Unsupervised Anomaly Detection of Information Operations Users via Behavioral and Language Patterns
Summary
Proposes TENSOR, an unsupervised anomaly detection method for identifying information operations users by analyzing temporal behavioral and language patterns using temporal point processes and LLM responses. Outperforms baselines on five real-world datasets.
arXiv:2607.05855v1 Announce Type: new
Abstract: Information Operations on social media networks have been identified as a significant threat to democracy and modern society, but they are challenging and expensive to detect by humans. Existing supervised IO detection methods fail to capture the dynamic nature of evolving IO user behavior, while existing unsupervised approaches rely on oversimplified assumptions of coordination among IO users that may not exist in practice. To overcome the limitations of existing methods, we formulate IO user detection as an anomaly detection problem and propose a novel unsupervised IO user detection approach called Temporal-bEhavior-laNguage Signals for information Operation Recognition (TENSOR), which leverages multimodal data, including temporal online user behavior, such as message posting activities, and the textual content of the messages. The motivation is that IO users are typically a very small fraction of all online users and have unique temporal behavioral and language patterns. Specifically, we train a Temporal Point Process (TPP) to capture abnormal temporal behavioral patterns of IO users because they are known to behave in a coordinated manner for IO campaigns. We further introduce a novel evidence function that converts LLM responses, which are generated from user post timelines, into quantitative scores to adjust the TPP outputs for better IO user detection. Experimental results show that TENSOR outperforms the baselines on five real-world IO datasets. Code is available at https://github.com/xiuzhenzhang/TENSOR.
View Cached Full Text
Cached at: 07/08/26, 04:45 AM
# Unsupervised Anomaly Detection of Information Operations Users via Behavioral and Language Patterns
Source: [https://arxiv.org/html/2607.05855](https://arxiv.org/html/2607.05855)
11institutetext:RMIT University, Melbourne, Victoria 3000, Australia11email:sishun\.liu@student\.rmit\.edu\.au, \{sajal\.halder,ke\.deng,xiuzhen\.zhang\}@rmit\.edu\.au22institutetext:Macquarie University, Sydney, New South Wales 2000, Australia22email:yan\.wang@mq\.edu\.au###### Abstract
Information Operations \(IOs\)onSocial Networks \(SNs\)have been identified as a significant threat to democracy and modern society, but they are challenging and expensive to detect by humans\. Existing supervisedIOdetection methods fail to capture the dynamic nature of evolvingIOuser behavior, while existing unsupervised approaches rely on oversimplified assumptions of coordination amongIOusers that may not exist in practice\. To overcome the limitations of existing methods, we formulateIOuser detection as an anomaly detection problem and propose a novel unsupervisedIOuser detection approach calledTemporal\-bEhavior\-laNguage Signals for information Operation Recognition \(TENSOR\), which leverages multimodal data, including temporal online user behavior, such as message posting activities, and the textual content of the messages\. The motivation is thatIOusers are typically a very small fraction of all online users and have unique temporal behavioral and language patterns\. Specifically, we train aTemporal Point Process \(TPP\)to capture abnormal temporal behavioral patterns ofIOusers because they are known to behave in a coordinated manner forIOcampaigns\. We further introduce a novelevidence functionthat convertsLLMresponses, which are generated from user post timelines, into quantitative scores to adjust theTPPoutputs for betterIOuser detection\. Experimental results show thatTENSORoutperforms the baselines on five real\-worldIOdatasets111Code is available at[https://github\.com/xiuzhenzhang/TENSOR](https://github.com/xiuzhenzhang/TENSOR)\.\.
## 1Introduction
The development ofSocial Networks \(SNs\)enables fast dissemination of critical information, large\-scale discussions, and joint actions about political and social issues becauseSNsconnect people\[[6](https://arxiv.org/html/2607.05855#bib.bib6)\]\. However, the capabilities ofSNscan be misused byInformation Operations \(IOs\), especially state\-sponsored ones\.IOsare organized attempts to tamper with the regular flow of information and influence public opinion through disinformation, hate speech, and other harmful content\.IOsare hard to detect because they are always initiated by a small group of users\[[28](https://arxiv.org/html/2607.05855#bib.bib28),[37](https://arxiv.org/html/2607.05855#bib.bib37)\]\. With targets including narrative manipulation and the fostering of division in online and real\-world communities, IOs have been identified as a significant threat to democracy, and the need for robust methods to detect these operations is urgent\[[7](https://arxiv.org/html/2607.05855#bib.bib7)\]\.
Researchers have proposedIOdetection approaches to identify whether individual posts or specific users are related to anIO\. The following discussion focuses on the detection ofIOusers, which is the primary interest of this study\.IOusersare motivated or incentivized to promoteIOs, while legitimate organic users are calledcontrol users\.IOuser detection approaches use patterns within user post timelines onSNs\. These patterns can be categorized asbehavioral patterns\(specifically,temporalbehavioral patterns because they describe the user activities onSNsover time\) andlanguage patterns, including speaking style and areas of interest\.
IOuser detection is challenging\. The biggest challenge is the generalization capability ofIOuser detection algorithms,*i\.e*\., their ability to detect unseenIOs\. In the wild,IOsevolve quickly, so existing labeled IO datasets always lag\. This means that existing supervised\[[1](https://arxiv.org/html/2607.05855#bib.bib1),[6](https://arxiv.org/html/2607.05855#bib.bib6),[18](https://arxiv.org/html/2607.05855#bib.bib18)\]and semi\-supervised\[[2](https://arxiv.org/html/2607.05855#bib.bib2),[20](https://arxiv.org/html/2607.05855#bib.bib20),[31](https://arxiv.org/html/2607.05855#bib.bib31),[36](https://arxiv.org/html/2607.05855#bib.bib36)\]IOuser detection methods suffer from a generalization issue,*i\.e*\., they cannot detect new, unseenIOs\. Recent zero\-shotLLMapproaches\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\]provide an unsupervised alternative, but their ability to capture complex temporal behavioral dynamics is still limited\.
This paper formulatesIOuser detection as an unsupervised anomaly detection problem and proposes a novel approach, calledTemporal\-bEhavior\-laNguage Signals for information Operation Recognition \(TENSOR\), to use behavioral and language patterns from user post timelines for unsupervisedIOuser detection\. Training anomaly detection models onIOdata is difficult because the training data, which are supposed to consist of timelines of control users, are contaminated byIOusers\. This harms the performance ofIOuser detection models\. To mitigate this issue, we use language patterns\. Specifically, first, we train aTemporal Point Process \(TPP\)on the contaminated data to identifyIOusers based on abnormal behavioral patterns, such as coordinated behavior among differentIOaccounts, which are commonly observed amongIOusers but rare among control users\. ATPPis a well\-defined stochastic process over temporal event sequences\[[4](https://arxiv.org/html/2607.05855#bib.bib4)\]\. Researchers have used the process to achieve state\-of\-the\-art performance in unsupervised identification of abnormal event sequences from normal ones\[[29](https://arxiv.org/html/2607.05855#bib.bib29)\]and outlier events from normal events\[[16](https://arxiv.org/html/2607.05855#bib.bib16)\]\. Second, we propose a novelevidence functionto adjustTPPinference\. This function convertsLLMresponses, which are generated from user post timelines, into quantitative scores that refine theTPPoutput for betterIOuser detection\. Experimental results show thatTENSORoutperforms other baselines by a significant margin on five real\-worldIOdatasets\.
## 2Related Work
Most existingIOdetection studies classifyIOusers based on pure behavioral patterns\[[6](https://arxiv.org/html/2607.05855#bib.bib6),[18](https://arxiv.org/html/2607.05855#bib.bib18)\], pure language patterns\[[1](https://arxiv.org/html/2607.05855#bib.bib1),[2](https://arxiv.org/html/2607.05855#bib.bib2),[8](https://arxiv.org/html/2607.05855#bib.bib8),[9](https://arxiv.org/html/2607.05855#bib.bib9),[10](https://arxiv.org/html/2607.05855#bib.bib10)\], or both\[[17](https://arxiv.org/html/2607.05855#bib.bib17),[20](https://arxiv.org/html/2607.05855#bib.bib20),[36](https://arxiv.org/html/2607.05855#bib.bib36)\]\. Vargas et al\.\[[36](https://arxiv.org/html/2607.05855#bib.bib36)\]classifyIOusers based on coordination behaviors, such as tweeting the same content within a timeframe, or tweeting the same hashtag within a timeframe, extracted from users’ post timelines\. Luceri et al\.\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\]investigateIOuser detection forLLMsunder few\-shot and fine\-tuning settings\. Minici et al\.\[[20](https://arxiv.org/html/2607.05855#bib.bib20)\]propose IOHunter, anIOdetector built upon a graph, in which the similarity of users’ behavioral traces, including tweets, hashtags, and time of posting, justifies connections between users\. Despite reported good performance on existing data, supervised and semi\-supervisedIOuser detection methods require labelledIOdatasets, which are limited in size and lag behind real\-worldIOpractices\. This raises concerns about the generalization capabilities of supervised and semi\-supervisedIOuser\-detection algorithms in real\-world scenarios\.
