QUIT: A Human-in-the-Loop Platform for AI Research Automation
Query · Understand · Implement · Tell
✦ Quit the Old Way of Doing Research ✦
What is QUIT?
QUIT is a transparent, human-controllable AI research pipeline — not a black box.
Give it a research topic, and it searches papers, generates ideas, writes experiment code, runs it, and produces a full LaTeX paper.
You stay in control at every step: inspect, edit, steer, or resume from any point in the pipeline.
QUIT also fully supports end-to-end mode — just set a topic and let the pipeline run to completion without any intervention. End-to-end cost with DeepSeek-V4-Pro: roughly ¥10 ($1.5) per manuscript.
✨ Key Features
| Feature | Details | |
|---|---|---|
| 🏆 | Human-in-the-Loop | Stop, inspect, edit artifacts, and resume at any pipeline stage |
| 🚀 | End-to-End Mode | Fully automated run from topic → paper with one click |
| 🖥️ | Web UI | Clean browser interface — no coding required to operate |
| 💰 | Cost-Efficient | Full workflow with DeepSeek-V4-Pro: |
| 🔌 | Multi-Provider | OpenAI · Anthropic · DeepSeek · Qwen · vLLM · any OpenAI-compatible endpoint |
| 📦 | Artifact-Driven | Every output is a plain file — traceable, reproducible, and resumable |
🔄 Pipeline Overview
PLAN → VALIDATE → RETRIEVE → READ → IDEATE → IDEA_EVAL → BUILD_SPEC → CODE → CODE_EVAL → WRITE → WRITE_EVAL → EXTRACT
| Stage | Agent | What happens |
|---|---|---|
| 🗺️ Query | PlannerAgent | Turns your topic into a validated ResearchBrief |
| 💡 Understand | ResearchAgent | Searches papers, extracts evidence cards, generates & ranks ideas |
| 🔧 Implement | BuilderAgent | Writes experiment code, runs it, audits results |
| 📝 Tell | BuilderAgent + ReviewerAgent | Drafts LaTeX paper from actual outputs and reviews it |
A ReviewerAgent independently audits ideas, code quality, and the paper draft at each evaluation stage.
🏗️ Architecture
Four specialized agents coordinated by a central Orchestrator / State Machine.
No agent shares conversational memory across invocations — everything communicates through named artifact files on disk.
This makes the entire pipeline traceable, reproducible, and fully resumable.
🖥️ Web UI
- Configure your topic, LLM provider, and pipeline range from the browser
- Watch live log output as the pipeline runs
- Browse and preview all generated artifacts (evidence cards, code, results, PDF paper)
- Stop at any state, edit intermediate files, and resume — without rerunning earlier stages
🙋 Human-in-the-Loop in Practice
Example workflow:
Run untilIDEA_EVAL→ readIdeaLibrary.jsonl→ delete weak ideas → edit the best one → resume fromBUILD_SPEC.
| Action | How |
|---|---|
| 👁️ Inspect | Open any artifact in the Results panel or your editor |
| ✏️ Edit | Modify BuildSpec.json, EvidenceCards.jsonl, generated code, etc. |
| ⏸️ Stop | Click Cancel, or set a Stop-after state before running |
| ▶️ Resume | Set Start-at to any state and click Start Run — prior artifacts are reused |
| 🔁 Rerun one stage | Set Start-at = Stop-after = <state> and click Start Run |
📄 Demo Papers (Generated End-to-End)
Offline RL Trust-Region Rectified Flow Offline Actor for Low-Latency Generative Policy Learning |
Robot & Energy Online Pricing and Energy-Aware Dispatch for Mobile Charging Robot-Mediated V2G Services |
3D Generation Decoupled Macro–Micro Generation for Editable Large-Scale 3D Surface Patterns |
⚡ Quick Start
Requirements: Python 3.11+, Git, LaTeX (texlive + latexmk)
git clone https://github.com/Mr-XcHan/QUIT.git
cd QUIT
bash setup.sh
Start the Web UI:
source .venv/bin/activate
cd Quit_v0_3_web
python server.py --port 7862
Open http://localhost:7862, fill in your research topic and API key, and click ▶ Start Run.
🔌 Supported LLM Providers
OpenAI · Anthropic · DeepSeek · Qwen · Moonshot · GLM · LMStudio · vLLM · any OpenAI-compatible endpoint
🤝 We're Looking for Testers!
Try the pipeline with your own research idea.
Full auto-run results are welcome — but we're especially excited to feature human-in-the-loop works where researchers stepped in to guide and improve the output.
If you'd like your result showcased, open an issue or PR on GitHub! 🙌
📜 License
Apache 2.0 — free to use and build upon, including commercially.