NEXUS rejects generic, repetitive content and delivers only what you truly care about ✅ — integrating real-world dynamic factors, leveraging multi-perspective reasoning to break conventional thinking patterns, and generating genuinely novel insights 💥. With a 60% faster analysis engine, NEXUS delivers deep intelligence in under 30 minutes at a fraction of the cost 🔍

NEXUS Framework: Domain Fine-Tuning × Deep Search × Attention-Driven Analysis × Digital Cognitive Twin — a Multi-Agent collaborative framework that breaks template-based analysis and reshapes intelligent content generation.

🔍 View Highlight Detail

NEXUS bridges "information analysis" and "social system simulation", grounded in verified, real-world information flows.

NEXUS is a Multi-Agent Cognitive Simulation framework for complex real-world information systems. Driven by user Preference Anchors, NEXUS actively extracts deep insights and strategic intelligence aligned with individual cognitive goals. It systematically integrates high-fidelity, highly-relevant signals from social sentiment, policy dynamics, and financial intelligence — building an end-to-end simulation pipeline that reconstructs causal chains between key events in high-dimensional space, enabling dynamic scenario forecasting and genuine cognitive breakthroughs.

• 🚀 Beyond Generic Outputs — Personalized Cognitive Maps 🔥🔥🔥

  • Deep search anchored to what users genuinely care about — not broad, generic information
  • Builds a user-exclusive Cognitive Map rather than one-size-fits-all analysis
  • 📖 Related: Farshidi et al. (2024), Understanding User Intent Modeling; Nguyen et al. (2018), Capsule Network Search Personalization

• 🔗 Domain-Specific Fine-Tuned Models over Generic LLMs 🔥🔥

  • Integrates domain-adapted fine-tuned models — not off-the-shelf LLMs
  • User attention anchor-driven information processing — not indiscriminate data handling

• 📈 Dynamic Trend Simulation over Static Prediction 🔥🔥

  • Incorporates real-time, dynamic influence factors — not static or idealized assumptions
  • Constructs a high-fidelity Digital Cognitive Twin world — not a simplified virtual environment
  • 📖 Related: Guo et al. (2025), Beyond Static Retrieval in GraphRAG; Zhu et al. (2025), Conversational Intent-Driven GraphRAG

• 💡 Multi-Perspective Deep Insight — Jump Outside the Box 🔥🔥🔥

  • Generates genuinely novel ideas — not rigid, template-driven analysis
  • Produces vivid simulated internal monologues — not superficial surface-level outputs
  • 📖 Related: Zhong et al. (2025), Beyond Sentiment; Saunders (2012), Autonomous Creative Systems

• 💡 Four Stakeholder Perspectives Simulated in Parallel 🔥🔥🔥

  • ⚖️ Policymakers & Regulators — e.g., government bodies, think tanks
  • 💰 Business & Market Actors — e.g., VC/PE firms, enterprises, market analysts
  • 🧑‍🤝‍🧑 Public & Community — e.g., KOLs, NGOs, social opinion amplifiers
  • 🧑‍🔬 Researchers & Experts — e.g., academics, industry specialists

• 🌟 60%+ Efficiency Gains over Prior Architecture 🔥🔥🔥

  • Versus the prior mirofish architecture: cost reduced by 60%, simulation time cut to 1/3
  • Achieved via deep integration of background knowledge priors + user attention anchor focusing, dramatically reducing token consumption while preserving analytical depth
  • Cost: Single simulation reduced from $5 → $2–3 (40–60% reduction)
  • Speed: Only 30 simulation rounds needed vs. prior 96 rounds; total time 18–20 min vs. 50 min (>60% faster)

NEXUS supports multi-domain, multi-role, multi-perspective, high-fidelity simulation of complex information systems and future scenario forecasting.

• 🔮 Ripple Effect Forecasting: Dynamically simulate the economic, social, and public safety chain reactions following policy announcements
• ⚠️ Early Risk Detection: Identify latent social conflicts and public sentiment crises; output proactive intervention strategies
• 📊 Causal Impact Evaluation: Rigorously decompose the causal effects of policy adjustments with evidence-backed analysis

• 🔮 Trend Intelligence: Fuse multi-source signals with user anchors to dynamically simulate industry inflection points and shifts in consumer mindset
• ⚔️ Competitive Landscape Simulation: Real-time modeling of competitor moves and market responses; assess the impact potential of new products, pricing, and campaigns
• 🧪 Strategy Validation: Run multi-scenario market simulations to pre-test revenue, market share, and brand reputation outcomes

• 📊 Emotion Heatmap Tracking: Real-time monitoring of social media discussions and the dynamic evolution of public sentiment
• 🎯 Key Voice Identification: Pinpoint opinion leaders (KOLs) and quantify their true influence weight on events
• ⚡ Sentiment Storm Early Warning: Anticipate potential viral events; simulate propagation scale, pathways, and risk levels

• 🧠 Knowledge Graph Construction: Integrate academic papers, policy reports, and cutting-edge technology into a queryable, reasoning-ready dynamic knowledge network
• 🔮 Frontier Trend Detection: Based on historical data and real-time signals, precisely capture the early emergence of new technologies, market opportunities, and research hotspots

💡 Not just prediction — simulation. Stay one cognitive step ahead, with every decision grounded in evidence.