Recent zero\-shotLLMapproaches\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\]provide an unsupervised alternative, but their ability to capture complex temporal behavioral dynamics is still limited\. Although not directly aimed atIOuser detection, other studies concerningIOusers verify coordination among users in the sameIO\[[12](https://arxiv.org/html/2607.05855#bib.bib12),[21](https://arxiv.org/html/2607.05855#bib.bib21)\]\. Although such methods reveal thatIOusers tend to form small and relatively well\-separated clusters compared with legitimate organic users, our study shows that it is an oversimplification to develop unsupervisedIOuser detection methods solely based on this observation\.
Anomaly Detection using contaminated data:Anomaly detection using contaminated data is possible if the anomalies occupy a small portion of the data\[[38](https://arxiv.org/html/2607.05855#bib.bib38)\]\. Most existing methods employ unsupervised identification of anomalies during training, allowing the model to either exclude these samples or leverage them to mitigate performance degradation\[[25](https://arxiv.org/html/2607.05855#bib.bib25),[27](https://arxiv.org/html/2607.05855#bib.bib27),[39](https://arxiv.org/html/2607.05855#bib.bib39)\]\. Patra et al\.\[[24](https://arxiv.org/html/2607.05855#bib.bib24)\]observed that these methods require prior knowledge about the dataset, such as the contamination ratio, which is usually unknown\. To solve these issues, they proposedEvidence\-based Post\-Hoc Adjustment Framework for Anomaly Detection \(EPHAD\), the first framework to adjust the prediction of an anomaly detector trained on contaminated data using an evidence function during test time\. Patra et al\.\[[24](https://arxiv.org/html/2607.05855#bib.bib24)\]focus on anomaly detection on visual data, but how to adaptEPHADto multimodal temporal and language data remains an open question\.
LLMsfor Social Media Analysis:LLM\-based social media analysis has been widely investigated, including post annotation\[[19](https://arxiv.org/html/2607.05855#bib.bib19),[34](https://arxiv.org/html/2607.05855#bib.bib34),[35](https://arxiv.org/html/2607.05855#bib.bib35)\], misinformation detection and mitigation\[[14](https://arxiv.org/html/2607.05855#bib.bib14),[26](https://arxiv.org/html/2607.05855#bib.bib26),[42](https://arxiv.org/html/2607.05855#bib.bib42),[43](https://arxiv.org/html/2607.05855#bib.bib43)\], andIOdetection\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\]\. Unlike\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\], which usesLLMsas the sole classifier based on behavioral and language patterns,TENSORleverages anLLMto provide additional evidence to enhance theTPP\-based anomaly detector\.
## 3Problem Definition
We formulateInformation Operation \(IO\)user detection as an unsupervised anomaly detection problem\. ASocial Network \(SN\)containsIOusers,*i\.e*\., users motivated or incentivized to promote anIO, and control users,*i\.e*\., normal genuine social media users\. The number ofIOusers is significantly smaller than that of normal users, and their behavioral and language patterns are noticeably different, as reflected in theirpost timeline\. The post timeline of useriiis𝒔i=\(si,1,si,2,⋯,si,Mi\)\\bm\{s\}\_\{i\}=\(s\_\{i,1\},s\_\{i,2\},\\cdots,s\_\{i,M\_\{i\}\}\), whereMiM\_\{i\}is the number of posts by this user\. Each post eventsi,k=\(ti,k,ci,k\)s\_\{i,k\}=\(t\_\{i,k\},c\_\{i,k\}\)consists of a timeti,kt\_\{i,k\}when this post is created and post contentci,kc\_\{i,k\}\. Post timelines of all users, including normal andIOusers, are denoted as𝑺=\{𝒔1,𝒔2,⋯,𝒔n\}\\bm\{S\}=\\\{\\bm\{s\}\_\{1\},\\bm\{s\}\_\{2\},\\cdots,\\bm\{s\}\_\{n\}\\\}\. For later use, we denote the respective sequences of all timestamps and contents as𝑻=\{𝒕1,𝒕2,…,𝒕n\}\\bm\{T\}=\\\{\\bm\{t\}\_\{1\},\\bm\{t\}\_\{2\},\\dots,\\bm\{t\}\_\{n\}\\\}and𝑪=\{𝒄1,𝒄2,…,𝒄n\}\\bm\{C\}=\\\{\\bm\{c\}\_\{1\},\\bm\{c\}\_\{2\},\\dots,\\bm\{c\}\_\{n\}\\\}, where𝒕i=\(ti,1,ti,2,⋯,ti,Mi\)\\bm\{t\}\_\{i\}=\(t\_\{i,1\},t\_\{i,2\},\\cdots,t\_\{i,M\_\{i\}\}\)and𝒄i=\(ci,1,ci,2,⋯,ci,Mi\)\\bm\{c\}\_\{i\}=\(c\_\{i,1\},c\_\{i,2\},\\cdots,c\_\{i,M\_\{i\}\}\)are sequences of timestamps and contents of userii\. We train a modelℳ\\mathcal\{M\}on𝑺\\bm\{S\}to identifyIOusers from control users\.IOusers are labeled 0, and control users are labeled 1, formulated as follows:
y=\{0,ℳ\(𝒕i,𝒄i\)⩾ϵ1,otherwisey=\\begin\{cases\}0,&\\mathcal\{M\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\\geqslant\\epsilon\\\\ 1,&\\text\{otherwise\}\\end\{cases\}\(1\)whereϵ\\epsilonis the threshold\. We show how we determine the value ofϵ\\epsilonin[Section˜5](https://arxiv.org/html/2607.05855#S5)\.
## 4Methodology
In this section, we introduceTemporal\-bEhavior\-laNguage Signals for information Operation Recognition \(TENSOR\), a novel anomaly detection approach that uses behavioral and language patterns from user action sequences for unsupervisedIOuser detection\.TENSORconsists of two core modules:Temporal Point Process \(TPP\)andLarge Language Model \(LLM\)\.TPPidentifiesIOusers using user action sequences, but its performance can be suboptimal because the training set is contaminated byIOusers\. To mitigate this issue, we propose a novel evidence function to convertLarge Language Model \(LLM\)outputs into a score and useEvidence\-based Post\-Hoc Adjustment Framework for Anomaly Detection \(EPHAD\)\[[24](https://arxiv.org/html/2607.05855#bib.bib24)\]to adjust the output of theTPPmodel using the evidence function\. In[Section˜4\.1](https://arxiv.org/html/2607.05855#S4.SS1)and[Section˜4\.2](https://arxiv.org/html/2607.05855#S4.SS2), we provide brief introductions toTPPandEPHAD\. In[Section˜4\.3](https://arxiv.org/html/2607.05855#S4.SS3), we present the structure and technical details ofTENSOR\.
### 4\.1TPPmodel
TheTemporal Point Process \(TPP\)describes a random process of an event sequence𝝉=\(τ1,τ2,⋯,τm\)\\bm\{\\tau\}=\(\\tau\_\{1\},\\tau\_\{2\},\\cdots,\\tau\_\{m\}\)whereτk\\tau\_\{k\}is the occurrence time\. This paper considers the simpleTPP, which only allows at most one event at any time, thusτk<τℓ\\tau\_\{k\}<\\tau\_\{\\ell\}ifk<ℓk<\\ell\. Given the history up to \(exclusive\) the current timett, denoted as𝓗\\bm\{\\mathcal\{H\}\}, theconditional intensity functionλ∗\(t\)\\lambda^\{\*\}\(t\)is the probability that an event will happen at timett\[[4](https://arxiv.org/html/2607.05855#bib.bib4)\]:222The asterisk denotes that this function conditions on history\.
