)≤0\.5T^\{\*\}\(t\_\{z\}\)\\leq 0\.5\) and high\-credibility videos \(Figure[6](https://arxiv.org/html/2610.27149#S6.F6)\)\. Across all topics, the two groups invoke authority at similar rates \(23%23\\%vs\.22%22\\%of claims\), but this hides a split by topic:in some topics, appeals to authority are more common among credible influencers, while in others they are more common among influencers who should not be trusted\.In topics about COVID\-19 vaccines and masks, authority is more common in low\-credibility videos \(by88–1818points\), where it is typically used to challenge institutions:

Low\-credibilityNatural immunity“Experts and the media hid evidence that natural immunity protects against severe COVID\-19\.”

In topics about treatments and diet, such as ivermectin, it is more common in high\-credibility videos \(by1111–1717points\), where it is typically used to cite institutions:

High\-credibilityIvermectin“A 2021 Cochrane review of 11 studies found no benefit of ivermectin for COVID\-19\.”

The same moral foundation can thus serve opposite purposes in different topics, so its presence does not indicate whether content should be believed\.

## 7Conclusion

We introduced a structured framework,GRID, which validates influencer content through collective inference over interconnected discourse and scientific evidence\.GRIDoutperforms competing baselines on content validation and enables analysis of how discourse patterns vary with credibility\.

## Ethical Considerations

##### Datasets\.

Our work analyzes publicly available health\-related influencer content retrieved through the TikTok Research API and expert assessments published by ScienceFeedback\.

##### Privacy and identifiability\.

Although the analyzed content is public, health communication can involve sensitive topics and identifiable individuals\. We therefore do not infer private attributes\. To reduce privacy risks associated with multimedia data, our model operates only on textual transcripts and do not use or release raw video, audio, images, or biometric signals such as faces or voices\. We also avoid reporting identifying information about individual creators and present our analyses in aggregate\.

##### Content\-level evaluation\.

Our framework evaluates the credibility of video content, not the overall trustworthiness, expertise, or intent of the creator\. A prediction that a video isincorrectorpartially correctshould thereforenotbe interpreted as a characterization of the individual who produced it\. This distinction is particularly important because creators may discuss different topics with varying levels of accuracy\.

##### Health information and scientific evidence\.

Our framework relies on retrieved biomedical literature and expert\-derived labels to evaluate health\-related claims\. Scientific evidence may be incomplete, contested, or change over time, and errors may arise from claim extraction, retrieval, relevance filtering, or stance classification\. Moreover, not all claims—such as personal experiences or subjective recommendations—can be meaningfully evaluated using biomedical literature\. We therefore restrict evidence retrieval to PubMed\-checkworthy claims and do not treat the absence of retrieved evidence as evidence that a claim is false\.

##### Potential misuse and deployment\.

Network\-level credibility inference could be misused for automated moderation, surveillance, or ranking individuals based on model\-generated credibility estimates\. Our framework is intended for research on discourse understanding and content validation\. It is not meant for medical decision\-making\. Any real\-world deployment should account for model uncertainty, evolving evidence, and the potential consequences of erroneous predictions, and should incorporate appropriate human oversight\.

## Limitations

Scope and representativeness\.Our study focuses on English\-language health videos from TikTok creators with at least1,0001\{,\}000followers\. The dataset is modest \(1,2451\{,\}245videos, of which569569are labeled\), and its coverage is shaped by: \(a\) the claims available fromScienceFeedback, \(b\) the selected search terms, and \(c\) the TikTok Research API\. The findings may therefore not generalize to smaller creators, other platforms, languages, or domains\. The framework itself can extend to settings with suitable external evidence, such as legal or financial content, but this requires domain\-specific sources, label definitions, and evaluation rather than direct transfer of the present results\.

Evaluation labels\.Our labels are transferred from expert\-reviewed claims to videos whose extracted takeaways are semantically aligned; they are not direct expert assessments of the complete videos\. Human evaluation finds84%84\\%takeaway–expert alignment, leaving room for transfer errors\. Moreover,4646of the5757correctexamples are proxy instances aligned with curatedMayoClinictakeaways, and the final evaluation set is imbalanced \(478478incorrect,3434partially correct, and5757correct\)\. Mapping fine\-grained source verdicts into three classes also removes distinctions among different types of error\. Larger datasets with direct, multi\-expert video annotations and stronger coverage of the minority classes would support more conclusive evaluation\.

Content and network coverage\.GRIDoperates on transcripts and captions, so it does not model visual demonstrations, or other multimodal signals that may alter a video’s meaning\. Its discourse network is obtained through keyword retrieval and relevance filtering and should be understood as a sampled view of related content, rather than the complete information environment encountered by users\. We also omit social and propagation context such as follower relations, comments, and engagement; moreover, the largest network in our experiments contains5757videos\. Future work could combine multimodal content with social context for substantially larger networks\.

Evidence and modeling assumptions\.Factual grounding is restricted to English PubMed title–abstract records published from 2014 through 2025\. This provides a consistent biomedical source but may miss older work, clinical guidance, and any evidence outside PubMed\. We also fix every retained PubMed node at maximum trust, which does not distinguish study design, methodological quality, retractions, or the strength of scientific consensus\. Incorporating multiple curated evidence sources and estimatingsource\-andstudy\-level reliability are important extensions\.

Several stages—takeaway and claim extraction, relation identification, evidence filtering, and discourse labeling—depend on LLM outputs, so errors can propagate through the graph\. Human validation is encouraging but limited to9090videos and270270claims; argumentative strategies and moral foundations also have lower raw agreement \(70%70\\%and71%71\\%\), reflecting their subjectivity\. Our reported results use unit rule weights, fixed decision thresholds, two particular LLMs, and a single model run\. Future work should test robustness across models, prompts, random seeds, inference settings, and independently annotated evaluation sets\.

Interpretation and use\.The argumentative\- and moral\-framing results are descriptive associations with trust inferred byGRID; they do not show that a framing choice causes credibility, nor that low inferred trust establishes falsehood\. Topic composition, evidence availability, and modeling choices can all affect these patterns\. Finally, performance remains imperfect, particularly for the minority classes, so the framework should support human review rather than serve as an automated tool of influencer content credibility\.

## 8Statement of AI Assistant Usage

The development of the codebase was partially assisted by an AI system, and all AI\-generated code was reviewed and verified by the authors\. An AI assistant was also used to help refine the manuscript’s writing\.

