AutoScholarLoop – AI‑assisted research loop

What it is – An open‑source framework that orchestrates a multi‑agent, multi‑stage research workflow. It mimics a small research group, turning a seed idea and optional references into a full draft paper with audit trails.

Key stages

1. S00 – Field Archive – builds a map of the field, paper cards, methods, datasets and an evidence bank.
2. S01 – Professor Decision – generates ideas, critiques them, probes novelty, ranks and selects a direction.
3. S02 – PhD Execution – plans baselines, runs experiments, analyses failures and writes review memos.
4. S03 – Writing – creates a manuscript plan, claim‑evidence table, drafts text, figures and iterates with reviews.
5. S04 – Quality Control – audits novelty, citations, reproducibility and decides if the draft is ready for submission.

Each stage writes structured Markdown checkpoints, so the whole process is auditable.

Main features

• Deterministic local provider for offline demos (no real model needed).
• OpenAI‑compatible adapter to plug in any OpenAI‑style LLM (e.g., OpenAI, DeepSeek).
• Literature adapters for Semantic Scholar and OpenAlex.
• Dry‑run and local shell execution back‑ends; generated code is saved under code/ and run on the host machine.
• Format‑aware manuscript generation for ACM, IEEE, Springer LNCS and Chinese thesis styles, with optional PDF compilation.
• Vue‑based web console for configuring models, uploading PDFs/notes, visualising loop progress and previewing checkpoints.
• Same pipeline available via CLI (autoscholarloop run …) and a Python web API.

Installation

# Python package (editable install with API, web UI and dev extras)
cd AutoScholarLoop
pip install -e ".[api,web,dev]"
# Front‑end (Vue, built with Vite 2)
cd web
npm install

Quick start (CLI)

autoscholarloop run \
  --seed "study retrieval‑augmented agents for scientific writing" \
  --loop-mode fast \
  --paper-format ieee \
  --workspace runs/demo

Typical options let you supply reference papers, choose the number of ideas, pick a literature source, select an execution backend (dry‑run or real shell), set the review ensemble size, and enable PDF compilation.

Web console

autoscholarloop web --host 127.0.0.1 --port 8000   # start Python API
cd web && npm run dev                               # launch Vue UI

The UI guides first‑time model configuration, uploads of PDFs/notes, selection of loop mode, and live visualisation of S00‑S04 progress.

Outputs – A reproducible workspace is created under the chosen --workspace directory, containing:

• Literature artefacts (field_map.md, paper_cards.md)
• Idea reports and chosen direction
• Experiment code, results and audit logs
• Writing artefacts (paper plan, claim‑evidence table, draft markdown/LaTeX)
• Quality‑gate reports and, if LaTeX is available, a compiled PDF
• A final release/README.md summarising the whole audit trail.

Status – The system provides a scaffold that can run end‑to‑end, but it does not guarantee scientific novelty, correct citations, valid experiments or submission‑ready papers. Human oversight is required.

License & responsible use – Distributed under the AI Scientist Source Code License (a Responsible AI‑type license). Users must disclose AI assistance in any resulting manuscript.

---

AutoScholarLoop is a research‑automation tool from the AI Group at the Chinese Academy of Sciences, aimed at making AI‑driven scientific writing auditable and extensible.