Beyond Individual Mimicry: Constructing Human-Like Social network with Graph-Augmented LLM Agents

Haoran Bu; Litian Zhang; Chuxuan Zhang; Zhanyuan Liu; Hui Pang; Xi Zhang

cs.SIMar 31, 2026

Beyond Individual Mimicry: Constructing Human-Like Social network with Graph-Augmented LLM Agents

Authors: Haoran Bu, Litian Zhang, Chuxuan Zhang, Zhanyuan Liu, Hui Pang, Xi Zhang

Organizations: Cyberspace Security, Beijing University of Posts and Telecommunications · Institute of Software, Chinese Academy of Sciences

Abstract

Driven by large language models (LLMs), social bot can autonomously engage in local interactions, whose human-like behaviors enable them to evade social bot detection. However, while these botnets exhibit realistic local social interactions, they fail to preserve human-like social network. This is because LLM-based bots are graph-unaware and cannot coordinate over global interactions, which makes those botnets vulnerable to graph neural network (GNN)-based detection. To address this limitation, we propose GraphMind, which equips LLM-driven social bots to explicitly learn and fit human-like social network structures. Building on this foundation, we further construct GraphMind-Botnet, a LLM-driven botnet designed to evaluate the performance of existing social bot detection algorithms. Experiments on datasets derived from GraphMind-Botnet show that both text-based and graph-based detection models show substantially degraded performance in distinguishing. Our results highlight the critical role of social link construction in LLM-driven social network generation, while exposing fundamental weaknesses in existing bot detection mechanisms.

Figures & tables

Figure 1: The comparation between existing LLM-driven social bot and GraphMind (ours). Our botnet resembles more human-like social networks, leading to significantly improved evasion performance against GNN-based detection.

Figure 2: Overview of the framework. (Left) GraphMind Framework, where two modules are employed to enable LLMs to generate diverse, strength-aware interactions and to construct multi-hop follow chains, thereby mitigating isolated nodes. (Right) Botnet simulation, in which GraphMind social bots autonomously build human-like social networks, improving structural realism and enhancing robustness against GNN-based detection methods.

Human	Bot	Edges	Edge Types	Communities
1,000	1000	41375	5	50

Table 1: The composition of the our dataset

Dataset	Metric	Metadata-based				Text-based	Homo-GNN		Heter-GNN
Dataset	Metric	AB	RF	DT	SVM	Wei et al.	GCN	GAT	BotRGCN	RGT	S-HGN
Cresci-15	Acc	96.9	96.0	94.7	95.4	94.78	98.2	97.7	96.3	97.2	96.9
Cresci-15	F1	95.5	95.6	94.6	95.3	84.25	96.0	97.0	96.3	96.5	95.9
Cresci-17	Acc	93.2	88.1	86.2	85.1	86.30	/	/	/	/	/
Cresci-17	F1	83.4	78.9	79.4	76.8	79.40	/	/	/	/	/
Twibot-20	Acc	85.9	85.4	81.1	85.7	78.26	76.8	83.2	86.8	87.4	85.9

Table 2: Bot detection performance across datasets (lower values indicate stronger evasion).

Panel of Figure 3 — Figure 3: Structural property analysis of different network

Panel of Figure 5 — Figure 5: Network visualizations comparing human and bot follow networks across different datasets.

Methods	S-HGN		BotRGCN		RGT
Methods	Acc	F1	Acc	F1	Acc	F1
All
Qwen3-1.7B	70.7	60.5	72.4	67.6	73.5	65.8
Qwen3-8B	69.1	68.4	70.5	62.6	69.6	65.3
w/o FIM
Qwen3-1.7B	78.9	69.5	79.8	74.2	80.5	73.1

Table 3: Ablation Study of GraphMind Modules on Detection Evasion (%).

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Level	Definition
Level 1 (Support clique)	Strongest ties;
	frequent contact
Level 2 (Sympathy group)	Close ties; frequent but
	not weekly contact
Level 3 (Affinity group)	Casual ties;
	occasional contact

Appendix

Table 4: Relationship Circle Definitions

Hyperparameter	Value
Optimizer	AdamW
Learning rate	$5\times 10^{-5}$
Training epochs	5
LR scheduler	Cosine
Warmup steps	100
Random seed	42

Appendix

Table 5: Key hyperparameters used for FIM.

Hyperparameter	Value
Optimizer	AdamW
Learning rate	$5\times 10^{-5}$
Training epochs	5
LR scheduler	Cosine
Warmup steps	100
Random seed	42

Appendix

Table 6: Key hyperparameters used for GSI.

Field	Content
Instruction	You are an expert in social media behavior analysis. Your task is to infer the social relationship level between two users from an ego-network perspective, and then generate a list of interaction actions that follow realistic behavioral patterns. Follow the steps below: Step 1: Infer the most plausible social relationship level within the ego network. Step 2: Based on the inferred relationship strength, determine the expected interaction intensity and cost.
Input	<user_profile> : {PROFILE} <tweets> : {TWEETS}
Think	The target user belongs to my ego network with relationship level {RELATIONSHIP_LEVEL}. Based on this relationship strength and the observed profile and interaction context, I infer appropriate interaction intensity and engagement patterns.
Output	A list of interaction actions in the following structured format: <action> <type> like / retweet / comment / follow </type> <tweet_id> {TWEET_ID} </tweet_id>

Appendix

Table 7: Prompt template used for FIM.