## References

- Akhynkoet al\.\(2025\)K\. Akhynko, O\. Kosovan, and M\. TrokhymovychHidden persuasion: detecting manipulative narratives on social media during the 2022 Russian invasion of Ukraine\.InProceedings of the Fourth Ukrainian Natural Language Processing Workshop \(UNLP 2025\),M\. Romanyshyn \(Ed\.\),Vienna, Austria \(online\),pp\. 194–202\.External Links:[Link](https://aclanthology.org/2025.unlp-1.19/),[Document](https://dx.doi.org/10.18653/v1/2025.unlp-1.19),ISBN 979\-8\-89176\-269\-5Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p2.1)\.
- Alhindiet al\.\(2018\)T\. Alhindi, S\. Petridis, and S\. MuresanWhere is your evidence: improving fact\-checking by justification modeling\.InProceedings of the First Workshop on Fact Extraction and VERification \(FEVER\),J\. Thorne, A\. Vlachos, O\. Cocarascu, C\. Christodoulopoulos, and A\. Mittal \(Eds\.\),Brussels, Belgium,pp\. 85–90\.External Links:[Link](https://aclanthology.org/W18-5513/),[Document](https://dx.doi.org/10.18653/v1/W18-5513)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p2.1),[§2](https://arxiv.org/html/2610.27149#S2.p3.1),[§3\.1](https://arxiv.org/html/2610.27149#S3.SS1.p1.1)\.
- Atanasovaet al\.\(2020\)P\. Atanasova, J\. G\. Simonsen, C\. Lioma, and I\. AugensteinGenerating fact checking explanations\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,D\. Jurafsky, J\. Chai, N\. Schluter, and J\. Tetreault \(Eds\.\),Online,pp\. 7352–7364\.External Links:[Link](https://aclanthology.org/2020.acl-main.656/),[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.656)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Bachet al\.\(2017\)S\. H\. Bach, M\. Broecheler, B\. Huang, and L\. GetoorHinge\-loss markov random fields and probabilistic soft logic\.J\. Mach\. Learn\. Res\.18\(1\),pp\. 3846–3912\.External Links:ISSN 1532\-4435Cited by:[§3\.2\.4](https://arxiv.org/html/2610.27149#S3.SS2.SSS4.p1.1),[§3\.2\.4](https://arxiv.org/html/2610.27149#S3.SS2.SSS4.p3.1)\.
- Bachet al\.\(2013\)S\. Bach, B\. Huang, B\. London, and L\. GetoorHinge\-loss markov random fields: convex inference for structured prediction\.InConference on Uncertainty in Artificial Intelligence \(UAI\),Bellvue, WA, USA\.Cited by:[§3\.2\.4](https://arxiv.org/html/2610.27149#S3.SS2.SSS4.p1.1)\.
- Barzilai\-Nahon \(2008\)K\. Barzilai\-NahonToward a theory of network gatekeeping: a framework for exploring information control\.Journal of the American Society for Information Science and Technology59\(9\),pp\. 1493–1512\.External Links:[Document](https://dx.doi.org/https%3A//doi.org/10.1002/asi.20857),[Link](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.20857),https://onlinelibrary\.wiley\.com/doi/pdf/10\.1002/asi\.20857Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p1.1)\.
- Beltagyet al\.\(2014\)I\. Beltagy, K\. Erk, and R\. MooneyProbabilistic soft logic for semantic textual similarity\.InProceedings of the 52nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),K\. Toutanova and H\. Wu \(Eds\.\),Baltimore, Maryland,pp\. 1210–1219\.External Links:[Link](https://aclanthology.org/P14-1114/),[Document](https://dx.doi.org/10.3115/v1/P14-1114)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p4.1),[§3\.2\.4](https://arxiv.org/html/2610.27149#S3.SS2.SSS4.p1.1)\.
- Bozdaget al\.\(2025\)N\. B\. Bozdag, S\. Mehri, X\. Yang, H\. Ha, Z\. Cheng, E\. Durmus, J\. You, H\. Ji, G\. Tur, and D\. Hakkani\-TürMust read: a systematic survey of computational persuasion\.External Links:2505\.07775,[Link](https://arxiv.org/abs/2505.07775)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p2.1)\.
- Devoset al\.\(2022\)S\. Devos, S\. Eggermont, and L\. VandenboschInstagram influencers as superwomen: influencers’ lifestyle presentations observed through framing analysis\.Media and Communication10\(1\)\.External Links:[Link](https://www.cogitatiopress.com/mediaandcommunication/article/view/4717)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p1.1)\.
- Dmonteet al\.\(2026\)A\. Dmonte, R\. R\. Oruche, M\. Zampieri, P\. Calyam, and I\. AugensteinClaim verification in the age of large language models: a survey\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 4: Student Research Workshop\),S\. T\.Y\.S\.S\., J\. D\. Rodriguez, and O\. de Gibert \(Eds\.\),San Diego, California, United States,pp\. 15–29\.External Links:[Link](https://aclanthology.org/2026.acl-srw.2/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-srw.2),ISBN 979\-8\-89176\-393\-7Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1),[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Durmuset al\.\(2019\)E\. Durmus, F\. Ladhak, and C\. CardieThe role of pragmatic and discourse context in determining argument impact\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),K\. Inui, J\. Jiang, V\. Ng, and X\. Wan \(Eds\.\),Hong Kong, China,pp\. 5668–5678\.External Links:[Link](https://aclanthology.org/D19-1568/),[Document](https://dx.doi.org/10.18653/v1/D19-1568)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Fabriset al\.\(2025\)S\. R\. Fabris, J\. A\. C\. N\. Júnior, M\. R\. Serafini, and L\. F\. D\. LopesSelf\-care or misinformation? investigating the herbal wellness movement on tiktok\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p1.1)\.
- Forouzan and Ihler \(2013\)S\. Forouzan and A\. IhlerLinear approximation to admm for map inference\.InProceedings of the 5th Asian Conference on Machine Learning,C\. S\. Ong and T\. B\. Ho \(Eds\.\),Proceedings of Machine Learning Research, Vol\.29,Australian National University, Canberra, Australia,pp\. 48–61\.External Links:[Link](https://proceedings.mlr.press/v29/Forouzan13.html)Cited by:[§C\.1](https://arxiv.org/html/2610.27149#A3.SS1.p9.1)\.
- Fraidin \(2004\)S\. N\. FraidinWhen is one head better than two? interdependent information in group decision making\.Organizational Behavior and Human Decision Processes93\(2\),pp\. 102–113\.External Links:ISSN 0749\-5978,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.obhdp.2003.12.003),[Link](https://www.sciencedirect.com/science/article/pii/S0749597803001171)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Freedmanet al\.\(2025\)G\. Freedman, A\. Dejl, D\. Gorur, X\. Yin, A\. Rago, and F\. ToniArgumentative large language models for explainable and contestable claim verification\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 14930–14939\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1),[§2](https://arxiv.org/html/2610.27149#S2.p3.1),[§3\.2](https://arxiv.org/html/2610.27149#S3.SS2.p1.1)\.
- Georgakopoulou and Garcés\-Conejos Blitvich \(2024\)A\. Georgakopoulou and P\. Garcés\-Conejos BlitvichInfluencer discourse: affective relations and identities\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p1.1)\.
- Grootendorst \(2022\)M\. GrootendorstBERTopic: neural topic modeling with a class\-based TF\-IDF procedure\.arXiv preprint arXiv:2203\.05794\.Cited by:[§3\.2\.2](https://arxiv.org/html/2610.27149#S3.SS2.SSS2.p2.1)\.