Goal: Anchor on user attention points; enhance model comprehension for specific domains (e.g., current affairs, social discourse, financial news, policy documents).

All-in-One Pipeline: Data Management → Data Augmentation → Model Fine-Tuning → Model Quantization → Local Deployment 📦

Powered by the personal pipeline project LLaMA Factory Pro:

• Input: User-defined unstructured data (PDFs, news articles, policy documents, etc.) → fully automated QA-pair generation
• Output: Enhanced fine-tuned model + online monitoring + deployment

1️⃣ Cloud Platform AutoDL ☁️ — Automated fine-tuning environment (Docker image config), .venv one-stop management, cloud storage support (AWS S3, Azure Blob, Hugging Face Hub, Ollama Registry)

2️⃣ Data Engineering — EasyDataset: auto-detects multimodal formats (text/image/mixed), auto-generates high-quality context-aware QA pairs

3️⃣ Training — Hyperparameter configuration → online supervised fine-tuning monitoring → auto-save to cloud storage upon completion

4️⃣ Deployment — Local model serving via Ollama / llama.cpp

Goal: Provide high-quality analytical priors for NEXUS through deep search, user-focus anchoring, multi-agent reasoning, and structured information modeling.

This module consists of two components: UPAIRS-DS (Pressure-Driven) and CognitiveTemp-DeepSearch Agents.

🔎 UPAIRS Deep Search Engine Built on UPAIRS Agents, this engine uses a Pressure-Driven paragraph-level iterative reflection architecture to continuously refine retrieval paths in complex information environments, ensuring high factual fidelity. 🔎 CognitiveTemp-DeepSearch Users can further customize search strategies via CognitiveTemp-DeepSearch-Agents to obtain results with stronger relevance, reliability, and timeliness. Key mechanisms: 🔄 Paragraph-level Iterative Reflection 🎯 Pressure-Driven Reflection Model 🧠 Multi-agent Collaborative Reasoning 📊 Uncertainty Quantification ⚡ Intelligent Budget Allocation 🎨 Source Quality Scoring 🔍 Noise-Filtered Re-ranking

Figure 3: Deep Search Interface (left) — Query Input and Search Result Exploration (right)

This module is composed of three sub-modules:

(1) User Attention Anchoring → (2) Multi-Agent Deep Analysis → (3) Structured Knowledge Output

Users perform targeted deep searches (e.g., academic topics like GraphRAG or social events) with customizable parameters: recency, relevance threshold, and source selection.

Figure 4: GraphRAG academic search (left) — Social event search: Bilibili streamer incident (right)

User Focus Annotation — Via the Hypothesis API Google Plugin, users highlight content of interest and add free-form natural language annotations (notes, questions, doubts). The system captures all highlights and annotations in real time, outputting structured JSON / JSONL data.

User Attention Annotation via Hypothesis API Users highlight content of interest; system captures in real time Free-form natural language annotations on any selected passage Output as structured JSON / JSONL for downstream agents Figure 6: Hypothesis API — Real-time user attention capture

Figure 7: GraphRAG Search Detail — Entity & Relation Visualization (left); Social Opinion Search — Semantic & Context Highlight (right)

In one line: Agent-A/B/C Chain Analysis + Agent-D Strategic Synthesis → High-relevance, high-timeliness, high-fidelity, highly-customized strategic reports

Analysis mode control: Express / Balanced / Deep | Relevance threshold configurable

Figure 8: Report Generation Interface (left) — User Input & Configuration; Generated Report Result (right) — Structured Summary & Insights

Goal: Within a high-fidelity virtual digital twin environment, simulate real-world public sentiment, event development, and user attention dynamics through multi-agent interactions for deep insight and predictive forecasting.

User-defined attention anchors (topics, key entities, specific events) are fused with multi-source information. Structured entity-relation extraction builds an enhanced GraphRAG representation with self-loop relations and multi-dimensional links, improving semantic depth and traceability.

