🔬 TrustResearcher: Automating Knowledge-Grounded and Transparent Research Ideation with Multi-Agent Collaboration

Elegant, Multi-Agent Research Ideation Prototype — from literature search to refined, distinct, well‑reviewed ideas — with clear logs, reproducible artifacts, and minimal setups.

---

📰 News

• [2026.02.09] 🎉 Major architecture overhaul with clearer orchestration, improved agent delegation, and modular skill system. Continuous improvements planned with emerging models and methods.
• [2026.01.12] 🏆 Accepted as WWW 2026 Demo Track paper! See you in Dubai.
• [2025.11.11] 📝 Revised paper submission with comprehensive appendix showcasing preliminary results.
• [2025.10.20] 📄 Initial paper submission now available on arXiv:2510.20844.

---

✨ Key Features & Pipeline

TrustResearcher integrates a fully literature-aware, multi-agent workflow that bridges retrieval, reasoning, and review — ensuring both novelty and evidence grounding.

The research pipeline consists of 5 main phases:

1. Literature Retrieval & Knowledge Graph Construction → Semantic Scholar API search to gather relevant papers, followed by knowledge graph construction to organize concepts, relationships, and research context.
2. Idea Generation → Planning module generates strategic research directions, followed by two complementary generation approaches (direct generation and variant generation), with preliminary critique for refinement and cross-pollination to enhance idea quality and diversity.
3. Preliminary Selection → Two-stage filtering: (a) External selection against literature for novelty, (b) Internal deduplication and diversity selection.
4. Detailed Review → Multi-criteria expert review evaluating ideas across novelty, feasibility, impact, and clarity dimensions.
5. Final Selection → Score-based ranking and selection of top ideas (≥3.5 threshold) for final output.

---

⚙️ Installation

Requirements

• Python 3.8+
• Network access for the model API and Semantic Scholar
• [Optional] Dedicated GPU server for real-time deduplication and KG relationship augmentation (may use preset embedding model)

Install

pip install -e .

Configure Credentials

Set up your API credentials in the following config files:

• configs/llm.yaml - LLM provider API key (OpenAI, Anthropic, etc.)
• configs/literature_search.yaml - Semantic Scholar API key

Other config files (idea_generation.yaml, reviewer.yaml, etc.) contain pipeline parameters and don't require credentials.

---

🚀 Quick Start

• CLI

# help
python -m src --help

# full pipeline (ensure configs/agent_config.yaml is set)
python -m src --topic "Design scalable and robust algorithms for the k-truss breaking problem that bypass global trussness updates via localized, incremental, and approximation methods, enabling near-real-time interventions on large-scale graphs." --num_ideas 2 --debug

python -m src --topic "Design scalable multi-agent systems for automated scientific ideation that replace linear generation with knowledge-graph-driven planning and RL-driven exploration, enabling grounded, explainable, and self-evaluated hypothesis generation via structured reasoning and adversarial peer-review loops, with publishable paper and great insights to the general AI field." --num_ideas 2 --debug

• Web UI

# process visualization UI
python -m src.ui_launcher --process-ui

# set UI host (default: localhost; use 0.0.0.0 for LAN)
python -m src.ui_launcher --process-ui --process-host 0.0.0.0

# set UI port (default: 7860)
python -m src.ui_launcher --process-ui --process-port 7861

---

📤 Outputs & 📜 Logs

• Results: outputs/{timestamp}.json with the complete pipeline output.
• Run logs: logs/session/{timestamp}.log (execution logs for each run).
• LLM logs: logs/llm/{timestamp}.jsonl (all LLM interactions with token & cost stats).
• Idea logs: logs/idea/{timestamp}.json (all generated ideas for each refinement stage).

---

🖼️ Case Study: Web UI in Action

Here’s what the interactive Web UI looks like when running a research session:

---

🧯 Troubleshooting

• Always run as a module: python -m src ... (avoid python src/main.py).
• Ensure write permissions for outputs/, logs/.

---

📚 Citation

If you find this work useful, please cite our paper:

@misc{zhou2025autoresearcher,
      title        = {{TrustResearcher}: Automating Knowledge-Grounded and Transparent Research Ideation with Multi-Agent Collaboration},
      author       = {Jiawei Zhou and Ruicheng Zhu and Mengshi Chen and Jianwei Wang and Kai Wang},
      year         = {2025},
      eprint       = {2510.20844},
      archivePrefix= {arXiv},
      primaryClass = {cs.MA},
      url          = {https://arxiv.org/abs/2510.20844}
}