- Haidt and Graham \(2007\)J\. Haidt and J\. GrahamWhen morality opposes justice: conservatives have moral intuitions that liberals may not recognize\.Social justice research20\(1\),pp\. 98–116\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p6.1)\.
- Hanet al\.\(2020\)Y\. Han, S\. Karunasekera, and C\. LeckieGraph neural networks with continual learning for fake news detection from social media\.ArXivabs/2007\.03316\.External Links:[Link](https://api.semanticscholar.org/CorpusID:220380817)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Hayes and Ben\-Shmuel \(2024\)A\. S\. Hayes and A\. T\. Ben\-ShmuelUnder the finfluence: financial influencers, economic meaning\-making and the financialization of digital life\.Economy and Society53\(3\),pp\. 478–503\.External Links:[Document](https://dx.doi.org/10.1080/03085147.2024.2381980),[Link](https://doi.org/10.1080/03085147.2024.2381980),https://doi\.org/10\.1080/03085147\.2024\.2381980Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p1.1)\.
- Kadavathet al\.\(2022\)S\. Kadavath, T\. Conerly, A\. Askell, T\. Henighan, D\. Drain, E\. Perez, N\. Schiefer, Z\. Dodds, N\. Dassarma, E\. Tran\-Johnson, S\. Johnston, S\. El\-Showk, A\. Jones, N\. Elhage, T\. Hume, A\. Chen, Y\. Bai, S\. Bowman, S\. Fort, D\. Ganguli, D\. Hernandez, J\. Jacobson, J\. Kernion, S\. Kravec, L\. Lovitt, K\. Ndousse, C\. Olsson, S\. Ringer, D\. Amodei, T\. B\. Brown, J\. Clark, N\. Joseph, B\. Mann, S\. McCandlish, C\. Olah, and J\. KaplanLanguage models \(mostly\) know what they know\.ArXivabs/2207\.05221\.External Links:[Link](https://api.semanticscholar.org/CorpusID:250451161)Cited by:[§3\.2\.1](https://arxiv.org/html/2610.27149#S3.SS2.SSS1.p3.1)\.
- Kotonya and Toni \(2020\)N\. Kotonya and F\. ToniExplainable automated fact\-checking for public health claims\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),B\. Webber, T\. Cohn, Y\. He, and Y\. Liu \(Eds\.\),Online,pp\. 7740–7754\.External Links:[Link](https://aclanthology.org/2020.emnlp-main.623/),[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.623)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Lawrence and Reed \(2019\)J\. Lawrence and C\. ReedArgument Mining: A Survey\.Computational Linguistics45\(4\),pp\. 765–818\.Note:Place: Cambridge, MA Publisher: MIT PressExternal Links:[Link](https://aclanthology.org/J19-4006/),[Document](https://dx.doi.org/10.1162/coli%5Fa%5F00364)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1)\.
- Lee and Theokary \(2021\)M\. T\. Lee and C\. TheokaryThe superstar social media influencer: Exploiting linguistic style and emotional contagion over content?\.Journal of Business Research132,pp\. 860–871\.External Links:ISSN 0148\-2963,[Link](https://www.sciencedirect.com/science/article/pii/S0148296320307657),[Document](https://dx.doi.org/10.1016/j.jbusres.2020.11.014)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p1.1),[§2](https://arxiv.org/html/2610.27149#S2.p2.1)\.
- Liet al\.\(2025\)H\. Li, X\. Han, H\. Wang, Y\. Wang, M\. Wang, R\. Xing, Y\. Geng, Z\. Zhai, P\. Nakov, and T\. BaldwinLoki: an open\-source tool for fact verification\.InProceedings of the 31st International Conference on Computational Linguistics: System Demonstrations,O\. Rambow, L\. Wanner, M\. Apidianaki, H\. Al\-Khalifa, B\. D\. Eugenio, S\. Schockaert, B\. Mather, and M\. Dras \(Eds\.\),Abu Dhabi, UAE,pp\. 28–36\.External Links:[Link](https://aclanthology.org/2025.coling-demos.4/)Cited by:[§C\.3](https://arxiv.org/html/2610.27149#A3.SS3.p2.1),[§2](https://arxiv.org/html/2610.27149#S2.p3.1),[§3\.2\.3](https://arxiv.org/html/2610.27149#S3.SS2.SSS3.p1.1),[§5\.1](https://arxiv.org/html/2610.27149#S5.SS1.p3.1),[Table 1](https://arxiv.org/html/2610.27149#S5.T1.2.6.1)\.
- Liu and Zheng \(2024\)X\. Liu and X\. ZhengThe persuasive power of social media influencers in brand credibility and purchase intention\.Humanities and Social Sciences Communications11\(1\),pp\. 1–12\.Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p2.1)\.
- Mehtaet al\.\(2022\)N\. Mehta, M\. L\. Pacheco, and D\. GoldwasserTackling fake news detection by continually improving social context representations using graph neural networks\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),S\. Muresan, P\. Nakov, and A\. Villavicencio \(Eds\.\),Dublin, Ireland,pp\. 1363–1380\.External Links:[Link](https://aclanthology.org/2022.acl-long.97/),[Document](https://dx.doi.org/10.18653/v1/2022.acl-long.97)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1),[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Mercier and Landemore \(2012\)H\. Mercier and H\. LandemoreReasoning is for arguing: understanding the successes and failures of deliberation\.Political Psychology33\(2\),pp\. 243–258\.External Links:[Document](https://dx.doi.org/https%3A//doi.org/10.1111/j.1467-9221.2012.00873.x),[Link](https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1467-9221.2012.00873.x),https://onlinelibrary\.wiley\.com/doi/pdf/10\.1111/j\.1467\-9221\.2012\.00873\.xCited by:[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Mochales and Moens \(2011\)R\. Mochales and M\. MoensArgumentation mining\.Artificial intelligence and law19\(1\),pp\. 1–22\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1),[§3\.2](https://arxiv.org/html/2610.27149#S3.SS2.p1.1)\.
- Montiet al\.\(2019\)F\. Monti, F\. Frasca, D\. Eynard, D\. Mannion, and M\. M\. BronsteinFake news detection on social media using geometric deep learning\.arXiv preprint arXiv:1902\.06673\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1),[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Nakshatriet al\.\(2026\)N\. S\. Nakshatri, E\. Caplan, R\. Pujari, and D\. GoldwasserTAIGR: towards modeling influencer content on social media via structured, pragmatic inference\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 40224–40248\.External Links:[Link](https://aclanthology.org/2026.acl-long.1868/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1868),ISBN 979\-8\-89176\-390\-6Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p2.1),[§2](https://arxiv.org/html/2610.27149#S2.p3.1),[§3\.1](https://arxiv.org/html/2610.27149#S3.SS1.p1.1),[§5\.1](https://arxiv.org/html/2610.27149#S5.SS1.p4.1),[§5\.2](https://arxiv.org/html/2610.27149#S5.SS2.p2.1)\.
- Nget al\.\(2025\)M\. P\. Ng, J\. Jiang, G\. Freedman, A\. Rago, and F\. ToniMArgE: meshing argumentative evidence from multiple large language models for justifiable claim verification\.External Links:2508\.02584,[Link](https://arxiv.org/abs/2508.02584)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- OpenAI \(2026\)OpenAIGPT\-5\.4 Thinking System Card\.Note:[https://deploymentsafety\.openai\.com/gpt\-5\-4\-thinking](https://deploymentsafety.openai.com/gpt-5-4-thinking)Cited by:[§5](https://arxiv.org/html/2610.27149#S5.p1.1)\.
- Powell and Pring \(2024\)J\. Powell and T\. PringThe impact of social media influencers on health outcomes: systematic review\.Social Science & Medicine340,pp\. 116472\.External Links:ISSN 0277\-9536,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.socscimed.2023.116472),[Link](https://www.sciencedirect.com/science/article/pii/S0277953623008298)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p1.1)\.