Field	Content
Instruction	You are an expert in social media behavior analysis. Given a source node and a target node, generate a multi-hop following path based on the input social network structure. First, select a set of potential intermediate nodes $\{n\}$ from the given network, and then predict and generate potential social links according to their profile information. Follow the steps below: Step 1: Select candidate intermediate nodes from the input graph based on profile similarity to the source node. Step 2: Sequentially infer plausible following relationships between adjacent nodes to form a multi-hop path. Requirements: 1. The path must contain $\geq 3$ hops (Source $\rightarrow$ Mediator 1 $\rightarrow$ Mediator 2 $\rightarrow$ Target). 2. All mediator nodes must be selected from the given network information. 3. Path inference must be based on feature similarities (bio, age, education, location, etc.). 4. The connection rationale must be analyzed for each hop.
Input	<graph> <nodes> {users} </nodes> <edges> {follow_edges} </edges> </graph> <source_node> {user_A} </source_node> <target_node> {user_B} </target_node>
Think	<think> [Link 1] <user>1</user> follow <user>2</user> <reason>...</reason> [Link ..] </think>
Output	Based on the analysis of user profiles and social relationships, the most plausible multi-hop following path is: user_A $\rightarrow$ mediator_1 $\rightarrow$ mediator_2 $\rightarrow$ user_B .

Appendix

Table 8: Prompt template used for GSI.

Figure 8: Botnet visualization of Twibot-20.

Figure 9: Human network visualization of Twibot-20.

Figure 10: Botnet visualization of Twibot-22.

Figure 11: Visualization of EvoBot (human + bot).

Figure 12: Botnet Visualization of EvoBot.

Explore similar work

TRACE-Bot: Detecting Emerging LLM-Driven Social Bots via Implicit Semantic Representations and AIGC-Enhanced Behavioral Patterns
Apr 2, 2026Zhongbo Wang, Zhiyu Lin, Zhu Wang +1LLM Security
School of Cyber Science and Engineering, Sichuan University, Sichuan 610207, China · School of Software and Microelectronics, Peking University, Beijing 102600, China · College of Computer Science, Sichuan University, Sichuan 610065, China +1
CSC: Calibrated Simplicity for Conflict-Aware Social Bot Detection in the LLM Era
Sep 20, 2026Yipeng Qian, Pengjie Zhao, Chaoxi NiuSocial Media Analysis Probability Calibration
Minzu University of China Beijing, China · City University of Macau Macau, China
RABot: Reinforcement-Guided Graph Augmentation for Imbalanced and Noisy Social Bot Detection
Feb 25, 2026Longlong Zhang, Xi Wang, Haotong Du +3Data Augmentation Class-Imbalanced Learning
School of Artificial Intelligence, Optics and Electronics (iOPEN), Northwestern Polytechnical University · Shenzhen Research Institute, Northwestern Polytechnical University · School of Automation Science and Engineering, Xi’an Jiaotong University +1

Apr 2, 2026J/K move · Enter open · S save

TRACE-Bot: Detecting Emerging LLM-Driven Social Bots via Implicit Semantic Representations and AIGC-Enhanced Behavioral Patterns

Zhongbo Wang, Zhiyu Lin, Zhu Wang, Haizhou Wang

School of Cyber Science and Engineering, Sichuan University, Sichuan 610207, China · School of Software and Microelectronics, Peking University, Beijing 102600, China · College of Computer Science, Sichuan University, Sichuan 610065, China+1

LLM Security

Open paper AskPDF arXiv

Large Language Model-driven (LLM-driven) social bots pose a growing threat to online discourse by generating human-like content that evades conventional detection. Existing methods suffer from limited detection accuracy due to overreliance on single-modality signals, insufficient sensitivity to the specific generative patterns of Artificial Intelligence-Generated Content (AIGC), and a failure to adequately model the interplay between linguistic patterns and behavioral dynamics. To address these limitations, we propose TRACE-Bot, a unified dual-channel framework that jointly models implicit semantic representations and AIGC-enhanced behavioral patterns. TRACE-Bot constructs fine-grained representations from heterogeneous sources, including personal information data, interaction behavior data and tweet data. A dual-channel architecture captures linguistic representations via a pretrained language model and behavioral irregularities via multidimensional activity features augmented with signals from state-of-the-art (SOTA) AIGC detectors. The fused representations are then classified through a lightweight prediction head. Experiments on two public LLM-driven social bot datasets demonstrate SOTA performance, achieving accuracies of 98.46% and 97.50%, respectively. The results further indicate strong robustness against advanced bot strategies, highlighting the effectiveness of jointly leveraging implicit semantic representations and AIGC-enhanced behavioral patterns for emerging LLM-driven social bot detection.

Abstract

Explore similar work

TRACE-Bot: Detecting Emerging LLM-Driven Social Bots via Implicit Semantic Representations and AIGC-Enhanced Behavioral Patterns

CSC: Calibrated Simplicity for Conflict-Aware Social Bot Detection in the LLM Era

RABot: Reinforcement-Guided Graph Augmentation for Imbalanced and Noisy Social Bot Detection