λ∗\(t\)=limΔt→0P\(τ∈\(t,t\+Δt\]\|𝓗\)Δt\\lambda^\{\*\}\\left\(t\\right\)=\\lim\_\{\\Delta t\\rightarrow 0\}\{\\dfrac\{P\\left\(\\tau\\in\(t,t\+\\Delta t\]\\middle\|\\bm\{\\mathcal\{H\}\}\\right\)\}\{\\Delta t\}\}\(2\)Withλ∗\(t\)\\lambda^\{\*\}\(t\), we can define the joint probability distributionp∗\(t\)p^\{\*\}\(t\)of the next event attt\.
p∗\(t\)=λ∗\(t\)exp\(−∫tltλ∗\(τ\)dτ\)p^\{\*\}\\left\(t\\right\)=\\lambda^\{\*\}\\left\(t\\right\)\\exp\\left\(\-\\int\_\{t\_\{l\}\}^\{t\}\{\\lambda^\{\*\}\(\\tau\)\\mathrm\{d\}\\tau\}\\right\)\(3\)Thenegative log\-likelihood \(NLL\)loss on𝝉\\bm\{\\tau\}observed in a time interval\[t0,T\]\[t\_\{0\},T\]is:
L=−logp\(𝝉\)=−∑k=1Mlogλ∗\(τk\)\+∫t0Tλ∗\(u\)duL=\-\\log p\(\\bm\{\\tau\}\)=\-\\sum\_\{k=1\}^\{M\}\{\\log\\lambda^\{\*\}\(\\tau\_\{k\}\)\}\+\\int\_\{t\_\{0\}\}^\{T\}\{\\lambda^\{\*\}\(u\)\\mathrm\{d\}u\}\(4\)whereMMis the number of events in𝝉\\bm\{\\tau\}\.[Eq\.˜4](https://arxiv.org/html/2607.05855#S4.E4)is the training loss ofTPPmodels\. RecentTPPmodels are based on neural networks \(see\[[30](https://arxiv.org/html/2607.05855#bib.bib30)\]for a comprehensive survey\)\.TPPhas been used to detect abnormal events or sequences from normal events or event sequences\[[16](https://arxiv.org/html/2607.05855#bib.bib16),[29](https://arxiv.org/html/2607.05855#bib.bib29),[41](https://arxiv.org/html/2607.05855#bib.bib41)\]\.
### 4\.2Evidence\-based Post\-Hoc Adjustment Framework for Anomaly Detection
Anomaly Detection \(AD\)algorithms need a training dataset𝑫=\{x1,x2,⋯,xn\}\\bm\{D\}=\\\{x\_\{1\},x\_\{2\},\\cdots,x\_\{n\}\\\}containing normal samples to train anADclassifierf\(x\)f\(x\)\. If𝑫\\bm\{D\}is contaminated by abnormal data, the trainedADmodelf\(x\)f\(x\)will treat anomalies as normal\. Patra et al\.\[[24](https://arxiv.org/html/2607.05855#bib.bib24)\]observed thatADmodels trained on contaminated𝑫\\bm\{D\}can be corrected at test time using a predefined evidence functionT\(x\)T\(x\)\. TheT\(x\)T\(x\)contains external knowledge about anomalies, which is not captured byf\(x\)f\(x\)\. The method is calledEvidence\-based Post\-Hoc Adjustment Framework for Anomaly Detection \(EPHAD\)\. Specifically,EPHADrevisesf\(x\)f\(x\)usingexponential tiltingas:
f^\(x\)=f\(x\)exp\(T\(x\)/β\)Z\\hat\{f\}\(x\)=\\frac\{f\(x\)\\exp\\left\(T\(x\)/\\beta\\right\)\}\{Z\}\(5\)whereβ\>0\\beta\>0is the temperature, andZ=∫Df\(x\)exp\(T\(x\)/β\)𝑑xZ=\\int\_\{D\}\{f\(x\)\\exp\\left\(T\(x\)/\\beta\\right\)dx\}is the normalizing constant\. This method is theoretically grounded in Korbak et al\.\[[13](https://arxiv.org/html/2607.05855#bib.bib13)\]\. They treat the Post\-Hoc adjustment as a KL\-regularized optimization problem, where the objective is to find an adjusted distributionf^\(x\)\\hat\{f\}\(x\)that minimizes the KL divergence from the original distributionf\(x\)f\(x\)while maximizes the expected evidenceT\(x\)T\(x\), as shown in the following optimization problem:
maxf^𝔼x∼f^\[T\(x\)\]−βDKL\(f^∥f\)\\max\_\{\\hat\{f\}\}\\mathbb\{E\}\_\{x\\sim\\hat\{f\}\}\[T\(x\)\]\-\\beta D\_\{\\textrm\{KL\}\}\(\\hat\{f\}\\\|f\)\(6\)whereDKL\(f^∥f\)D\_\{\\textrm\{KL\}\}\(\\hat\{f\}\\\|f\)is the KL\-divergence betweenf^\\hat\{f\}andff\. They show that[Eq\.˜5](https://arxiv.org/html/2607.05855#S4.E5)is the solution to this optimization problem\. BecauseADonly depends on the relative ordering of samples, the intractable normalizing constantZZin[Eq\.˜5](https://arxiv.org/html/2607.05855#S4.E5)can be omitted, which simplifies[Eq\.˜5](https://arxiv.org/html/2607.05855#S4.E5)into:
f^\(x\)=f\(x\)exp\(T\(x\)/β\)\\hat\{f\}\(x\)=f\(x\)\\exp\\left\(T\(x\)/\\beta\\right\)\(7\)This is more practical for implementingADmodels\. In our approach elaborated in the following section, we use[Eq\.˜7](https://arxiv.org/html/2607.05855#S4.E7)to adjustf\(x\)f\(x\)based onTPPmodels trained on contaminated training data using the evidence function derived fromLLMoutputs for betterIOuser detection\.
### 4\.3Temporal\-bEhavior\-laNguage Signals for information Operation Recognition
Figure 1:Architecture ofTemporal\-bEhavior\-laNguage Signals for information Operation Recognition \(TENSOR\)\.In this section, we proposeTemporal\-bEhavior\-laNguage Signals for information Operation Recognition \(TENSOR\), which is sketched in[Fig\.˜1](https://arxiv.org/html/2607.05855#S4.F1)\.TENSORchecks whether a user is anIOuser based on their𝒕i\\bm\{t\}\_\{i\}and𝒄i\\bm\{c\}\_\{i\}\. The abstract inputxxin[Eq\.˜7](https://arxiv.org/html/2607.05855#S4.E7)is instantiated as a userxi=\(𝒕i,𝒄i\)x\_\{i\}=\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\), where𝒕i\\bm\{t\}\_\{i\}and𝒄i\\bm\{c\}\_\{i\}represent the sequence of timestamps and contents of userii, respectively\. It has two main components\. The first component is a frozenTPPmodel trained on𝑻\\bm\{T\}, which contains bothIOand control users\. TheTPPmodel captures the difference between control andIOusers using behavior patterns, including the coordinated behavior betweenIOaccounts or an excessive number of posts compared with control users\. The differences are reflected in the value−1Milogp\(𝒕i\)\-\\frac\{1\}\{M\_\{i\}\}\\log p\(\\bm\{t\}\_\{i\}\)\. Thus,f\(xi\)=exp\(−1Milogp\(𝒕i\)\)f\(x\_\{i\}\)=\\exp\\left\(\-\\frac\{1\}\{M\_\{i\}\}\\log p\(\\bm\{t\}\_\{i\}\)\\right\)is theTPP\-based score computed from the timestamp component ofxix\_\{i\}\. The second component is the averaged evidence functionT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\), backed by a frozenLLM\. It determines whether a user is anIOuser based on behavioral and language patterns in𝒄i\\bm\{c\}\_\{i\}and𝒕i\\bm\{t\}\_\{i\}\. That is,T\(xi\)=T¯N\(𝒕i,𝒄i\)T\(x\_\{i\}\)=\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)is the averaged evidence function overNNsamples derived from a frozenLLM\.−1Milogp\(𝒕i\)\-\\frac\{1\}\{M\_\{i\}\}\\log p\(\\bm\{t\}\_\{i\}\)andT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)are later fused usingEPHADto mitigate the effect ofTPPtrained on contaminated𝑻\\bm\{T\}\. The outputs of the two components are fused using[Eq\.˜7](https://arxiv.org/html/2607.05855#S4.E7)to mitigate the effect of data contamination inTPP\.
TENSORusesTPPmodels to classifyIOusers using the behavioral patterns collected from𝒕i\\bm\{t\}\_\{i\}\. Because there are more control users thanIOusers,TPPmodels yield higher probabilitiesp\(𝒕i\)p\(\\bm\{t\}\_\{i\}\)for control users compared toIOusers\. This makesTPPa validADclassifier:
y=\{0,if−1Milogp\(𝒕i\)⩾ϵ1,otherwisey=\\begin\{cases\}0,&\\text\{if\}\\ \-\\frac\{1\}\{M\_\{i\}\}\\log p\(\\bm\{t\}\_\{i\}\)\\geqslant\\epsilon\\\\ 1,&\\text\{otherwise\}\\end\{cases\}\(8\)whereIOusers are labeled0, control users are labeled11,ϵ\\epsilonis the threshold, andMiM\_\{i\}is the length of𝒕i\\bm\{t\}\_\{i\}\.