- Pradeepet al\.\(2021\)R\. Pradeep, X\. Ma, R\. Nogueira, and J\. LinScientific claim verification with VerT5erini\.InProceedings of the 12th International Workshop on Health Text Mining and Information Analysis,E\. Holderness, A\. Jimeno Yepes, A\. Lavelli, A\. Minard, J\. Pustejovsky, and F\. Rinaldi \(Eds\.\),online,pp\. 94–103\.External Links:[Link](https://aclanthology.org/2021.louhi-1.11/)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Radfordet al\.\(2023\)A\. Radford, J\. W\. Kim, T\. Xu, G\. Brockman, C\. McLeavey, and I\. SutskeverRobust speech recognition via large\-scale weak supervision\.InProceedings of the 40th International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.202,pp\. 28492–28518\.Cited by:[§3](https://arxiv.org/html/2610.27149#S3.p2.1)\.
- Reimers and Gurevych \(2019\)N\. Reimers and I\. GurevychSentence\-BERT: sentence embeddings using Siamese BERT\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),K\. Inui, J\. Jiang, V\. Ng, and X\. Wan \(Eds\.\),Hong Kong, China,pp\. 3982–3992\.External Links:[Link](https://aclanthology.org/D19-1410/),[Document](https://dx.doi.org/10.18653/v1/D19-1410)Cited by:[§3\.2\.2](https://arxiv.org/html/2610.27149#S3.SS2.SSS2.p2.1)\.
- Santhanamet al\.\(2022\)K\. Santhanam, O\. Khattab, J\. Saad\-Falcon, C\. Potts, and M\. ZahariaColBERTv2: effective and efficient retrieval via lightweight late interaction\.InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,M\. Carpuat, M\. de Marneffe, and I\. V\. Meza Ruiz \(Eds\.\),Seattle, United States,pp\. 3715–3734\.External Links:[Link](https://aclanthology.org/2022.naacl-main.272/),[Document](https://dx.doi.org/10.18653/v1/2022.naacl-main.272)Cited by:[§3\.2\.3](https://arxiv.org/html/2610.27149#S3.SS2.SSS3.p3.1)\.
- Sundriyalet al\.\(2023\)M\. Sundriyal, T\. Chakraborty, and P\. NakovFrom chaos to clarity: claim normalization to empower fact\-checking\.InFindings of the Association for Computational Linguistics: EMNLP 2023,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 6594–6609\.External Links:[Link](https://aclanthology.org/2023.findings-emnlp.439/),[Document](https://dx.doi.org/10.18653/v1/2023.findings-emnlp.439)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p2.1),[§2](https://arxiv.org/html/2610.27149#S2.p2.1)\.
- Sviridovaet al\.\(2025\)E\. Sviridova, E\. Cabrio, and S\. VillataMining implicit arguments for reasoning: A survey\.Argument & Computation,pp\. 19462174251344764\(EN\)\.Note:Publisher: SAGE PublicationsExternal Links:ISSN 1946\-2166,[Link](https://doi.org/10.1177/19462174251344764),[Document](https://dx.doi.org/10.1177/19462174251344764)Cited by:[§3\.2](https://arxiv.org/html/2610.27149#S3.SS2.p1.1)\.
- Thagard and Verbeurgt \(1998\)P\. Thagard and K\. VerbeurgtCoherence as constraint satisfaction\.Cognitive Science22\(1\),pp\. 1–24\.External Links:ISSN 0364\-0213,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/S0364-0213%2899%2980033-0),[Link](https://www.sciencedirect.com/science/article/pii/S0364021399800330)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Thagard \(1989\)P\. ThagardExplanatory coherence\.Behavioral and brain sciences12\(3\),pp\. 435–467\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p2.1),[§2](https://arxiv.org/html/2610.27149#S2.p4.1),[§3\.1](https://arxiv.org/html/2610.27149#S3.SS1.p1.1),[§3](https://arxiv.org/html/2610.27149#S3.p1.1)\.
- Vladikaet al\.\(2024\)J\. Vladika, P\. Schneider, and F\. MatthesHealthFC: verifying health claims with evidence\-based medical fact\-checking\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),N\. Calzolari, M\. Kan, V\. Hoste, A\. Lenci, S\. Sakti, and N\. Xue \(Eds\.\),Torino, Italia,pp\. 8095–8107\.External Links:[Link](https://aclanthology.org/2024.lrec-main.709/)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Waddenet al\.\(2020\)D\. Wadden, S\. Lin, K\. Lo, L\. L\. Wang, M\. van Zuylen, A\. Cohan, and H\. HajishirziFact or fiction: verifying scientific claims\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),B\. Webber, T\. Cohn, Y\. He, and Y\. Liu \(Eds\.\),Online,pp\. 7534–7550\.External Links:[Link](https://aclanthology.org/2020.emnlp-main.609/),[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.609)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Waltonet al\.\(2008\)D\. N\. Walton, C\. Reed, F\. Macagno,et al\.Argumentation schemes\.Vol\.1,Cambridge University Press Cambridge\.Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p6.1)\.
- Wanget al\.\(2019\)X\. Wang, W\. Shi, R\. Kim, Y\. Oh, S\. Yang, J\. Zhang, and Z\. YuPersuasion for good: towards a personalized persuasive dialogue system for social good\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,A\. Korhonen, D\. Traum, and L\. Màrquez \(Eds\.\),Florence, Italy,pp\. 5635–5649\.External Links:[Link](https://aclanthology.org/P19-1566/),[Document](https://dx.doi.org/10.18653/v1/P19-1566)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1),[§2](https://arxiv.org/html/2610.27149#S2.p2.1)\.
- Wanget al\.\(2025\)Y\. Wang, M\. Wang, H\. Iqbal, G\. N\. Georgiev, J\. Geng, I\. Gurevych, and P\. NakovOpenFactCheck: building, benchmarking customized fact\-checking systems and evaluating the factuality of claims and LLMs\.InProceedings of the 31st International Conference on Computational Linguistics,O\. Rambow, L\. Wanner, M\. Apidianaki, H\. Al\-Khalifa, B\. D\. Eugenio, and S\. Schockaert \(Eds\.\),Abu Dhabi, UAE,pp\. 11399–11421\.External Links:[Link](https://aclanthology.org/2025.coling-main.755/)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Yanget al\.\(2025\)A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§5](https://arxiv.org/html/2610.27149#S5.p1.1)\.
- Zhaoet al\.\(2021\)X\. Zhao, E\. Durmus, H\. Zhang, and C\. CardieLeveraging topic relatedness for argument persuasion\.InFindings of the Association for Computational Linguistics: ACL\-IJCNLP 2021,C\. Zong, F\. Xia, W\. Li, and R\. Navigli \(Eds\.\),Online,pp\. 4401–4407\.External Links:[Link](https://aclanthology.org/2021.findings-acl.386/),[Document](https://dx.doi.org/10.18653/v1/2021.findings-acl.386)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p4.1)\.
- Zhenget al\.\(2024\)L\. Zheng, C\. Li, X\. Zhang, Y\. Shang, F\. Huang, and H\. JiaEvidence retrieval is almost all you need for fact verification\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 9274–9281\.External Links:[Link](https://aclanthology.org/2024.findings-acl.551/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.551)Cited by:[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.
- Zhuet al\.\(2025\)Y\. Zhu, N\. Potyka, D\. Hernández, Y\. He, Z\. Ding, B\. Xiong, D\. Zhou, E\. Kharlamov, and S\. StaabArgRAG: explainable retrieval augmented generation using quantitative bipolar argumentation\.External Links:2508\.20131,[Link](https://arxiv.org/abs/2508.20131)Cited by:[§1](https://arxiv.org/html/2610.27149#S1.p3.1),[§2](https://arxiv.org/html/2610.27149#S2.p3.1)\.