# Entity & Relation Extraction Example (GraphRAG Academic Scenario)
Technical Foundations
├─ RAG (Retrieval-Augmented Generation)
├─ Knowledge Graph
│   ├─ is_a ...
│   └─ used_in ...
Frameworks
├─ GraphRAG
│   ├─ uses → RAG
│   └─ built_on → Knowledge Graph
Application Systems
├─ MED-COPILOT
│   ├─ uses → GraphRAG
│   ├─ uses → Knowledge Graph
│   └─ has_risk → ⚠️ Hallucination / Calibration Bias
...

Figure 9a: Entity & Relation Extraction — Extracting entities and relations from text and event streams to provide high-quality priors for GraphRAG.	Figure 9b: GraphRAG Construction — Generating a reasoning-ready knowledge graph from extracted entities and relations, supporting multi-dimensional linkage.

Figure 10a: Graph Build — Initial stage: entity and relation import.	Figure 10b: Graph Expansion — Adding links and multi-dimensional connections.

Figure 10c: Graph Visualization — Global view of all nodes and relationship network structure.	Figure 10d: Complete Knowledge Graph — Full graph supporting multi-hop reasoning and querying.

Based on the information priors from Step 1, the high-fidelity virtual world Canyon is initialized with trending topics, contextual background, and activation sequences. Multiple virtual agents are generated, each with unique knowledge bases, cognitive structures, behavioral tendencies, and data associations.

Each virtual agent persona includes:

• Basic Attributes: Age, occupation, domain expertise
• Cognitive Characteristics: Information processing style, decision-making approach, risk preference
• Behavioral Tendencies: Social inclination, information-sharing habits
• Influence Level: Scope and weight of impact within the simulated ecosystem

Sample Virtual Agents (GraphRAG Academic Scenario):

Figure 11a: Agent Persona Build — Initial stage: basic attribute definition.	Figure 11b: Agent Persona Expansion — Adding preferences and behavioral inclinations.

Figure 11c: Agent Persona Refinement — Adjusting cognitive strategies and interaction styles.	Figure 11d: Agent Persona Completion — Full cognition and behavior patterns seeded from real-world data.

Figure 11e: Agent Persona Configuration — Customizing attributes and preferences per user requirements.	Figure 11f: Agent Persona Lookup — Inspecting each agent's event cognition and behavioral characteristics.

Within the Canyon digital twin world, the system simulates the evolution of real-world public sentiment by injecting environmental factors to guide information flows and opinion dynamics. This step models sentiment evolution, hot-topic aggregation, and information diffusion patterns.

Figure 12a: Hot Topic Detection — Identifying high-attention topics within the Canyon virtual world.	Figure 12b: Canyon Simulation Overview — System-wide operation and event flow structure.

Figure 12c: Virtual Agent Dialogue Sample 1 — Agent-to-agent discussion of trending topics.	Figure 12d: Virtual Agent Dialogue Sample 2 — Diverse agent reactions and cognitive simulations.

The system conducts multi-round iterative simulation, generating event development reports covering outbreak dynamics, sentiment diffusion trends, and critical node analysis. Each virtual agent's reactions and behavior patterns are recorded, enabling deep insight queries and predictive analysis.

This project compares two contrasting scenarios:

🧑‍🏫 Simulated Face-to-Face Interviews:

• Academic Scenario: GraphRAG technology development
• Social Media Scenario: Bilibili streamer account-destruction incident

Figure 13a: Academic Agent Interview 1 — Five-year GraphRAG development trajectory and key beneficiary sectors	Figure 13b: Social Opinion Streamer-Agent Interview 1 — Platform training necessity and technical countermeasures against violations

Figure 13c: Academic Agent Interview 2 — GraphRAG challenges in healthcare and expert priorities in RAG research	Figure 13d: Social Opinion Victim-Agent Interview 2 — User trust erosion and emerging virtual property protection policy discourse

Figure 13e: Academic Institution-Agent Interview 3 — Financial institution's GraphRAG application outlook and risk assessment	Figure 13f: Social Opinion Policymaker-Agent Interview 3 — Clarifying virtual item ownership under property law and debt law

Figure 13g: Academic Institution-Agent Interview 4 — Long-term impact of GraphRAG on information infrastructure	Figure 13h: Social Opinion Policymaker-Agent Interview 4 — Balancing content creation freedom with platform accountability

🔍 Deep Insight

The academic research scenario (GraphRAG technology) and the social media scenario (Bilibili account incident) represent two informationally distinct ecosystems. Through Agent Interview Simulation in Canyon, agents exhibited markedly different cognitive frameworks, information processing approaches, reasoning pathways, and behavioral responses to identical topics or events. Academic agents tended toward structured knowledge integration, technical evolution analysis, and theoretical reasoning, while social media agents focused on emotional propagation, collective responses, and opinion evolution dynamics. This contrastive simulation reveals the underlying cognitive and propagation mechanisms across different information domains.