You are an advanced social media analyst specialized in detecting Information Operations \(IO\)\. Your objective is to analyze user profiles and posting behaviors to distinguish between state\-sponsored Information Operation \(IO\) accounts and control users based on the following behavioral frameworks:Role A: Information Operation \(IO\) AccountIdentity: These accounts are verified as part of inauthentic, coordinated efforts backed by state actors to manipulate public debate\.Tactics:\* Strategic Manipulation: They employ tactics like hashtag hijacking, artificial amplification, and the dissemination of propaganda or disinformation\.\* Targeting: They focus on specific audience communities, often using coordinated actions such as flooding through political cartoons or memes\.\* Profile Composition: They may consist of human\-operated accounts, automated bots, or compromised profiles that have been repurposed for a campaign\.Role B: Control AccountIdentity: These represent legitimate, organic users who act as a baseline for "normal" social media behavior\.Selection Context: These users are identified by their engagement in the same topics and hashtags as IO accounts during the same time frames, but without coordination\.Behavioral Characteristics:\* Authentic Engagement: They discuss similar political or social topics without endorsing or participating in an orchestrated state agenda\.\* Content Diversity: Their timelines include posts on unrelated personal or general topics, whereas IO accounts are often more single\-mindedly focused on campaign goals\.IO accounts are rare\. Most accounts are control accounts\.Please do your analysis step by step\. If you can think, think as thoroughly as possible to make your decision since the results are quite important\. Your thinking process can be long, but the final response should be concise: only answer whether this account is an IO account or a control account\.
Figure 2:Prompts for zero\-shotIOuser detectionHowever, the dataset𝑺\\bm\{S\}is contaminated byIOusers, which could harm detection accuracy\. To mitigate this,TENSORadjusts−1Milogp\(𝒕i\)\-\\frac\{1\}\{M\_\{i\}\}\\log p\(\\bm\{t\}\_\{i\}\)using aLLM\-based evidence functionT\(𝒕i,𝒄i\)T\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\. TheLLMtakes in the system prompt in[Fig\.˜2](https://arxiv.org/html/2607.05855#S4.F2)and a structured input of𝒕i\\bm\{t\}\_\{i\}and𝒄i\\bm\{c\}\_\{i\}\. It then generates a responseri=LLM\(𝒕i,𝒄i\)r\_\{i\}=\\mathrm\{LLM\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\), indicating whether the user is anIOuser or a control user by answering “IOaccount” or “Control account”\. To convertrir\_\{i\}intoT\(𝒕i,𝒄i\)T\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\), we use asemantic similarity mappingstrategy instead of enforcing strict text matching, which is prone to failure becauseLLMoutputs are non\-deterministic\. Specifically, we compute the semantic similarity between the rawLLMoutputrir\_\{i\}and reference texts representingIOor control users\. In our case, the reference texts are “IOaccount” forIOusers and “Control account” for control users, denoted asaioa\_\{io\}andacla\_\{cl\}\. By applying a𝚜𝚘𝚏𝚝𝚖𝚊𝚡\\mathtt\{softmax\}transformation to these similarity scores, we obtain a probability distribution over theIOand control users\.T\(𝒕i,𝒄i\)T\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)is defined as the probability of selecting theIOuser:
T\(𝒕i,𝒄i\)=exp\(sim\(aio,LLM\(𝒕i,𝒄i\)\)\)exp\(sim\(acl,LLM\(𝒕i,𝒄i\)\)\)\+exp\(sim\(aio,LLM\(𝒕i,𝒄i\)\)\)T\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)=\\frac\{\\exp\\left\(\\mathrm\{sim\}\(a\_\{io\},\\mathrm\{LLM\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\)\\right\)\}\{\\exp\\left\(\\mathrm\{sim\}\(a\_\{cl\},\\mathrm\{LLM\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\)\\right\)\+\\exp\\left\(\\mathrm\{sim\}\(a\_\{io\},\\mathrm\{LLM\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\)\\right\)\}\(9\)wheresim\(x,y\)\\mathrm\{sim\}\(x,y\)measures the similarity between input textxxandyy\. The output ofLLMsmay vary across multiple generations with the same input\. To ensure robustness, we drawNNsamples ofT\(𝒕i,𝒄i\)T\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)and compute their mean, denoted asT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\):
T¯N\(𝒕i,𝒄i\)=1N∑r=1NT\(r\)\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)=\\frac\{1\}\{N\}\\sum\_\{r=1\}^\{N\}\{T^\{\(r\)\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\}\(10\)whereT\(r\)\(𝒕i,𝒄i\)T^\{\(r\)\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)refers to therr\-th sample\. According to[Eq\.˜7](https://arxiv.org/html/2607.05855#S4.E7), the decision rule forTENSORis:
y=\{0,ifexp\(−1Milogp\(𝒕i\)\)exp\(T¯N\(𝒕i,𝒄i\)/β\)⩾ϵ1,otherwisey=\\begin\{cases\}0,&\\text\{if\}\\ \\exp\\left\(\-\\frac\{1\}\{M\_\{i\}\}\\log p\(\\bm\{t\}\_\{i\}\)\\right\)\\exp\\left\(\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)/\\beta\\right\)\\geqslant\\epsilon\\\\ 1,&\\text\{otherwise\}\\end\{cases\}\(11\)
## 5Experiments
This section \(i\) benchmarksTENSORagainst existing baselines, \(ii\) evaluates the impact of behavioral and language patterns onIOuser detection via ablation studies, \(iii\) comparesEPHADagainst alternative methods for integrating these patterns forIOuser detection, and \(iv\) analyzes the sensitivity ofTENSORto differentLLMsand temperature valuesβ\\beta\. We run each experiment on A100 and L40S GPUs three times with different random seeds, and their mean and standard deviation \(1\-sigma\) are reported\. The computational complexity ofTENSORfor one user isO\(Mi\)O\(M\_\{i\}\), because theTPPcomponent analyzes one sequence𝒔i\\bm\{s\}\_\{i\}inO\(Mi\)O\(M\_\{i\}\), while theLLMcomponent processes one sequence inO\(1\)O\(1\)333However,LLMscan be the speed bottleneck ofTENSORbecause they are usually much slower thanTPPs\. Hence, it is possible thatTENSORtakes a constant time to process a sequence, even though the computational complexity isO\(Mi\)O\(M\_\{i\}\)\.\.
TENSORis supposed to be training\-free because all its components are either frozen or deterministic computations\. However, to the best of our knowledge, largeTPPmodels pretrained on large event sequence data do not exist, so we must train aTPPmodel on𝑻\\bm\{T\}beforehand\.TENSORworks with any existingTPPmodel that providesp\(𝒕i\)p\(\\bm\{t\}\_\{i\}\)\. According to\[[15](https://arxiv.org/html/2607.05855#bib.bib15),[29](https://arxiv.org/html/2607.05855#bib.bib29)\], the performance gap between existing state\-of\-the\-artTPPmodels is small\. Without loss of generality, we choose theSelf\-attentive Hawkes Process \(SAHP\)\[[40](https://arxiv.org/html/2607.05855#bib.bib40)\]because of its relatively simple design and good performance\. The loss function for trainingSAHPon𝑻\\bm\{T\}is[Eq\.˜4](https://arxiv.org/html/2607.05855#S4.E4)\.
TENSORworks with all existingLLMs\. Recently, we have seen the rise of reasoning models\[[11](https://arxiv.org/html/2607.05855#bib.bib11)\]\. By enabling theLLMto think during test time, reasoning models consistently improve performance across various tasks, especially for solving mathematical problems and difficult logic problems\[[22](https://arxiv.org/html/2607.05855#bib.bib22)\]\. In this paper, we use an open\-weight reasoning modelgpt\-oss\-120B\(reasoning model, 117B parameters with 5\.1B active parameters\)\[[23](https://arxiv.org/html/2607.05855#bib.bib23)\]with a medium reasoning effort\. In[Section˜5\.4](https://arxiv.org/html/2607.05855#S5.SS4), we reportTENSOR’s performance withllama3\.3\-70B\(non\-reasoning model, 70B parameters\),qwen\-next\-80B\(reasoning model, 80B parameters with 3B active parameters\)\[[33](https://arxiv.org/html/2607.05855#bib.bib33)\],glm\-4\.5\-air\(reasoning model, 106B parameters with 12B active parameters\)\[[32](https://arxiv.org/html/2607.05855#bib.bib32)\], andmistral3\.2\-24b\(non\-reasoning model, 24B parameters\)444https://huggingface\.co/mistralai/Mistral\-Small\-3\.2\-24B\-Instruct\-2506\. As for the similarity function in[Eq\.˜9](https://arxiv.org/html/2607.05855#S4.E9), we use the bge\-m3\-v2 reranker\[[3](https://arxiv.org/html/2607.05855#bib.bib3)\]\. TheNNin[Eq\.˜10](https://arxiv.org/html/2607.05855#S4.E10)is 5\.
TENSORhas one hyperparameter: temperatureβ\\beta\. According to Patra et al\.\[[24](https://arxiv.org/html/2607.05855#bib.bib24)\], we setβ=0\.5\\beta=0\.5during the experiments\. In[Section˜5\.4](https://arxiv.org/html/2607.05855#S5.SS4), we investigate how temperature affects the detection performance ofTENSOR\.