## Appendix AExtended Results

### A\.1Ablation

Impact of Network Size\.

Table 4:Content\-validation performance after randomly removing related videos\. Results aggregate1010random trials\. Removing all related videos corresponds to thePer\-video Pragmatic Inferencebaseline in Table[1](https://arxiv.org/html/2610.27149#S5.T1)\.Table[4](https://arxiv.org/html/2610.27149#A1.T4)complements the observational trend in Figure[3](https://arxiv.org/html/2610.27149#S5.F3)with a controlled removal test\. F1\-macro decreases monotonically from0\.600\.60to0\.580\.58and0\.530\.53as we remove25%25\\%,50%50\\%, and75%75\\%of related videos; removing all related videos yields the per\-video score of0\.460\.46in Table[1](https://arxiv.org/html/2610.27149#S5.T1)\. Class\-wise F1 also decreases at every removal level\. Thus, the gain from reflects useful information supplied by the surrounding related videos\.

Varying Network Composition\.Figure[7](https://arxiv.org/html/2610.27149#A1.F7)separates the information\-flooding experiment in Figure[4](https://arxiv.org/html/2610.27149#S5.F4)by network size\. For networks of size22–1010, our method remains stable across flooding levels\. In contrast,Per\-video Score Aggregationdeclines steadily from0\.760\.76to0\.480\.48\. This widening gap shows that our method does not simply conform to the average trust score of the surrounding videos\.

For networks larger than1010, both methods initially lose prediction quality, but they diverge beyond50%50\\%flooding\. Specifically, compared to score aggregation baseline, we observe that our method improves at higher information flooding levels\. This improvement should not be interpreted as a general benefit of flooding: all targets in this region are labeledincorrect, and the larger\-network panel contains nocorrecttargets\. In these cases, additional opposing content can provide contrastive evidence against an incorrect target, which our structured inference can use while score aggregation simply follows the neighborhood average\. Subset \(or dataset\) construction for this experiment is detailed in Appendix[C\.4](https://arxiv.org/html/2610.27149#A3.SS4)\.

Figure 7:Median prediction quality across information\-flooding levels for small and larger networks; shaded regions show the25%−75%25\\%\-75\\%interquartile range\. The dashed line marks the point at which opposing perspectives become the majority\.
### A\.2Analysis of argumentative & moral framing by credibility

Table[3](https://arxiv.org/html/2610.27149#S6.T3)summarizes the discourse dimensions used in Section[6](https://arxiv.org/html/2610.27149#S6)\. An argumentative strategy is assigned only when a claim has an identifiable inferential role; a claim receives each moral\-foundation label that it invokes\. Consequently,28%28\\%of claims have no assigned strategy,31%31\\%have no assigned foundation, and claims may contribute to multiple foundation categories\.

Figure[8](https://arxiv.org/html/2610.27149#A1.F8)breaks down claim trust by each combination of argumentative strategy and moral foundation\. It supports the finding in Section[6](https://arxiv.org/html/2610.27149#S6)that the moral foundation matters more than the strategy: within every strategy, claims invokingcare/harmare among the most trusted, while those invokingfairness/cheating,loyalty/betrayal, orliberty/oppressionare among the least\.Authority/subversionvaries most with the accompanying strategy, ranging from well above average with associational/empirical reasoning to well below average with consequence/trade\-off or classification/definitional reasoning, consistent with its use for both citing and challenging institutions \(Figure[6](https://arxiv.org/html/2610.27149#S6.F6)\)\.

![Refer to caption](https://arxiv.org/html/2610.27149v1/fig_strategy_x_foundation.png)Figure 8:Does a strategy receive similar trust regardless of the moral foundation it invokes?No: the same strategy receives much higher trust when paired with care/harm than with fairness, loyalty, or liberty\. Cells show mean claim trust for each strategy–foundation pair relative to the mean over all claims; N/A marks pairs with fewer than3030claims\. Sign/diagnostic is omitted because most of its pairs fall below this threshold\.

## Appendix BHuman Evaluation

We conduct human validation of intermediate model outputs, which were used in dataset construction, claim\-prior estimation, and analysis of discourse patterns\. At the video level, annotators judge whether the takeaway identified bygpt\-5\.4\-minicaptures the video’s central message, and whether it aligns with the associated expert\-reviewed claim\. Together, these two tasks test whether the expert label can be transferred to the video\. Annotators are shown the transcript, caption, predicted takeaway, and expert\-reviewed claim\.

At the claim level, annotators judge thespecificityandsource invoked, as predicted byQwen3\.5\-9B, which determine the claim prior\. They also judge the model predicted argumentative\-strategy and moral\-foundation assignments used for discourse characterization \(Section[6](https://arxiv.org/html/2610.27149#S6)\)\. Annotators are shown the claim, takeaway, and all four predicted labels, with access to the transcript, and caption for context\. We also provide definitions and examples for each annotation task\.