Figure 14a: Entity-Level Insight (Academic – GraphRAG). Structural knowledge associations and latent research trajectories revealed through entity-relation analysis.	Figure 14b: Entity-Level Insight (Social Opinion). Key persons, events, and organizations identified; influence weights and propagation patterns mapped.

Figure 14c: Question-Level Insight (Academic – GraphRAG). Multi-hop knowledge retrieval and integration for deep research insight generation.	Figure 14d: Question-Level Insight (Social Opinion). Question-driven reasoning to understand core public concerns and cross-perspective interaction dynamics.

After simulated interviews and deep analysis, users can interactively engage with the generated virtual world reports and individual virtual agents. Through multi-turn dialogue and strategy exploration, users gain access to each virtual entity's latent reasoning paths, internal concerns, and projected behavioral trends.

Structured Comparative Analysis — Academic vs. Social Media:

Official Position vs. Inner Monologue (NEXUS's unique insight capability — revealing authentic sentiment):

📑 Detailed analytical reports:

Figure 15a: GraphRAG Analytical Report — Comprehensive report for the academic research scenario: structured knowledge integration, multi-hop reasoning, systematic summary of GraphRAG trends, research directions, and application domains.	Figure 15b: Social Opinion Event Report — Sentiment analysis report for the Bilibili streamer incident: event background, opinion propagation, platform ecosystem impact, and structured synthesis of key public concerns.

Figure 15c: GraphRAG Agent Mindset — Academic agent's internal reasoning and cognitive synthesis post-interview: understanding of GraphRAG's future trajectory, latent research challenges, and potential breakthroughs.	Figure 15d: Social Opinion Agent Mindset — Emotional and cognitive response records from social opinion agents: attitudinal divergence among virtual individuals and reflections on platform governance.

1. Nguyen, N. T., et al. (2018). A Capsule Network-based Embedding Model for Search Personalization. arXiv:1804.04266. 📄 https://arxiv.org/abs/1804.04266

2. Farshidi, F., et al. (2024). Understanding User Intent Modeling for Conversational Recommender Systems. User Modeling and User-Adapted Interaction, Springer Nature. 🔗 https://link.springer.com/article/10.1007/s11257-024-09398-x

3. Guo, Q., et al. (2025). Beyond Static Retrieval: Opportunities and Pitfalls of Iterative Retrieval in GraphRAG. arXiv:2509.25530. 📄 https://arxiv.org/abs/2509.25530

4. Zhu, Y., et al. (2025). Conversational Intent-Driven GraphRAG. arXiv:2506.19385. 📄 https://arxiv.org/abs/2506.19385

5. Zhong, V., et al. (2025). Beyond Sentiment: Cognitive and Narrative Open-Ended Generation. arXiv:2502.13925v1. 📄 https://arxiv.org/html/2502.13925v1

6. Saunders, R. (2012). Towards Autonomous Creative Systems: A Computational Approach. Cognitive Computation, 4(3), 216–225. 🔗 https://www.researchgate.net/publication/257788397

@article{nguyen2018capsule,
  title={A Capsule Network-based Embedding Model for Search Personalization},
  author={Nguyen, Nam Tran and et al.},
  journal={arXiv preprint arXiv:1804.04266},
  year={2018}
}

@article{farshidi2024understanding,
  title={Understanding User Intent Modeling for Conversational Recommender Systems},
  author={Farshidi, F. and et al.},
  journal={User Modeling and User-Adapted Interaction},
  year={2024},
  doi={10.1007/s11257-024-09398-x}
}

@article{guo2025graphrag,
  title={Beyond Static Retrieval: Opportunities and Pitfalls of Iterative Retrieval in GraphRAG},
  author={Guo, Q. and et al.},
  journal={arXiv preprint arXiv:2509.25530},
  year={2025}
}

@article{zhu2025conversational,
  title={Conversational Intent-Driven GraphRAG},
  author={Zhu, Y. and et al.},
  journal={arXiv preprint arXiv:2506.19385},
  year={2025}
}

@article{zhong2025beyond,
  title={Beyond Sentiment: Cognitive and Narrative Open-Ended Generation},
  author={Zhong, V. and et al.},
  journal={arXiv:2502.13925v1},
  year={2025}
}

@article{saunders2012autonomous,
  title={Towards Autonomous Creative Systems: A Computational Approach},
  author={Saunders, R.},
  journal={Cognitive Computation},
  volume={4},
  number={3},
  pages={216--225},
  year={2012}
}