##### Baseline Models:
We compareTENSORwith three baselines\. Although not directly aimed atIOuser detection, two studies concerningIOusers verify coordination among users in the sameIO\[[12](https://arxiv.org/html/2607.05855#bib.bib12),[21](https://arxiv.org/html/2607.05855#bib.bib21)\]\. These studies motivate the first two baselines\.
- •User Embedding Clustering \(Clustering\)\[[12](https://arxiv.org/html/2607.05855#bib.bib12)\]represents each user via a user embedding, generated by averaging the embeddings of all events for that user extracted from a model trained on𝒯\\mathcal\{T\}\. Then, these user\-level representations are fed into a clustering model to partition the users into two distinct groups\. Users in the smaller group areIOusers\.
- •Behavioral Languages for Online Characterization \(BLOC\)\[[21](https://arxiv.org/html/2607.05855#bib.bib21)\]suggests describing user behaviors as strings of symbols using formal languages specified by rules\. Next, these sequences are converted into user embeddings using a TF\-IDF model\. Users whose embeddings differ from the majority are consideredIOusers\.
- •LLM\-basedIOuser detection \(LLM\)\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\]usesLLMsto detect in a zero\-shot setting whether a user is anIOuser based on behavioral and language patterns\. This method first checks whether a postxix\_\{i\}belongs to anIO\. Then users whose posts are mostlyIOposts are labelled asIOusers\.
##### Datasets:
We use theIOuser dataset collected by Seckin et al\.\[[28](https://arxiv.org/html/2607.05855#bib.bib28)\]\. This dataset contains various identifiedIOsin 26 labelled datasets across 15 countries\. We evaluateTENSORand baselines on Egypt, China\_1, Iran\_1, Russia\_1, and UAE\. These datasets have the complete activity timelines of allIOusers, but control users are limited to their last 100 posts on days when they are involved inIOs\. To mitigate this bias, we apply the same strategy toIOusers by limiting the number of posts fromIOusers to 100 per day\. The data statistics of curated datasets are available in[Table˜1](https://arxiv.org/html/2607.05855#S5.T1)\. We assign 80% of data to the training set, 10% to the validation set, and 10% to the test set\. Please note that although these datasets include labels indicating which users areIOusers, they are used solely for performance evaluations\.TENSORand baselines cannot see the labels\.
Table 1:Statistics of the curatedIOdatasetsNumber ofIOusersNumber ofcontrol usersNumber ofIOpostsNumber ofcontrol postsEgypt21924266,57714,204China\_153728,445169,2871,718,924Iran\_15433,291247,386192,674Russia\_12,83020,9611,335,0641,595,514UAE3,3376,6351,244,984366,873Total7,46659,5743,063,2983,888,189
##### Evaluation Metrics:
We evaluateTENSORand baselines using precision, recall, F1\-score,Area Under the receiver operating characteristic Curve \(AUC\), andArea Under the Precision\-Recall Curve \(AUPRC\)555AUPRCis often referred to as Average Precision \(AP\) in machine learning toolkits\. We prefer the termAUPRCas it more accurately describes the geometric calculation of the metric\.\. We only reportAUPRCfor experiments in[Section˜5\.2](https://arxiv.org/html/2607.05855#S5.SS2),[Section˜5\.3](https://arxiv.org/html/2607.05855#S5.SS3), and[Section˜5\.4](https://arxiv.org/html/2607.05855#S5.SS4)to simplify the comparison of results\. We addAUPRCbecause Davis et al\.\[[5](https://arxiv.org/html/2607.05855#bib.bib5)\]suggestAUPRCas an alternative toAUCfor tasks with a largely imbalanced class distribution, whereAUCresults can be overly optimistic\. TheIOdataset, as shown in[Table˜1](https://arxiv.org/html/2607.05855#S5.T1), is highly imbalanced, with fewIOusers and many control users, which justifies the use ofAUPRC\.
Computing precision, recall, and F1\-score requires a thresholdϵ\\epsilonto decide the labelyy\. In this work, the threshold is adjusted by maximizing the F1 on the validation set\. In practice, the scores ofIOand control users are often modeled as two class\-conditional normal distributions\. The threshold is the score that separates them, that is, most sequences on one side belong toIOusers and most sequences on the other side belong to control users\.
### 5\.1ComparingTENSORwith baselines
This section comparesTENSORwith three baselines,Clustering,BLOC, andLLMon fiveIOdatasets\. The metrics used are precision, recall, F1\-score,AUCandAUPRC\. The results are presented in[Tables2](https://arxiv.org/html/2607.05855#S5.T2)to[6](https://arxiv.org/html/2607.05855#S5.T6)\.
Table 2:The precision of TENSOR and baselines on five real\-world IO datasets \(higher is better\)\.TENSORClusteringBLOCLLMEgypt0\.6138±\\pm0\.00550\.8056±\\pm0\.06000\.4761±\\pm0\.00000\.6882±\\pm0\.0316China\_10\.8361±\\pm0\.01830\.2649±\\pm0\.34950\.0321±\\pm0\.00000\.1190±\\pm0\.0018Iran\_10\.8332±\\pm0\.01460\.9198±\\pm0\.04430\.1440±\\pm0\.00000\.2928±\\pm0\.0046Russia\_10\.6644±\\pm0\.02900\.1218±\\pm0\.00490\.1127±\\pm0\.00000\.2106±\\pm0\.0075UAE0\.7864±\\pm0\.00790\.5824±\\pm0\.14520\.3357±\\pm0\.00000\.4556±\\pm0\.0048Average0\.74680\.53890\.22010\.3532Table 3:The recall of TENSOR and baselines on five real\-world IO datasets \(higher is better\)\.TENSORClusteringBLOCLLMEgypt0\.9394±\\pm0\.02140\.6818±\\pm0\.03710\.9091±\\pm0\.00000\.8636±\\pm0\.0000China\_10\.4691±\\pm0\.01750\.8889±\\pm0\.09440\.0926±\\pm0\.00000\.6605±\\pm0\.0087Iran\_10\.5152±\\pm0\.01710\.5697±\\pm0\.14651\.0000±\\pm0\.00000\.6848±\\pm0\.0086Russia\_10\.6419±\\pm0\.01920\.9435±\\pm0\.07250\.8657±\\pm0\.00000\.2874±\\pm0\.0159UAE0\.7784±\\pm0\.00980\.6766±\\pm0\.10451\.0000±\\pm0\.00000\.3832±\\pm0\.0000Average0\.66880\.75210\.77350\.5759Table 4:The F1 score of TENSOR and baselines on five real\-world IO datasets \(higher is better\)\.TENSORClusteringBLOCLLMEgypt0\.7424±\\pm0\.01070\.7381±\\pm0\.04420\.6250±\\pm0\.00000\.7656±\\pm0\.0194China\_10\.6006±\\pm0\.01120\.2764±\\pm0\.34150\.0476±\\pm0\.00000\.2017±\\pm0\.0026Iran\_10\.6366±\\pm0\.01650\.6895±\\pm0\.11780\.2517±\\pm0\.00000\.4102±\\pm0\.0055Russia\_10\.6521±\\pm0\.00440\.2154±\\pm0\.00550\.1995±\\pm0\.00000\.2430±\\pm0\.0103UAE0\.7823±\\pm0\.00640\.6247±\\pm0\.12820\.5026±\\pm0\.00000\.7656±\\pm0\.0194Average0\.68280\.50880\.32530\.4772Table 5:The AUC of TENSOR and baselines on five real\-world IO datasets \(higher is better\)\.TENSORClusteringBLOCLLMEgypt0\.7806±\\pm0\.00730\.7570±\\pm0\.03950\.3636±\\pm0\.00000\.7558±\\pm0\.0306China\_10\.9338±\\pm0\.00190\.3645±\\pm0\.42070\.5872±\\pm0\.00000\.9057±\\pm0\.0064Iran\_10\.8558±\\pm0\.00720\.6676±\\pm0\.09960\.4905±\\pm0\.00000\.7698±\\pm0\.0071Russia\_10\.9213±\\pm0\.00260\.3017±\\pm0\.09670\.3946±\\pm0\.00000\.7019±\\pm0\.0024UAE0\.9263±\\pm0\.00050\.7403±\\pm0\.12540\.3426±\\pm0\.00000\.7558±\\pm0\.0306Average0\.88360\.56620\.43570\.7778Table 6:The AUPRC of TENSOR and baselines on five real\-world IO datasets \(higher is better\)\.TENSORClusteringBLOCLLMEgypt0\.7690±\\pm0\.00870\.8232±\\pm0\.01600\.3839±\\pm0\.00000\.6859±\\pm0\.0284China\_10\.5654±\\pm0\.00690\.2712±\\pm0\.36940\.0288±\\pm0\.00000\.1670±\\pm0\.0053Iran\_10\.6910±\\pm0\.03360\.6711±\\pm0\.09480\.1352±\\pm0\.00000\.3075±\\pm0\.0071Russia\_10\.6952±\\pm0\.01330\.0812±\\pm0\.01030\.0941±\\pm0\.00000\.1848±\\pm0\.0006UAE0\.8562±\\pm0\.00030\.6783±\\pm0\.14740\.2475±\\pm0\.00000\.6859±\\pm0\.0284Average0\.71540\.50500\.17790\.4062The results show thatTENSORis the overall best approach forIOuser detection, as measured byAUCandAUPRC\. One outlier is Egypt, whereClusteringperforms better onAUPRC\. The reason is the temperatureβ\\beta\. Later results in[Section˜5\.4](https://arxiv.org/html/2607.05855#S5.SS4.SSS0.Px2)show thatβ=0\.5\\beta=0\.5from\[[24](https://arxiv.org/html/2607.05855#bib.bib24)\]is not the best choice\. By lowering the temperature, the overall performance ofTENSORacross all datasets significantly improves and can outperformClusteringon all datasets\. We do not discuss how to optimizeβ\\betawithout any labels in this paper\.