We randomly sample9090videos\. The two video\-level tasks are annotated for every video; for the four claim\-level tasks, we randomly select up to three extracted claims per video, yielding270270claim items\. Each item is independently evaluated by two of the1212annotators\. The annotators had an average age of approximately2828, held at least a bachelor’s degree in computer science, and were primarily graduate students\. Each task asks whether the predicted output is correct or aligned\. For the multi\-label moral\-foundation task, a prediction is correct when at least one assigned foundation is appropriate\. Table[5](https://arxiv.org/html/2610.27149#A2.T5)reports the percentage of positive judgments and raw annotator agreement\.

Table 5:Human validation results and raw inter\-annotator agreement\. The result column reports the percentage judged correct, aligned, or yes, as applicable\.At the video level, the extracted takeaway aligns with the expert validation fromScienceFeedbackin84%84\\%of cases, with80%80\\%annotator agreement\. Moreover, the extracted takeaway itself is judged correct in91%91\\%of videos\. Together, these results support the reliability of transferring expert validation to the corresponding influencer content\. Thespecificityandsource invokedannotations similarly support their use in defining claim priors\. Argumentative strategies and moral foundations are also validated for most claims, although their lower raw agreement reflects the greater subjectivity of these discourse\-level categories\.

## Appendix CImplementation Details

This section reports the settings needed to reproduce our framework and baselines\. Prompt templates are provided in Appendix[D](https://arxiv.org/html/2610.27149#A4)\.

### C\.1Our Framework

LLM settings\.We usegpt\-5\.4\-minifortakeawayand claim extraction,Qwen3\.5\-9Bfor the remaining LLM components\. For local Qwen inference, we disable extended thinking and use temperature0\.70\.7, top\-p=0\.8p=0\.8, and the top\-100100token log\-probabilities\.

Influencer content collection\.The associated prompts are[D\.1](https://arxiv.org/html/2610.27149#A4.SS1)–[D\.1](https://arxiv.org/html/2610.27149#A4.SS1)in Appendix[D](https://arxiv.org/html/2610.27149#A4)\. We represent each video with its transcript and caption\. We transcribe the audio withfaster\-whisper333[https://pypi\.org/project/faster\-whisper/](https://pypi.org/project/faster-whisper/)\(small\.en\)\. After identifying thetakeaway\(core audience\-facing recommendation\) for the video, for increased recall using TikTok Research API, we use exactly two short, nonredundant terms–typically an entity and an outcome, for anANDquery over TikTok caption metadata\. From the obtained topic\-specific candidate videos, we extract takeaways and retain a candidate only when its central message addresses the same health issue and can meaningfully reinforce or challenge the targettakeaway; topical overlap alone is insufficient\. Duplicate video IDs and the target video itself are removed\.

Individual argumentation graphs\.Here, we discuss the details associated with claim extraction, and graph construction\. The associated prompts are[D\.1](https://arxiv.org/html/2610.27149#A4.SS1)–[D\.1](https://arxiv.org/html/2610.27149#A4.SS1)in Appendix[D](https://arxiv.org/html/2610.27149#A4)\.

Claim extraction: We use the transcript, caption, and previously extractedtakeawayfor this task\. The takeaway is supplied only to resolve the video’s intended message and prioritize claims relevant to that message; it is not treated as evidence for inventing additional claims\.

Graph construction details: claims are compared with a target node in batches of five\. For estimating edge weight, we normalize the log\-probabilities over the classes \{direct support,weak support, andnone\}, and computeP⁡\(direct\)\+0\.5​P​\(weak\)P\(\\mathrm\{direct\}\)\+0\.5P\(\\mathrm\{weak\}\)\. We require a score of at least0\.50\.5for claim\-to\-takeaway edges\. This stricter first\-level threshold prevents a merely relevant claim from becoming a direct justification for the video’s central message\. Once a claim is attached, it becomes a target for the remaining unattached claims; recursive claim\-to\-claim edges use a threshold of0\.10\.1so that incomplete but useful intermediate reasoning is not discarded\. Expansion ends when no new claim can be attached\. Claims that remain unattached receive no within\-video support edge\.

Networked argumentation discourse\.The associated prompt is[D\.1](https://arxiv.org/html/2610.27149#A4.SS1)in Appendix[D](https://arxiv.org/html/2610.27149#A4)\. We compare each takeaway with every other takeaway in the same related videos network\. For claims, we useall\-MiniLM\-L6\-v2444[https://huggingface\.co/sentence\-transformers/all\-MiniLM\-L6\-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2)embeddings and cluster using BERTopic to reduce the number of comparisons with clearly unrelated content\. For each claim, we rank cross\-video claims in the same semantic cluster by cosine similarity and retain at most the top2020candidates, as described in Section[3\.2\.2](https://arxiv.org/html/2610.27149#S3.SS2.SSS2)\. We retain every non\-nonerelation in the network\. For the edge score

Factual grounding\.The associated prompts are[D\.1](https://arxiv.org/html/2610.27149#A4.SS1)–[D\.1](https://arxiv.org/html/2610.27149#A4.SS1)in Appendix[D](https://arxiv.org/html/2610.27149#A4)\. Our retrieval index contains approximately1414million English\-language PubMed title\-\-abstract records published from 2014 through 2025\. We build the index from the 2025 PubMed annual baseline555[https://ftp\.ncbi\.nlm\.nih\.gov/pubmed/baseline/](https://ftp.ncbi.nlm.nih.gov/pubmed/baseline/)and use ColBERTv2 for semantic retrieval\. This date restriction prioritizes recent biomedical research and guidance while retaining good coverage\.

Only claims labeled PubMed\-checkworthy enter retrieval\. For each such claim, we generate up to five supporting and five opposing queries\. We retrieve2020documents per query, deduplicate documents by PMID, and retain the top2020documents separately, by their retrieval score, for the supporting and opposing pools\. An LLM relevance pass removes documents that do not match the claim and the intended pool direction\. Abstracts are truncated to300300words for this pass\. We then classify the remaining articles by stance and retain at most three articles from each pool, preferring strong over weak evidence and excluding neutral or irrelevant articles\. A PMID retained from one pool is not duplicated in the other\.

Trust inference\.We implement the model withpslpython666[https://pypi\.org/project/pslpython/](https://pypi.org/project/pslpython/)\. All six rules in Section[3\.2\.4](https://arxiv.org/html/2610.27149#S3.SS2.SSS4)use weight11and linear hinge loss\. PubMed nodes are observed at11; claim priors are added as observedTrustPriorvalues through Rules R5–R6; and takeaways are unobserved targets initialized at00\. We run ADMM[Forouzan and Ihler \(2013\)](https://arxiv.org/html/2610.27149#bib.bib23)inference with seed77, at most5,0005\{,\}000iterations\. For the reported full model, trust scores in\[0,0\.6\]\[0,0\.6\],\(0\.6,0\.65\]\(0\.6,0\.65\], and\(0\.65,1\]\(0\.65,1\]map toincorrect,partially correct, andcorrect, respectively\.

### C\.2Annotated Dataset

Source alignment\.We use expert\-reviewed health claims fromScienceFeedbackas the primary label source\. For each expert\-claim, we select up to two representative keywords and retrieve candidate videos through the TikTok Research API, restricting the set to English\-language videos from creators with at least1,0001\{,\}000followers\. After topical filtering, we extract each video’s takeaway and retain the video only when its takeaway is semantically equivalent to the source claim from the expert\. This stricter alignment is necessary to transfer the source verdict to the video’s central message\.

Due to the class\-imbalance issue found in expert\-annotatedScienceFeedbackclaims, especially for claims that receive afully correctverdict, we augment this class using manually curated takeaways from trustedMayoClinicvideos as gold\-standard correct claims\. We apply the same retrieval and semantic\-equivalence criterion to find TikTok videos expressing those claims; the resulting matches serve as proxy instances of thecorrectclass\. This procedure contributes4646of the final5757correct examples\.

Label mapping and data curation\.Among the fine\-grained verdicts represented in our source claims \(from experts\), we map \{inaccurate,mostly inaccurate,incorrect,flawed reasoning,unsupported, andmisleading\} toincorrectclass; \{lacks context,imprecise,partially correct,mostly accurate, andmostly correct\} topartially correct; and \{accurateandcorrect\} tocorrectclass\. Overall, the final evaluation set contains569569unique labeled videos:478478incorrect,3434partially correct, and5757correct\-labeled videos\. Table[6](https://arxiv.org/html/2610.27149#A3.T6)shows the resulting dataset by influencer popularity\.

Table 6:Dataset distribution for labeled \(expert\-annotated\) and unlabeled videos by influencer popularity\.
### C\.3Baselines

Direct prompting\.The associated prompt is[D\.2](https://arxiv.org/html/2610.27149#A4.SS2)in Appendix[D](https://arxiv.org/html/2610.27149#A4)\. We provideQwen3\.5\-9Bonly the target transcript and request one of the three task labels with a short rationale\. No retrieved evidence, related\-video content, extracted takeaway, or gold metadata is included\.

LOKI\.The associated prompts are[D\.2](https://arxiv.org/html/2610.27149#A4.SS2)–[D\.2](https://arxiv.org/html/2610.27149#A4.SS2)in Appendix[D](https://arxiv.org/html/2610.27149#A4)\. We adapt LOKI[Li et al\. \(2025\)](https://arxiv.org/html/2610.27149#bib.bib8)using their source code777[https://github\.com/Libr\-AI/OpenFactVerification](https://github.com/Libr-AI/OpenFactVerification)\. It extracts a flat set of atomic factual claims from the caption and transcript, verifies each claim independently, and aggregates the results to verify the overall document\. To ensure fair comparison across all methods, we retrieve evidence for every query using our indexed PubMed data, rather than relying onSERPER\-APIas in their source code\. Note that we generate two supporting and two opposing retrieval queriesper claim\. Retrieval keeps five hits per query, deduplicates the pooled results by PMID, retains the top fiveper claim, and sends the top three to the evidence\-relation classifier\. Claim factuality is the fraction of directional evidence labeledsupportsamong evidence labeledsupportsorrefutes; the video score is the mean over claims with directional evidence\. Cases with no verified claim remain uncovered rather than receiving an imputed prediction\. For this baseline, we conduct a threshold sweep using a grid from\{0\.05,0\.1,…,0\.95\}\\\{0\.05,0\.1,\.\.\.,0\.95\\\}to identify threshold pairs optimized for obtaining maximum F1\-macro score\. The reported thresholds are0\.90\.9and0\.950\.95for three ordered classes\. Specifically,≤0\.9\\leq 0\.9forincorrect,\(0\.9,0\.95\]\(0\.9,0\.95\]forpartially correctclass,\>0\.95\>0\.95forcorrectclass\.

Per\-video pragmatic inference\.This baseline uses the same extraction, within\-video graph construction, factual grounding, priors, and PSL rules as our full system, but creates one graph containing only the target video\. The reported thresholds are0\.550\.55and0\.750\.75for the three ordered classes, obtained based on a threshold sweep and optimized for maximum F1\-macro score\.

Per\-video score aggregation\.We first run the preceding single\-video procedure independently for every video including the target and its related videos in the network, then assign the target the arithmetic mean of their takeaway trust scores\. The reported results are with thresholds0\.600\.60and0\.700\.70\(obtained after a threshold sweep, optimized for maximum F1\-macro score\)\. Thus, this baseline receives the same videos and scientific\-grounding machinery as our method but cannot propagate support, contradiction, or equivalence across videos\.

### C\.4Ablation

Varying Network Composition\.For each labeled target video, we begin with the full related\-video graph\. LetAAcontain every othertakeaway\-bearing video in this graph\. We partitionAAinto videos whose takeaways are aligned with the target video’s perspective and videos whose takeaways contradict it\. These judgements are made using an LLM \(Qwen\-3\.5\-9B\)\. Each ablation instance retains the target video and a subsetS⊆AS\\subseteq A\. We define its relative size as\(\|S\|\+1\)/\(\|A\|\+1\)\(\|S\|\+1\)/\(\|A\|\+1\)and its flooding level as\|Sopp\|/\(\|S\|\+1\)\|S\_\{\\mathrm\{opp\}\}\|/\(\|S\|\+1\), where the\+1\+1accounts for the target video\. Thus,0%0\\%flooding contains onlyalignedrelated videos, while higher values retain an increasingly opposing network\. In this manner, we construct different subsets by varying relative size and flooding level, yielding3,6763\{,\}676subsets for analysis\.

## Appendix DPrompt Specifications

We provide the operative prompt instructions below\. To keep the appendix compact, we omit illustrative demonstrations and repeated JSON\-only output reminders to the model; all prompts preserve the supplied identifiers, and return one decision for every input item\. Bracketed text denotes an input field\.

### D\.1Framework Prompts

P1: Takeaway extraction\.

> Input:\[transcript; caption\]\. Instruction:Extract exactly one central recommendation, warning, suggestion, protocol, or audience\-directed implication\. Select the message the audience is primarily meant to retain or act upon; if several are equally strong, select the first\. Write it as a self\-contained, audience\-facing statement, resolving pronouns when the input permits\. Preserve the original meaning, scope, and certainty\. Do not add medical qualifications, factual judgments, or outside knowledge\. Return the takeaway, whether it is explicit or implicit, a short selection rationale, and the exact grounding span\. If the input contains no coherent takeaway, returnNO\_TAKEAWAY\_FOUND\.

P2: Search\-keyword selection\.