We also observe that the results ofClusteringhave a large standard deviation across multiple datasets\. The reason is the data imbalance with many control users and fewIOusers\. As shown in[Fig\.˜3](https://arxiv.org/html/2607.05855#S5.F3),Clusteringon imbalanced data is sensitive to random seeds, leading to the misclassification of control users asIOusers\. This instability results in inconsistent performance, particularly on highly imbalanced datasets such as China\_1 and Russia\_1\.
Figure 3:Impact of data imbalance onClustering’s performance\. Data points represent control users \(C0\) andIOusers \(C1\)\. Results indicate that significant data imbalance shifts cluster centroids with some random seeds, leading to performance inconsistency\.
### 5\.2The impact of behavioral and language patterns onIOuser detection
If we only use the behavioral or language patterns of one user forIOuser detection, we may get suboptimal results\. To demonstrate this, we compareTENSORwith the following ablation baselines: \(i\)TENSORwithout the evidence functionT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\), which is described in[Eq\.˜8](https://arxiv.org/html/2607.05855#S4.E8), \(ii\)TENSORwithoutTPP, detectingIOusers byT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\), \(iii\)TENSORbutT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)returns random samples from the uniform distribution𝒰\(0,1\)\\mathcal\{U\}\(0,1\), and \(iv\) the Random model, which assigns scores sampled from𝒰\(0,1\)\\mathcal\{U\}\(0,1\)to users\. We include Random becauseAUPRCis sensitive to the ratio of control andIOusers, whereas theAUCremains at 0\.5\. The results are reported in[Table˜7](https://arxiv.org/html/2607.05855#S5.T7)\.
Table 7:TheAUPRCofTENSORand ablation baselines on five real\-worldIOdatasets \(temperatureβ=0\.5\\beta=0\.5,LLMisgpt\-oss\-120B, higher is better\)\.TENSORTENSORwithoutLLMTENSORwithoutTPPTENSORwith randomT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)RandomEgypt0\.7690±\\pm0\.00870\.6916±\\pm0\.00300\.5839±\\pm0\.01930\.6770±\\pm0\.05070\.5083±\\pm0\.0720China\_10\.5654±\\pm0\.00690\.4740±\\pm0\.00460\.0669±\\pm0\.00440\.4721±\\pm0\.01200\.0212±\\pm0\.0047Iran\_10\.6910±\\pm0\.03360\.6438±\\pm0\.02520\.1989±\\pm0\.01800\.6085±\\pm0\.01910\.1558±\\pm0\.0220Russia\_10\.6953±\\pm0\.01330\.6425±\\pm0\.00730\.1707±\\pm0\.00090\.6302±\\pm0\.01220\.1216±\\pm0\.0073UAE0\.8562±\\pm0\.00030\.8009±\\pm0\.00020\.4080±\\pm0\.00610\.7619±\\pm0\.00600\.3390±\\pm0\.0152Average0\.71540\.65060\.28570\.62990\.2292
The ablation results demonstrate that jointly considering behavioral and language patterns forIOuser detection is better than considering only one, as theAUPRCofTENSORis significantly better thanTENSORwithoutLLMorTPP\. The experiment results also show that a standaloneLLMperforms poorly at zero\-shotIOuser detection\. Another observation is thatTENSORoutperformsTENSORwith randomT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\), whose performance is even worse thanTENSORwithoutLLM\. This demonstrates thatLLMstill provides useful information from𝒄i\\bm\{c\}\_\{i\}and𝒕i\\bm\{t\}\_\{i\}forIOuser detection despite its poor performance in this task\.
### 5\.3The benefit ofLLM\-backedT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)
EPHADenablesTENSORto jointly consider behavioral and language patterns forIOuser detection, but is it the overall best way in terms of performance? In this section, we investigate several intuitive and existing alternatives to train aIOuser detector based on behavioral and language patterns\. Specifically, they are: \(i\) behavior\- and language\-awareTPPmodel\. The input ofTPPis𝒕i\\bm\{t\}\_\{i\}and𝒄i\\bm\{c\}\_\{i\}\. The training loss is−logp\(𝒕i,𝒄i\)\-\\log p\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\. Then we use[Eq\.˜8](https://arxiv.org/html/2607.05855#S4.E8)forIOuser detection, \(ii\)BLOC\[[21](https://arxiv.org/html/2607.05855#bib.bib21)\], \(iii\)LLM\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\], and \(iv\) the Random model\. The results are reported in[Table˜8](https://arxiv.org/html/2607.05855#S5.T8)\.
Table 8:TheAUPRCofTENSORand other behavior\- and language\-aware baselines \(temperatureβ=0\.5\\beta=0\.5,LLMisgpt\-oss\-120B, higher is better\)\.TENSORbehavior\- andlanguage\-awareTPPBLOC\[[21](https://arxiv.org/html/2607.05855#bib.bib21)\]LLM\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\]RandomEgypt0\.7690±\\pm0\.00870\.6949±\\pm0\.00170\.3838±\\pm0\.01600\.6859±\\pm0\.02840\.5083±\\pm0\.0720China\_10\.5654±\\pm0\.00690\.4674±\\pm0\.00420\.0288±\\pm0\.00000\.1670±\\pm0\.00530\.0212±\\pm0\.0047Iran\_10\.6910±\\pm0\.03360\.6290±\\pm0\.00520\.1352±\\pm0\.00000\.3075±\\pm0\.00710\.1558±\\pm0\.0220Russia\_10\.6953±\\pm0\.01330\.6242±\\pm0\.01450\.0940±\\pm0\.00000\.1848±\\pm0\.00060\.1216±\\pm0\.0073UAE0\.8562±\\pm0\.00030\.8011±\\pm0\.00120\.2475±\\pm0\.00000\.4187±\\pm0\.00200\.3390±\\pm0\.0152Average0\.71540\.64330\.17790\.35280\.2292
We see thatTENSORoutperforms other alternatives by a significant margin on all datasets\. A detailed analysis shows thatBLOCperforms worse than the Random baseline\. A potential reason is thatBLOCrequires more detailed information about posts, such as who the post is sent to and whether it contains a picture, which is missing from the dataset\. This results in most generatedBLOCsequences containing only one symbol referring to a post, which significantly harms their classification capability\.LLMand behavior\- and language\-awareTPPoutperform Random but are outperformed byTENSOR\. On closer inspection, we find that the results of the behavior\- and language\-awareTPPare basically the same as those ofTPP\. These results demonstrate that adding language data to theTPPmodel does not improve itsIOdetection capability\.
### 5\.4Sensitivity to differentLLMsandβ\\beta
All reported results ofTENSORin previous sections are based on temperatureβ=0\.5\\beta=0\.5inherited from theEPHADpaper andgpt\-oss\-120B\. However,TENSORworks with all existingLLMsand temperaturesβ\>0\\beta\>0\. In this section, we evaluate the stability ofTENSORunder differentLLMsand temperatures\.
##### Sensitivity toLLMs:
Table 9:TheAUPRCofTENSORwith differentLLMsinT¯N\(𝒕i,𝒄i\)\\bar\{T\}\_\{N\}\(\\bm\{t\}\_\{i\},\\bm\{c\}\_\{i\}\)\(temperatureβ=0\.5\\beta=0\.5, higher is better\)\.TENSORw/gpt\-oss\-120BTENSORw/llama3\.3\-70BTENSORw/qwen\-next\-80BTENSORw/glm\-4\.5\-airTENSORw/mistral3\.2\-24bTENSORwithoutLLMEgypt0\.7690±\\pm0\.00870\.7079±\\pm0\.00190\.7373±\\pm0\.00840\.7467±\\pm0\.00170\.6806±\\pm0\.00570\.6916±\\pm0\.0030China\_10\.5654±\\pm0\.00690\.5224±\\pm0\.00620\.5251±\\pm0\.00300\.5609±\\pm0\.00350\.5085±\\pm0\.00440\.4740±\\pm0\.0046Iran\_10\.6910±\\pm0\.03360\.6622±\\pm0\.02690\.6711±\\pm0\.09480\.6737±\\pm0\.03450\.6480±\\pm0\.03530\.6438±\\pm0\.0252Russia\_10\.6953±\\pm0\.01330\.6530±\\pm0\.01030\.6358±\\pm0\.01310\.6475±\\pm0\.01220\.6495±\\pm0\.00810\.6425±\\pm0\.0073UAE0\.8562±\\pm0\.00030\.8154±\\pm0\.00220\.7850±\\pm0\.00180\.8152±\\pm0\.00070\.7804±\\pm0\.00090\.8009±\\pm0\.0002Average0\.71540\.66700\.67090\.68880\.65340\.6506
[Table˜9](https://arxiv.org/html/2607.05855#S5.T9)shows the performance ofTENSORwith differentLLMs\. Besidesgpt\-oss\-120B, we pickllama3\.3\-70B,qwen\-next\-80B,glm\-4\.5\-air, andmistral3\.2\-24b\. We observe that configurations usinggpt\-oss\-120B,llama3\.3\-70B, andglm\-4\.5\-airconsistently outperformTENSORwithoutLLMacross all datasets\. Althoughqwen\-next\-80Bandmistral3\.2\-24bunderperform on some datasets because of their smaller model or active parameter sizes, their overall average is still better thanTENSORwithoutLLM\. These results show thatTENSORis stable across differentLLMs\.
##### Sensitivity toβ\\beta:
[Fig\.˜4](https://arxiv.org/html/2607.05855#S5.F4)shows how temperatureβ\\betaaffects the performance ofTENSOR\. We observe that although the optimal temperature varies across datasets, it is consistently around 0\.3 forgpt\-oss\-120B\. Generally, increasing the temperature from 0\.05 brings a quick performance improvement to a maximum, followed by a gradual decline\. For China\_1, Iran\_1, Russia\_1, and UAE, the performance drop is relatively small, while for Egypt, theAUPRCdrops significantly from over 0\.85 to below 0\.70\. Table[10](https://arxiv.org/html/2607.05855#S5.T10)comparesTENSORat its optimal temperature against the default configuration and other baselines\.TENSORconsistently outperforms other approaches at the optimal temperature\. These results indicate that unsupervised optimization ofβ\\betais a promising approach to further improve the performance ofTENSOR\. We leave this as future work\.
Egypt
China\_1
Iran\_1
Russia\_1
UAE
Figure 4:AUPRCofTENSORon all fiveIOdatasets with different temperatures from 0\.05 to 2\.Table 10:TheAUPRCofTENSORand other behavior\- and language\-aware baselines \(temperatureβ=0\.5\\beta=0\.5forTENSORresults in the second column,LLMisgpt\-oss\-120B, higher is better\)\.TENSORwith bestβ\\betaTENSORClustering\[[12](https://arxiv.org/html/2607.05855#bib.bib12)\]BLOC\[[21](https://arxiv.org/html/2607.05855#bib.bib21)\]LLM\[[17](https://arxiv.org/html/2607.05855#bib.bib17)\]RandomEgypt0\.8568±\\pm0\.00810\.7690±\\pm0\.00870\.8232±\\pm0\.01600\.3838±\\pm0\.01600\.6859±\\pm0\.02840\.5083±\\pm0\.0720China\_10\.5796±\\pm0\.01750\.5654±\\pm0\.00690\.2712±\\pm0\.36940\.0288±\\pm0\.00000\.1670±\\pm0\.00530\.0212±\\pm0\.0047Iran\_10\.7108±\\pm0\.02710\.6910±\\pm0\.03360\.6711±\\pm0\.09480\.1352±\\pm0\.00000\.3075±\\pm0\.00710\.1558±\\pm0\.0220Russia\_10\.7045±\\pm0\.00920\.6953±\\pm0\.01330\.0812±\\pm0\.01030\.0940±\\pm0\.00000\.1848±\\pm0\.00060\.1216±\\pm0\.0073UAE0\.8743±\\pm0\.00070\.8562±\\pm0\.00030\.6783±\\pm0\.14740\.2475±\\pm0\.00000\.4187±\\pm0\.00200\.3390±\\pm0\.0152Average0\.74520\.71540\.50500\.17790\.35280\.2292
## 6Conclusion
In this study, we addressed the critical challenge of identifyingIOusers within social networks\. While traditional supervised models struggle with the evolving nature ofIOuser behaviors, unsupervised models rely on rigid assumptions of coordination,TENSORoffers a more resilient alternative by framing detection as a multimodal anomaly problem\.TENSORshifts the focus toward the inherent friction betweenIOcampaigns and organic user behaviors\. By deploying aTPP, we successfully captured the behavioral patterns that distinguish coordinated influence from genuine social interaction\. A key contribution of this work is the mitigation of “training set contamination\.” By integrating a novel evidence function that translatesLLMresponses, which are generated from users’ post timelines, into quantitative scores, we effectively adjust the output ofTPPto mitigate the noise caused byIOusers embedded in the training data\. Experimental results show thatTENSORoutperforms other baselines by a significant margin on five real\-worldIOdatasets\.\{credits\}
#### 6\.0\.1Acknowledgements
This research is supported in part by the Australian Research Council \(ARC\) Discovery Projects DP200101441 and DP210100743\. We appreciate the compute and LLM services provided by RMIT RACE Hub\.
## References
- \[1\]Addawood, A\., Badawy, A\., Lerman, K\., Ferrara, E\.: Linguistic Cues to Deception: Identifying Political Trolls on Social Media\. Proceedings of the International AAAI Conference on Web and Social Media13, 15–25 \(2019\)
- \[2\]Alizadeh, M\., Shapiro, J\.N\., Buntain, C\., Tucker, J\.A\.: Content\-based features predict social media influence operations\. Science Advances6\(30\) \(2020\)
- \[3\]Chen, J\., Xiao, S\., Zhang, P\., Luo, K\., Lian, D\., Liu, Z\.: M3\-Embedding: Multi\-Linguality, Multi\-Functionality, Multi\-Granularity Text Embeddings Through Self\-Knowledge Distillation\. In: ACL 2024\. pp\. 2318–2335 \(2024\)
- \[4\]Daley, D\.J\., Vere\-Jones, D\. \(eds\.\): An Introduction to the Theory of Point Processes Volume I: Elementary Theory and Methods\. Probability and its Applications, Springer, 2 edn\. \(2003\)
- \[5\]Davis, J\., Goadrich, M\.: The relationship between Precision\-Recall and ROC curves\. In: ICML 2006\. pp\. 233–240 \(2006\)
- \[6\]Ezzeddine, F\., Ayoub, O\., Giordano, S\., Nogara, G\., Sbeity, I\., Ferrara, E\., Luceri, L\.: Exposing influence campaigns in the age of LLMs: a behavioral\-based AI approach to detecting state\-sponsored trolls\. EPJ Data Science12\(1\), 46 \(2023\)
- \[7\]Forum, W\.E\.: The global risks report 2024: insight report\. World Economic Forum, 19th ed edn\. \(2024\), oCLC: 1499197716
- \[8\]Haider, S\., Luceri, L\., Deb, A\., Badawy, A\., Peng, N\., Ferrara, E\.: Detecting Social Media Manipulation in Low\-Resource Languages\. In: WWW 2023 Companion\. pp\. 1358–1364 \(2023\)
- \[9\]Im, J\., Chandrasekharan, E\., Sargent, J\., Lighthammer, P\., Denby, T\., Bhargava, A\., Hemphill, L\., Jurgens, D\., Gilbert, E\.: Still out there: Modeling and Identifying Russian Troll Accounts on Twitter\. In: WebSci 2020\. pp\. 1–10 \(2020\)
- \[10\]Jachim, P\., Sharevski, F\., Treebridge, P\.: TrollHunter \[Evader\]: Automated Detection \[Evasion\] of Twitter Trolls During the COVID\-19 Pandemic \(2020\), arXiv:2012\.02586
- \[11\]Ke, Z\., Jiao, F\., Ming, Y\., Nguyen, X\.P\., Xu, A\., Long, D\.X\., Li, M\., Qin, C\., Wang, P\., Savarese, S\., Xiong, C\., Joty, S\.: A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems\. Transactions on Machine Learning Research \(2025\)
- \[12\]Kong, Q\., Calderon, P\., Ram, R\., Boichak, O\., Rizoiu, M\.A\.: Interval\-censored Transformer Hawkes: Detecting Information Operations using the Reaction of Social Systems\. In: Proceedings of the ACM Web Conference 2023\. pp\. 1813–1821\. WWW ’23 \(2023\)
- \[13\]Korbak, T\., Perez, E\., Buckley, C\.: RL with KL penalties is better viewed as Bayesian inference\. In: EMNLP 2022\. pp\. 1083–1091\. Association for Computational Linguistics \(2022\)