> Input:\[takeaway; optional caption/transcript context\]\. Instruction:Select exactly two simple, atomic terms for an AND search over social\-media captions\. Prefer common words likely to occur in captions, usually one for the main intervention/entity and one for the condition, outcome, or core issue\. Avoid redundant variants, unnecessary modifiers, and overly specific phrases\. Balance recall with precision\. Optionally return one auxiliary term for later filtering, but do not place it in the search conjunction\.

P3: Retrieved\-video relevance\.

> Input:\[target takeaway; candidate takeaway\]\. Instruction:Decide whether the candidate addresses the same underlying health issue as the target and expresses a message that is equivalent to, supports, or opposes the target\. Recommendations may differ, but the overlapping factual issue must be identifiable\. Shared entities or topic words alone are insufficient\. Returnrelevantwith the applicable relation, ornot\_relevant, with one\-sentence justification\.

P4: Claim and metadata extraction\.

> Input:\[transcript; caption; takeaway\]\. Instruction:Extract every atomic, testable proposition stated or strongly implied in the transcript or caption that helps characterize the video’s message\. Each claim must be self\-contained, preserve concrete qualifiers \(population, dose, duration, comparison, source, and outcome\), and include an exact grounding span\. Do not invent a claim from the takeaway\. Label each claim as explicit or implicit; specificity ashighonly when concrete bounded details materially narrow the proposition, otherwiselow; and the primary source invoked asinstitutional/scientific evidence,experiential evidence,social testimony, orunsupported assertion\. These source labels describe presentation, not truth\. Return an empty list when no claim is present\.

P5: Within\-video support\.

> Input:\[target claim; candidate statements from the same video\]\. Instruction:Classify the directed relation from each statement to the target claim asdirect support,weak support, ornone\. Direct support clearly entails, restates, justifies, or gives evidence for the target\. Weak support makes the target more credible but requires an additional inference\. Use none for topical overlap, missing assumptions, or a highly indirect connection\. Do not infer unstated facts\. Return a one\- or two\-sentence rationale for each statement\.

P6: Cross\-video relation\.

> Input:\[source claim; candidate claims from other videos\]\. Instruction:Classify the directed relation from the source to each candidate asequivalent,contradicts,supports, ornone, in that decision order\. Equivalent requires the same core proposition, polarity, entities, and material scope\. Contradiction requires conflicting assertions about the same issue\. Support requires the source to provide a direct reason for the candidate and is directional\. Use none for compatible statements about different aspects, reverse\-direction support, vague relatedness, or uncertainty\. Use only the supplied text and provide one sentence of rationale\.

P7: PubMed checkworthiness\.

> Input:\[transcript; extracted claims\]\. Instruction:Label each claimPubMed\-checkworthywhen it makes a biomedical or health\-related factual assertion that is generalizable beyond the immediate anecdote and can in principle be evaluated using PubMed literature\. Label personal, rhetorical, non\-biomedical, unresolvably vague, or context\-specific statementsnot PubMed\-checkworthy\. Judge verifiability rather than truth; when genuinely uncertain, prefer checkworthy\. Give one brief explanation per claim\.

P8: Bidirectional query expansion\.

> Input:\[transcript; PubMed\-checkworthy claim\]\. Instruction:Generate up to five supporting and five opposing semantic\-search queries\. Each query must be a concise, standalone natural\-language phrase or question suitable for embedding retrieval, without Boolean syntax or field tags\. Cover distinct, claim\-compatible angles such as outcome, mechanism, population, route, study design, safety, adverse events, or null effects\. Opposing queries should seek contrary, null, or adverse evidence\. Preserve the claim’s population, intervention, and outcome; do not introduce a new effect or broaden its scope\. Return a one\-sentence justification for each query\.

P9: PubMed relevance filtering\.

> Input:\[takeaway; claim; intended supporting/opposing pool; article titles and abstracts\]\. Instruction:For each article, first determine whether it addresses a matching exposure/intervention, outcome, population, or comparison; then determine whether its reported findings could serve the intended pool direction\. Returnrelevantonly when the article is useful for evaluating this claim in that pool\. Returnnot\_relevantfor off\-topic work, mismatched outcomes, insufficient evidence, or an article that belongs to the opposite pool\. Do not treat keyword overlap or retrieval rank as evidence, and do not infer unstated results\. When uncertain, choose not relevant\.

P10: PubMed evidence stance\.

> Input:\[claim; article titles and abstracts\]\. Instruction:Classify each article asSS\(direct, strong support\),WS\(limited, indirect, qualified, or context\-specific support\),WO\(limited, indirect, qualified, or context\-specific opposition\),SO\(direct, strong opposition\),N\(related but nondirectional\), orI\(irrelevant\)\. Evaluate the exact claim and distinguish population, dose, route, timing, comparison, and outcome\. Use strong labels only for directly aligned evidence\. For a negative claim, a null result supports it only when the article evaluates the same effect\. Use only the title and abstract and give one concise rationale\.

### D\.2Baseline Prompts

P11: Direct content validation\.

> Input:\[target transcript\]\. Instruction:Assess the transcript as a whole using general medical and scientific knowledge\. Returnfully\_correctwhen all material claims and recommendations are accurate without a consequential missing qualification;partially\_correctwhen accurate material content co\-occurs with an important error, overstatement, unsupported generalization, or missing qualification; andincorrectwhen the central claim or recommendation is false, unsupported, or substantially misleading and any accurate details are incidental\. Preserve distinctions in population, timing, dose, route, comparison, certainty, and outcome\. Return the label and a concise rationale\.

P12–P15: LOKI adaptation\.

> Claim extraction input:\[caption; transcript\]\. Extract concise, atomic, independently checkable health claims\. Resolve references using only the input; preserve material qualifiers; split independently assessable conjunctions; remove semantic duplicates; and exclude recommendations, opinions, rhetorical questions, and purely anecdotal experiences unless they contain a concrete factual assertion\. Checkworthiness input:\[claims\]\. Mark each claimCHECKWORTHYwhen PubMed evidence could objectively support or refute its empirical medical, biological, epidemiological, treatment, prevention, diagnostic, or risk assertion\. Otherwise markNOT\_CHECKWORTHY\. Judge investigability, not truth\. Query input:\[checkworthy claims\]\. For each claim, generate exactly two supporting\-direction and two opposing\-direction sentence\-like PubMed queries\. Use established terminology and retain material population, intervention/exposure, and outcome constraints\. Do not repeat the original claim, use Boolean fragments, assume a truth label, omit a claim, or invent one\. Evidence input:\[claims; retrieved article titles/abstracts\]\. For every claim–article pair, returnSUPPORTSwhen the evidence directly backs the claim at a compatible scope,REFUTESwhen it directly contradicts a material part or reports a compatible null/opposite result, andIRRELEVANTwhen it is indirect, mismatched, or insufficient\. Use only the supplied evidence and give a brief rationale\.