- \[14\]Kumar, R\., Goddu, B\., Saha, S\., Jatowt, A\.: Silver Lining in the Fake News Cloud: Can Large Language Models Help Detect Misinformation? IEEE Transactions on Artificial Intelligence6\(1\), 14–24 \(2025\)
- \[15\]Lin, H\., Wu, L\., Zhao, G\., Pai, L\., Li, S\.Z\.: Exploring Generative Neural Temporal Point Process\. TMLR \(2022\)
- \[16\]Liu, S\., Hauskrecht, M\.: Event Outlier Detection in Continuous Time\. In: Proceedings of the 38th International Conference on Machine Learning\. pp\. 6793–6803\. PMLR \(2021\)
- \[17\]Luceri, L\., Boniardi, E\., Ferrara, E\.: Leveraging Large Language Models to Detect Influence Campaigns on Social Media\. In: Companion Proceedings of the ACM Web Conference 2024\. pp\. 1459–1467 \(2024\)
- \[18\]Luceri, L\., Giordano, S\., Ferrara, E\.: Detecting Troll Behavior via Inverse Reinforcement Learning: A Case Study of Russian Trolls in the 2016 US Election\. Proceedings of the International AAAI Conference on Web and Social Media14, 417–427 \(2020\)
- \[19\]Malik, U\., Bernard, S\., Pauchet, A\., Chatelain, C\., Picot\-Clémente, R\., Cortinovis, J\.: Pseudo\-Labeling With Large Language Models for Multi\-Label Emotion Classification of French Tweets\. IEEE Access12, 15902–15916 \(2024\)
- \[20\]Minici, M\., Luceri, L\., Fabbri, F\., Ferrara, E\.: IOHunter: Graph Foundation Model to Uncover Online Information Operations\. Proceedings of the AAAI Conference on Artificial Intelligence39\(27\), 28258–28266 \(2025\)
- \[21\]Nwala, A\.C\., Flammini, A\., Menczer, F\.: A language framework for modeling social media account behavior\. EPJ Data Science12\(1\), 33 \(2023\)
- \[22\]OpenAI: OpenAI o1 System Card \(2024\), arXiv:2412\.16720
- \[23\]OpenAI: gpt\-oss\-120b & gpt\-oss\-20b Model Card \(2025\), arXiv:2508\.10925
- \[24\]Patra, S\., Taieb, S\.B\.: An Evidence\-Based Post\-Hoc Adjustment Framework for Anomaly Detection Under Data Contamination\. In: NeurIPS 2025 \(2025\)
- \[25\]Perini, L\., Bürkner, P\.C\., Klami, A\.: Estimating the Contamination Factor’s Distribution in Unsupervised Anomaly Detection\. In: Proceedings of the 40th International Conference on Machine Learning\. pp\. 27668–27679\. PMLR \(2023\)
- \[26\]Qi, P\., Yan, Z\., Hsu, W\., Lee, M\.L\.: SNIFFER: Multimodal Large Language Model for Explainable Out\-of\-Context Misinformation Detection\. In: CVPR\. pp\. 13052–13062 \(2024\)
- \[27\]Qiu, C\., Li, A\., Kloft, M\., Rudolph, M\., Mandt, S\.: Latent Outlier Exposure for Anomaly Detection with Contaminated Data\. In: Proceedings of the 39th International Conference on Machine Learning\. pp\. 18153–18167\. PMLR \(2022\)
- \[28\]Seckin, O\.C\., Pote, M\., Nwala, A\.C\., Yin, L\., Luceri, L\., Flammini, A\., Menczer, F\.: Labeled Datasets for Research on Information Operations\. Proceedings of the International AAAI Conference on Web and Social Media19, 2567–2574 \(2025\)
- \[29\]Shchur, O\., Turkmen, A\.C\., Januschowski, T\., Gasthaus, J\., Günnemann, S\.: Detecting Anomalous Event Sequences with Temporal Point Processes\. In: NeurIPS 2021 \(2021\)
- \[30\]Shchur, O\., Türkmen, A\.C\., Januschowski, T\., Günnemann, S\.: Neural Temporal Point Processes: A Review\. In: IJCAI 2021\. pp\. 4585–4593 \(2021\)
- \[31\]Smith, S\.T\., Kao, E\.K\., Mackin, E\.D\., Shah, D\.C\., Simek, O\., Rubin, D\.B\.: Automatic detection of influential actors in disinformation networks\. Proceedings of the National Academy of Sciences118\(4\) \(2021\)
- \[32\]Team, G\.\.: GLM\-4\.5: Agentic, Reasoning, and Coding \(ARC\) Foundation Models \(2025\), arXiv:2508\.06471
- \[33\]Team, Q\.: Qwen3 technical report \(2025\),[https://arxiv\.org/abs/2505\.09388](https://arxiv.org/abs/2505.09388)
- \[34\]Tekumalla, R\., Banda, J\.M\.: Leveraging Large Language Models and Weak Supervision for Social Media Data Annotation: An Evaluation Using COVID\-19 Self\-reported Vaccination Tweets\. In: HCI International 2023 \- Late Breaking Papers\. pp\. 356–366\. Springer Nature Switzerland \(2023\)
- \[35\]Törnberg, P\.: Large Language Models Outperform Expert Coders and Supervised Classifiers at Annotating Political Social Media Messages\. Social Science Computer Review43\(6\), 1181–1195 \(2025\)
- \[36\]Vargas, L\., Emami, P\., Traynor, P\.: On the Detection of Disinformation Campaign Activity with Network Analysis\. In: CCSW 2020\. pp\. 133–146 \(2020\)
- \[37\]Vishnuprasad, P\.S\., Nogara, G\., Cardoso, F\., Cresci, S\., Giordano, S\., Luceri, L\.: Tracking Fringe and Coordinated Activity on Twitter Leading Up to the US Capitol Attack\. Proceedings of the International AAAI Conference on Web and Social Media18, 1557–1570 \(2024\)
- \[38\]Wang, S\., Zeng, Y\., Liu, X\., Zhu, E\., Yin, J\., Xu, C\., Kloft, M\.: Effective End\-to\-end Unsupervised Outlier Detection via Inlier Priority of Discriminative Network\. In: Advances in Neural Information Processing Systems\. vol\. 32 \(2019\)
- \[39\]Yoon, J\., Sohn, K\., Li, C\.L\., Arik, S\.O\., Lee, C\.Y\., Pfister, T\.: Self\-supervise, Refine, Repeat: Improving Unsupervised Anomaly Detection\. Transactions on Machine Learning Research \(2022\)
- \[40\]Zhang, Q\., Lipani, A\., Kirnap, O\., Yilmaz, E\.: Self\-Attentive Hawkes Process\. In: Proceedings of the 37th International Conference on Machine Learning\. pp\. 11183–11193\. PMLR \(2020\)
- \[41\]Zhang, S\., Zhou, C\., Zhang, P\., Liu, Y\., Li, Z\., Chen, H\.: Multiple Hypothesis Testing for Anomaly Detection in Multi\-type Event Sequences\. In: ICDM\. pp\. 808–817 \(2023\)
- \[42\]Zhang, Y\., Sharma, K\., Du, L\., Liu, Y\.: Toward Mitigating Misinformation and Social Media Manipulation in LLM Era\. In: WWW 2024 Companion\. pp\. 1302–1305\. Association for Computing Machinery \(2024\)
- \[43\]Zhou, X\., Sharma, A\., Zhang, A\.X\., Althoff, T\.: Correcting misinformation on social media with a large language model \(2026\), arXiv:2403\.11169Similar Articles
Detecting the Undetectable: Enhancing Unsupervised time series Anomaly Detection via Active Learning
Proposes a novel framework combining active learning with masked reconstruction and minimax strategies to improve unsupervised time series anomaly detection, achieving 12.39% AUC improvement over baselines across 28 test cases.
Seeing the Needle in the Haystack: Towards Weakly-Supervised Log Instance Anomaly Localization via Counterfactual Perturbation
This paper introduces LogMILP, a weakly-supervised framework for log instance anomaly localization that uses prototype-guided structural modeling and counterfactual perturbation consistency regularization to improve detection and interpretability with only bag-level labels.
Rethinking Structural Anomaly Detection: From Decision Boundaries to Projection Operators
This paper rethinks structural anomaly detection by shifting from decision boundaries to projection operators onto the low-dimensional manifold of normal data, showing that projection-aligned methods outperform existing boundary-based and reconstruction-based approaches.
Modeling Spectral Energy Shifts in Spatio-Temporal Graph Anomaly Detection
Proposes a node-level spectral energy formulation for detecting camouflaged anomalies in graphs, extending to spatio-temporal settings with energy-driven message passing. Demonstrates effectiveness on large-scale benchmarks.
Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers
The article introduces SAGE, a multi-agent LLM framework for time-series anomaly detection that uses specialized analyzers to improve interpretability and reliability. It demonstrates superior performance over baselines on three benchmarks and enhances diagnostic reporting through structured evidence consolidation.